ACMMM2026

Abstract:
Multiview learning aims to exploit cross-view consistency and complementarity to improve prediction performance. In practice, however, asynchronous acquisition, transmission errors, and sensor malfunctions may introduce \emphnoisy correspondence, where some views within a sample are misaligned, inevitably breaking the cross-view alignment assumption underlying standard multiview learning and corrupting both training and inference. Existing methods typically learn to downweight mismatched views at inference time, while assuming correctly aligned views during training. As a result, they overlook the biased learning induced by mismatched training views, which act as view-level noisy supervision, thus leading to inaccurate reliability estimation and performance degradation. To address this issue, we propose CARF, a Consistency-Aware Robust Fusion framework for robust multiview learning under noisy correspondence, which integrates two key components. Specifically, we introduce a robust Hellinger-based objective that aligns each view with the fused prediction (\ie consensus) while mitigating the adverse impact of mismatched views, thereby promoting cross-view consistency among correctly matched views and stabilizing multiview fusion. To further suppress the influence of noisy views, we estimate view reliability by jointly modeling \emphview consistency and \emphlabel consistency, which measure decision-level agreement across different views and fidelity to the assigned label, respectively. The resulting reliability scores are then used to reweight the view-specific optimization during training and to perform reliability-aware fusion during both training and inference, thus yielding more robust consensus learning and more reliable final prediction. Extensive experiments on nine datasets demonstrate that CARF consistently outperforms state-of-the-art methods on both clean and noisy data. Code is available at \urlhttps://github.com/XLearning-SCU/2026-ACM-MM-CARF.

Abstract:
Multi-view classification aims to exploit complementary and consistent information across multiple views to learn discriminative representations. However, in real-world scenarios, multi-view data frequently suffer from arbitrary view missing, thus Incomplete Multi-View Classification (IMVC) becomes a widespread research hot. Existing imputation-based IMVC methods typically treat view completion and downstream classification task as separate optimization objectives, which neglect the crucial information of task identity during imputation. Consequently, the generated pseudo-features often lack discriminative semantics, which inevitably introduces semantic noise and disrupts cross-view information, thereby decreasing classification performance. To address the above limitation, we propose a MetA-learning Guided Imputation framework for Incomplete multi-view Classification (MAGIC), which reformulates view imputation as a task-driven adaptive optimization process guided by downstream classification objective. Specifically, MAGIC comprises three core components, i.e., Multi-view Quality Evaluation (MQE), Meta-learning Missing Imputation (MMI), and Adaptive Cross-view Contrastive (ACC). MQE quantifies the reliability of each view to mitigate interference from low-quality data. MMI utilizes the task feedback to learn how to impute with high quality for incomplete multi-view data. To preserve distinct decision boundaries, ACC employs a dynamic modulation of negative-sample repulsion, thereby boosting the model's discriminative power. Extensive experiments on seven benchmark datasets demonstrate that our MAGIC achieves superior performance over state-of-the-art IMVC methods under different missing rates. The code is released at https://github.com/7ktsy/MAGIC.

Abstract:
Multi-view clustering (MVC) has become more prevalent with the increasing multi-view data in real-world applications. Nonetheless, modeling the underlying data relationships and learning the discriminative latent embedding remain challenging. The local geometric structure and similarity among samples encode rich relational information that is crucial for spectral embedding. Moreover, a strategy for assigning an individual weight to each view is necessary. To address these challenges, we propose a novel Dual-Alignment Multi-view Clustering via Adaptive Tensor Learning (DAAT), where a point-wise and super-relational alignment mechanism is proposed to jointly model first-order and higher-order coherence among samples. This design not only enforces sample-level alignment but also captures super-similarity, thereby preserving the intrinsic manifold structure of the data. Furthermore, an adaptive tensor learning is developed, which incorporates an exponential mapping to enhance the discrimination between significant and trivial singular values. Meanwhile, a dynamic weighting strategy is introduced to measure the reliability of each view, allowing the model to emphasize informative views and reduce the influence of less reliable ones. Experimental results on several benchmark datasets verify the effectiveness and superiority of DAAT over state-of-the-art methods. The code is available at https://github.com/jiaxinyang04/DAAT.

Abstract:
Current methods for incremental object detection typically assume either a fixed category set or a stationary data distribution, overlooking a more realistic scenario in which both the category space and domain evolve over time. In this work, we study domain-class dual-incremental object detection, where models must concurrently handle newly introduced categories with shifts in data distribution across domains, requiring both in-domain knowledge retention and cross-domain knowledge generalization. In response, we propose Semantic-Centric Cross-Domain Learning (SCL-IOD), a unified framework that leverages a domain-invariant semantic space to decouple domain bias while preserving transferable knowledge across incremental steps. Specifically, SCL-IOD consists of two key components: 1) Semantic-Centric Knowledge Transfer aligns region-level features with a domain-invariant semantic space to stabilize knowledge retention under evolving domains; and 2) Domain-Aware Prototype Reconstruction models domain-specific variations in the prototype space to synthesize features for unseen domain-category combinations, enhancing cross-domain generalization. Extensive experiments on challenging DuIOD benchmarks demonstrate the effectiveness and superiority of the proposed method. Code is available at \hrefhttps://github.com/LuowenDragon/SCL-IODhttps://github.com/LuowenDragon/SCL-IOD.

Abstract:
Task decomposition is crucial for enabling robots to master long-horizon compositional environments. Towards this goal, skill primitive learning offers a promising approach. However, solely relying on a single inferred skill abstraction creates an expression bottleneck, rendering it insufficient to capture the complex, heterogeneous kinematic behavior combinations required by tasks even with simple semantic descriptions. To address this, we propose SkillAttention: a framework that dynamically synthesizes context-aware latent skills to ensure precise alignment with task-specific kinematic requirements. These synthesized skills, jointly conditioned with the robot's dynamic state, serve as guidance for a downstream diffusion module to generate coherent motion trajectories tailored to the current task execution status. Additionally, we introduce a language-primed staged fusion module that reinforces the guiding role of textual instructions, effectively mitigating instruction-following failures caused by modality imbalance. Extensive experiments on compositional long-horizon benchmark LOReL demonstrate that SkillAttention achieves significant improvements in both success rates and skill learning efficiency. Transcending simulation-centric state-of-the-art models, our framework successfully bridges the sim-to-real gap, maintaining consistent performance in real-world precise manipulation tasks that are significantly more challenging than their simulated counterparts. Code is available at \urlhttps://github.com/is-aHuan/SkillAttention.

Abstract:
\beginabstract Flow-based vision-language-action (VLA) policies offer strong expressivity for action generation, but suffer from a fundamental inefficiency: multi-step inference is required to recover action structure from uninformative Gaussian noise, leading to a poor efficiencyquality trade-off under real-time constraints. We address this issue by rethinking the role of the starting point in generative action modeling. Instead of shortening the sampling trajectory, we propose CF-VLA, a coarse-to-fine two-stage formulation that restructures action generation into a coarse initialization step that constructs an action-aware starting point, followed by a single-step local refinement that corrects residual errors. Concretely, the coarse stage learns a conditional posterior over endpoint velocity to transform Gaussian noise into a structured initialization, while the fine stage performs a fixed-time refinement from this initialization. To stabilize training, we introduce a stepwise strategy that first learns a controlled coarse predictor and then performs joint optimization. Experiments on CALVIN and LIBERO show that our method establishes a strong efficiencyperformance frontier under low-NFE (Number of Function Evaluations) regimes: it consistently outperforms existing NFE=2 methods, matches or surpasses the NFE=10 \pi_0.5 baseline on several metrics, reduces action sampling latency by 73.2%, and achieves the best hard-task real-robot success rate of 74.7%, outperforming MIP by 26.0 points and reaching comparable or slightly better performance than \pi_0.5 with a fivefold smaller inference budget. These results suggest that structured, coarse-to-fine generation enables both strong performance and efficient inference. Our code is available at \urlhttps://github.com/EmbodiedAI-RoboTron/CF-VLA. \endabstract

Abstract:
In zero-shot 3D instance segmentation, recent methods primarily lift multi-view 2D masks into 3D point clouds. Despite strong performance, tracking-based approaches rely on an initial single-view mask and sequential mask propagation, making them prone to error accumulation. We argue that stable multi-view mask association should be grounded in geometric consistency rather than purely frame-to-frame propagation. More importantly, we find that SAM2's predicted IoU score is positively correlated with 3D mask quality, indicating that SAM2 implicitly knows which masks are more reliable for 3D clustering. However, it only reflects individual mask quality and not cross-view correspondences. Based on this insight, we present SAM2Scene, a 3D instance segmentation framework that leverages the predicted IoU score to guide multi-view mask association and 3D clustering. First, we propose Multi-View Proposal Association, where 3D projection establishes cross-view correspondences and the predicted IoU score is used to rank and group proposals across views. By grounding mask association in geometry, this design reduces drift and leads to more stable cross-view mask consistency. Second, we introduce Adaptive-Fidelity Clustering, in which the predicted IoU score serves as an adaptive fidelity prior for graph construction and 3D clustering. This allows the clustering process to prioritize more reliable clusters, resulting in more coherent and accurate 3D instances. Comprehensive evaluations on ScanNetV2, ScanNet200, and ScanNet++ demonstrate the effectiveness of our method and show consistent improvements over previous methods. Our project page is at https://jihuaizhaohd.github.io/SAM2Scene.

Abstract:
Visual object tracking aims to continuously locate specific targets within sequential frames, evolving from single-modal methods to multi-modal ones. However, existing multi-modal trackers are typically designed for fixed modality combinations, requiring separate models for different inputs. This leads to a poor adaptability to missing or imperfect modalities, and limited generalization. To address these issues, we propose a novel unified framework called AnyTrack for object tracking with any modalities. Specifically, we design a Modality-aware Interaction Module (MIM) to facilitate dynamic interaction across diverse modalities. This module bridges modality discrepancies and aggregates temporal cues to maintain spatio-temporal consistency during cross-modal interaction. Furthermore, we introduce a Context Understanding Module (CUM) to establish spatial correspondence between visual features and target locations via global-local prompts. This module employs target-aware context modeling to enhance foreground-background discrimination for precise localization. Finally, to support the training and evaluation under diverse modalities, we extend existing multi-modal object tracking benchmarks by incorporating grayscale images, language descriptions, and audio clips. Extensive experiments with both complete and missing modality settings demonstrate that our AnyTrack achieves state-of-the-art performance, validating its effectiveness and flexibility. The source code is available at https://github.com/IdolLab/AnyTrack.

Abstract:
Group-based reinforcement learning methods for multimodal large language models typically rely on trajectory-level credit assignment that applies a single advantage to all tokens in a response. However, multimodal reasoning involves substantially higher perceptual uncertainty than text-only settings: the model must repeatedly re-examine visual information to verify intermediate interpretations, while different visual groundings can lead to divergent reasoning paths. Consequently, such uniform credit assignment is particularly inadequate and causes relative advantages to progressively degenerate toward zero. To address these challenges, we propose Multi-Branch Policy Optimization (MBPO), a tree-based framework that constructs reasoning trees at vision-language decision boundaries, enabling sibling branches to explore diverse visual hypotheses and assigning segment-level credit through branch-relative advantages. We further introduce a temporal replay buffer to reuse informative segments while controlling policy staleness. Experiments on several multimodal reasoning benchmarks show that MBPO outperforms representative baselines, improving both learning signal quality and optimization efficiency. The code is publicly available at https://github.com/ShuaiLyu0110/MBPO.

Abstract:
Multimodal large language models (MLLMs) increasingly rely on long chain-of-thought reasoning for complex tasks. However, as reasoning sequences lengthen, models may gradually rely less on visual evidence and more on accumulated textual context, leading to visual forgetting. Existing approaches do not directly constrain how visual evidence is used and maintained along the original reasoning trajectory, leaving long-context visual forgetting insufficiently addressed. To address this issue, we propose Remember-R1, a reinforcement learning framework that mitigates long-context visual forgetting by applying process-level supervision directly on the original reasoning trajectory. Specifically, Remember-R1 introduces rewards that encourage broader coverage of annotated visual evidence, sustained visual dependence in later reasoning steps, and greater focus on question-relevant image regions. Experiments across multiple model scales and diverse multimodal benchmarks demonstrate that Remember-R1 consistently improves reasoning performance. Additional analyses further show that it slows the decline of visual attention during generation, supporting its effectiveness in mitigating long-context visual forgetting. The code is available: https://github.com/Ch921-cell/Remember-R1.

Abstract:
Denoising generative models deliver high-fidelity generation but remain bottlenecked by inference latency due to the many iterative denoiser calls required during sampling. Training-free acceleration methods reduce latency by either sparsifying the model architecture or shortening the sampling trajectory. Current training-free acceleration methods are more complex than necessary: higher-order predictors amplify error under aggressive speedups, and architectural modifications hinder deployment. Beyond 2x acceleration, step skipping creates structural scarcity---at most one fresh evaluation per local window---leaving the computed output and its backward difference as the only causally grounded information. Based on this, we propose ZEUS, an acceleration method that predicts reduced denoiser evaluations using a second-order predictor, and stabilizes aggressive consecutive skipping with an interleaved scheme that avoids back-to-back extrapolations. ZEUS adds essentially zero overhead, no feature caches, and no architectural modifications, and it is compatible with different backbones, prediction objectives, and solver choices. Across image and video generation, ZEUS consistently improves the speed--fidelity performance over recent training-free baselines, achieving up to 3.2x end-to-end speedup while maintaining perceptual quality. The code is publicly available at \urlhttps://github.com/Ting-Justin-Jiang/ZEUS.

Abstract:
Glass surface detection (GSD) is critical for scene understanding and reconstruction, and yet remains challenging due to the transparency and reflectivity of glass surfaces. Existing GSD methods typically rely on 2D appearance cues, which may fail in geometrically ambiguous scenes. In this paper, we propose a paradigm shift: grounding GSD in 3D visual geometry to explicitly model the physical existence of glass surfaces. Our method first distills rich 3D priors from the visual geometry grounded transformer (VGGT) and generates glass-aware 3D representations. It then exploits multi-tasking learning with a novel glass detection head, consisting of two core modules: a Frequency Self-Attention Module (FSAM) that identifies glass-specific spectral features for glass surface localization, and a Geometry Grounding Block (GeGB) that selectively grounds 2D features in 3D geometry for glass surface segmentation. Extensive experiments demonstrate that our method achieves state-of-the-art performance across seven standard GSD benchmarks, generalizes well to video/multi-modal data, and substantially improves reconstruction in glass scenes. Code is available in \urlhttps://github.com/YT3DVision/VGGT_GLASS.

Authors: Huakang Chen (Northwestern Polytechnical University), Jingbin Hu (Northwestern Polytechnical University), Liumeng Xue (Nanjing University), Qirui Zhan (Northwestern Polytechnical University), Wenhao Li (Northwestern Polytechnical University), Guobin Ma (Northwestern Polytechnical University), Hanke Xie (Northwestern Polytechnical University), Dake Guo (Northwestern Polytechnical University), Linhan Ma (Northwestern Polytechnical University), Yuepeng Jiang (Northwestern Polytechnical University), Bengu Wu (Yutu Zhineng), Pengyuan Xie (Lingguang Zhaxian Technology), Chuan Xie (Lingguang Zhaxian Technology), Qiang Zhang (Lingguang Zhaxian Technology), Lei Xie (Northwestern Polytechnical University)

Abstract:
Instruction-following text-to-speech (TTS) has emerged as an important capability for controllable and expressive speech generation, yet its evaluation remains underdeveloped due to limited benchmark coverage, weak diagnostic granularity, and insufficient multilingual support. We present MINT-Bench, a comprehensive multilingual benchmark for instruction-following TTS. MINT-Bench is built upon a hierarchical multi-axis taxonomy, a scalable multi-stage data construction pipeline, and a hierarchical hybrid evaluation protocol that jointly assesses content consistency, instruction following, and perceptual quality. Experiments across ten languages show that current systems remain far from solved: frontier commercial systems lead overall, while leading open-source models become highly competitive and can even outperform commercial counterparts in localized settings such as Chinese. The benchmark further reveals that harder compositional and paralinguistic controls remain major bottlenecks for current systems. We release MINT-Bench together with the data construction and evaluation toolkit to support future research on controllable, multilingual, and diagnostically grounded TTS evaluation. The leaderboard and demo are available at \urlhttps://aslp-lab.github.io/MINT-Bench-Demo/.

Abstract:
Video Temporal Grounding (VTG) faces a cross-modal semantic gap that often leads to background features being incorrectly aligned with the query, while directly matching the query to moments results in insufficient discriminability and consistency of temporal semantics. To address this issue, we propose MLLM-Assisted Semantic-Relational Consistent Alignment (MASRA), a training-time MLLM-based optimization framework for VTG. MASRA leverages an MLLM during training to produce two forms of textual priors, namely event-level descriptions with temporal spans and clip-level captions, and instantiates two MLLM-assisted alignments. Event Semantic Temporal Alignment (ESTA) aligns temporal context with event semantics to explicitly strengthen the correspondence between semantics and temporal events and improve span-level separability. Local Relational Consistency Alignment (LRCA) constructs a textual relation matrix derived from clip-level captions and aligns it with the temporal feature similarity matrix in the model, enhancing temporal consistency while capturing local structural information. MASRA includes two simple supporting modules, semantic-guided enhancement and second-order relational attention, to better utilize the learned semantic context and relational structure. Moreover, we introduce Decoupled Alignment Interaction (DAI) with a context-aware codebook to adaptively absorb query-irrelevant semantics and alleviate the cross-modal gap. The MLLM is only invoked during training and is not used at inference. Extensive experiments show that MASRA outperforms existing methods, and ablation studies validate its effectiveness. The code is available at https://github.com/Evangelion09/MASRA.

Abstract:
Large Audio-Language Models (ALMs) have recently demonstrated remarkable capabilities in holistic audio understanding, yet they remain unreliable for temporal grounding, i.e., the task of pinpointing exactly when an event occurs within long-form audio. This limitation stems from two factors: training data dominated by clip-level supervision lacking precise timestamps, and benchmarks that fail to simulate real-world scenarios where short events are obscured by dense background sounds. In this paper, we introduce SpotSound, an audio language model designed for grounding audio events. SpotSound incorporates a novel training objective, specifically designed to suppress hallucinated timestamps for events absent from the input. Additionally, we present SpotSound-Bench, a challenging temporal grounding benchmark where target events occupy less than 10% of each clip, creating a rigorous 'needle-in-a-haystack' evaluation. Experiments demonstrate that SpotSound achieves state-of-the-art results on temporal grounding benchmarks while maintaining robust performance across general downstream audio-language tasks. Code, models and benchmark are released on https://loiesun.github.io/spotsound/

Abstract:
Electronic Theater Programs (ETPs) serve as critical promotional media in the performing arts, comprising a multi-page collection of heterogeneous visual assets such as theatrical posters, performance details, and character portraits. However, existing text-to-image paradigms struggle with such complex design tasks due to their inability to comprehend long-context narratives and maintain visual consistency across multiple distinct pages. To address this, we introduce ETPDesigner, a collaborative Multi-Agent framework that directly synthesizes high-quality ETPs from raw dramatic scripts. Emulating a professional design pipeline, our framework orchestrates specialized agents for semantic script analysis, core poster synthesis, functional background generation, and the stratified composition of character assets. Central to ETPDesigner is a global style anchor mechanism that extracts visual priors from the core poster to enforce strict aesthetic uniformity across all generated components. Furthermore, we elevate the ETP from a static publication to an interactive companion. By integrating portrait animation, customized speech synthesis, and persona-grounded Large Language Models (LLMs), our system enables users to engage in real-time, voice-enabled conversations with the generated virtual characters. To rigorously benchmark this task, we construct ETP-Pro, a domain-specific benchmark dataset of professional theater posters and high-quality character portraits. Extensive evaluations demonstrate our method's superiority in producing semantically faithful, aesthetically consistent, and highly interactive program sets. Project page is available at \urlhttps://umunmun.github.io/ETPDesigner.

Abstract:
Recent text-to-image models have improved global realism, but text rendering remains a persistent failure mode: images may look convincing overall, yet local typography often contains malformed glyphs, broken strokes, irregular spacing, and other artifacts that humans heavily penalize. We formulate Text-in-Image Quality Assessment (TIQA), a no-reference task that estimates a human-aligned perceptual quality score for detected text regions independently of semantic correctness. We also introduce two datasets. TIQA-Crops contains 120k text crops from 36k AI-generated images from 12 generators, with 10k mean-opinion-score (MOS) labels and 110k proxy labels for pretraining. TIQA-Images contains 1,500 text-heavy images from 10 recent generators, including proprietary systems, with paired overall-quality and text-quality subjective scores. We also propose an AI-generated No-reference Text-in-Image Quality Assessment (ANTIQA) model, a lightweight predictor with text-specific inductive biases. Across crop-level and image-level evaluations, ANTIQA achieves the best alignment with human judgments, reaching PLCC/SROCC of 0.942/0.935 on TIQA-Crops and 0.842/0.837 for text-quality MOS on unseen generators in TIQA-Images. These results establish perceptual text quality as a distinct evaluation target for modern text-to-image generation. The code and dataset are available at \urlhttps://github.com/koltsov-cmc/antiqa.

Abstract:
Tactile feedback is fundamental to Hand-Object Interaction (HOI), governing contact formation, force regulation, and stable manipulation, making it essential for achieving true human-like dexterous manipulation. Yet, current human-to-robot dexterous transfer pipelines primarily rely on kinematic trajectories, resulting in motion imitation without physically grounded interaction. To address this, we introduce TactiDex, a real-world tactile-guided benchmark specifically designed to move dexterous manipulation beyond kinematic mimicry toward contact-level human-likeness. TactiDex provides a comprehensive dataset that elegantly aligns whole-hand tactile signals with multi-granularity kinematic and object states, coupled with standardized evaluation metrics. Building upon this data paradigm, we propose a tactile-driven transfer framework that effectively translates human demonstrations into physically plausible robotic execution. We introduce TactiSkill, a framework built upon a novel tri-component tactile reward that innovatively uses tactile signals as structured supervision. This reward unifies guidance, human-like alignment, and contact constraints into a single objective. Through comprehensive experiments on both single and bimanual tasks, we demonstrate that TactiSkill achieves superior performance in manipulation success and physical realism. This work lays a crucial foundation for advancing tactile-aware dexterous manipulation. Our project page at \urlhttps://tactidex.github.io/.

Abstract:
Generating spatially coherent stereo audio from text is an important yet underexplored problem in multi-modal generation. While text-to-audio systems have achieved strong performance, conventional neural codecs and generative frameworks often struggle to preserve intricate inter-channel dependencies for stereo signals, resulting in collapsed soundstage. In this paper, we present a unified framework for text-to-stereo audio generation. We first introduce StereoCodec, a discrete audio codec that jointly encodes dual-channel signals. By incorporating psychoacoustic-informed reconstruction objectives and dedicated architecture designs, StereoCodec significantly improves spatial fidelity while maintaining high reconstruction quality. Building upon these discrete representations, we propose StereoGen based on masked generative modeling. It utilizes a hybrid transformer combining multi-modal and single-modal blocks to ensure robust cross-modal alignment and global spatial-temporal coherence. To address the limitations of existing metrics in complex acoustic scenes, we further develop StereoCLAP, a robust evaluation framework that utilizes hard negative mining to measure semantic-spatial consistency. Extensive subjective and objective evaluations demonstrate that our approach consistently outperforms existing state-of-the-art methods across diverse scenarios, offering superior spatial accuracy and improved inference efficiency. Samples are availabel at https://stereogen-web.github.io/.

Abstract:
Fine-tuning vision foundation models (VFMs) has become the dominant paradigm for underwater instance segmentation (UIS), yet existing methods overlook the fact that underwater degradation further aggravates the confusion between target instances and visually cluttered backgrounds, a challenge that general adaptation strategies fail to address effectively. Although frequency-domain style alignment methods have shown some promise, they typically apply uniform operations across the entire spectrum. In contrast, we find that the effects of different degradations are concentrated in different frequency bands of the amplitude spectrum, making band-decomposed correction a more reasonable strategy. Based on this, we propose Band-Decomposed Adapter (BDA), a parameter-efficient fine-tuning method. Specifically, BDA employs Gaussian functions to partition the amplitude into multiple frequency bands and constructs an independent subspace for each band to perform targeted degradation correction. Furthermore, we design a dynamic routing mechanism that adaptively allocates the contribution of each frequency band according to the global amplitude distribution, thereby enabling robust handling of mixed degradations. Extensive experiments on UIIS and USIS10K show that BDA consistently outperforms existing state-of-the-art methods, validating the effectiveness of band-decomposed frequency-domain adaptation for UIS. Code is available at \urlhttps://github.com/Marinus47/BDA.

Abstract:
Graphics Program Synthesis is pivotal for interpreting and editing visual data, effectively facilitating the reverse-engineering of static visuals into editable TikZ code. While TikZ is the de facto standard for scientific schematics due to its programmatic flexibility, its requirement for rigorous spatial precision presents a significant challenge for Multimodal Large Language Models. Progress is currently stifled by two primary gaps: (1) Data Quality Gap: existing image-TikZ corpora often lack strict executability and reliable visual alignment; (2) Evaluation Gap: a lack of benchmarks for both structural and visual fidelity. To address these, we present a closed-loop framework featuring: SciTikZ-230K, a large-scale, high-quality dataset from our Execution-Centric Data Engine covering 11 diverse scientific disciplines; SciTikZ-Bench, a multifaceted benchmark spanning from basic geometric constructs to intricate hierarchical schematics to evaluate both visual fidelity and structural logic. To further broaden the scope of visual-code optimization methodology, we introduce a novel Dual Self-Consistency Reinforcement Learning optimization paradigm, which utilizes Round-Trip Verification to penalize degenerate code and boost overall self-consistency. Empowered by these, our trained model SciTikZer-8B achieves state-of-the-art performance, consistently outperforming proprietary giants like Gemini-2.5-Pro and massive models like Qwen3-VL-235B-A22B-Instruct. Our project is available at https://github.com/JackieForest/SciTikZ.

Abstract:
Recent VLM-based agents aim to replicate OpenAI O3's "thinking with images" via tool use, yet most open-source methods lack tool-calling capability for multi-image inputs, limiting their applicability to real-world multi-image QA tasks. To address this gap, we propose IMAgent, an open-source visual agent trained with end-to-end reinforcement learning for fine-grained single/multi-image reasoning. During inference, VLMs tend to gradually neglect visual inputs; to mitigate this issue, we design two dedicated tools for visual reflection and confirmation, enabling the model to actively refocus attention on image content. Beyond that, we, for the first time, reveal how tool usage enhances agent performance from an attention perspective. Equipped with a carefully designed two-layer motion trajectory masking strategy and tool-use reward gain, IMAgent acquires an effective tool-use paradigm through pure reinforcement learning, eliminating the need for costly supervised fine-tuning data. To further unleash the inherent tool-usage potential of the base VLM and fill data gaps, we construct a challenging, visually enriched multi-image QA dataset via multi-agent system. Extensive experiments validate that IMAgent achieves SOTA performance across mainstream single and multi-image benchmarks, and our in-depth analysis offers actionable insights for the community. Code is available at: https://github.com/cq-dong/IMAgent.

Abstract:
Evaluating object removal in images and videos remains challenging because the task is inherently one-to-many, yet existing metrics frequently disagree with human perception. Full-reference metrics reward copy-paste behaviors over genuine erasure; no-reference metrics suffer from systematic biases such as favoring blurry results; and global temporal metrics are insensitive to localized artifacts within edited regions. To address these limitations, we propose RC (Removal Coherence), a pair of perception-aligned metrics: RC-S, which measures spatial coherence via sliding-window feature comparison between masked and background regions, and RC-T, which measures temporal consistency via distribution tracking within shared restored regions across adjacent frames. To validate RC and support community benchmarking, we further introduce \mboxPROVE-Bench, a two-tier real-world benchmark comprising \mboxPROVE-M, an 80-video paired dataset with motion augmentation, and \mboxPROVE-H, a 100-video challenging subset without ground truth. Together, RC metrics and PROVE-Bench form the PROVE (Perceptual RemOVal cohErence) evaluation framework for visual media. Experiments across diverse image and video benchmarks demonstrate that RC achieves substantially stronger alignment with human judgments than existing evaluation protocols. Project page: \urlhttps://xiaomi-research.github.io/prove/.

Abstract:
Large Multimodal Models (LMMs) such as LLaVA are typically trained with an autoregressive language modeling objective, providing only indirect supervision to visual tokens. This often yields weak internal visual representations and brittle behavior under distribution shift. Inspired by recent progress on latent denoising for learning high-quality visual tokenizers, we show that the same principle provides an effective form of visual supervision for improving internal visual feature alignment and multimodal understanding in LMMs. We propose a latent denoising framework that corrupts projected visual tokens using a saliency-aware mixture of masking and Gaussian noising. The LMM is trained to denoise these corrupted tokens by recovering clean teacher patch features from hidden states at a selected intermediate LLM layer using a decoder. To prevent representation collapse, our framework also preserves the teacher's intra-image similarity structure and applies intra-image contrastive patch distillation. During inference, corruption and auxiliary heads are disabled, introducing no additional inference-time overhead. Across a broad suite of standard multimodal benchmarks, our method consistently improves visual understanding and reasoning over strong baselines, and yields clear gains on compositional robustness benchmarks (e.g., NaturalBench). Moreover, under ImageNet-C-style non-adversarial common corruptions applied to benchmark images, our method maintains higher accuracy and exhibits reduced degradation at both moderate and severe corruption levels. Our code is available at \urlhttps://github.com/dhruvashp/latent-denoising-for-lmms.

Abstract:
Facial appearance editing powers popular applications like FaceApp and Photoshop. Generative Adversarial Networks (GANs) and 3D Morphable Models (3DMMs) have been widely used for facial editing. GANs can perform varied facial edits (e.g., changing hair color, hairstyle), but often produce unstable edits. 3DMMs produce stable edits, but can only alter pose and facial expression. Recently, text-guided diffusion models like Nano Banana have become popular for image editing. Text-guided models are a compelling alternative to GANs and 3DMMs since they can produce both stable and varied image edits. While text-guided models have been widely tested for whole-scene edits (e.g., ``make the woman play a guitar''), they have not been comprehensively tested for facial editing. We conducted the first large-scale evaluation (～1M images evaluated) of six popular text-guided models on a sequential facial editing task. We present Face-Edit-Attributes, the largest collection of 169 facial editing attributes focused on hair, accessories, and pose edits. We compared model performance using two popular celebrity face datasets: CelebA and CelebSET. Our results show that most models performed hair and accessory edits well, but struggled with editing pose. All models over-edit (e.g., changing hair color when asked only to change the hairstyle). We also evaluated demographic biases in each model. Our results show surprising biases in overediting: almost all models created more overedits for dark-skinned male faces and old faces. The code and data for our results (including our repository of ～ 1M images) can be accessed \hrefhttps://github.com/rahul1801/Face-Edit-Bench\textcolorbluehere. \endabstract

Abstract:
Gaze calibration establishes a mapping between raw data and the user's on-screen gaze coordinates. For practical deployment, calibration must be seamless and capable of generalizing across diverse users and environments to ensure reliable performance in unseen conditions, which is beyond current methods. To address this, we frame practical calibration as a generic coupling pair mining process, aimed at uncovering correspondences between raw gaze information and general video content. We first introduce a simulated pilot dataset to validate our hypothesis that smooth pursuits are most correlated with moving objects in the visual field and useful yet generic clues can be extracted for calibration. Building on this, we further propose a mining-by-searching method to extract useful pairs, which consists of trajectories from smooth pursuit and corresponding viewed objects, from real-world datasets seamlessly by effectively exploiting motion-related cues. Importantly, a mining-by-classification model trained on our simulated dataset learns to recognize valid coupling pairs from motion-related cues, enabling deployment on unseen real-world datasets with unknown users without additional training. Finally, a per-person transformation is learned to bridge the gap between these pairs. Evaluations on public real-world benchmarks as well as our pilot dataset show that our approaches, including two mining versions, are both generic and effective across various backbone architectures, consistently achieving lower angular error compared to SOTA baselines. Our code and pilot dataset will be released at https://github.com/BX-valor/Coupling-Pair-Mining.

Abstract:
Text-to-image person re-identification (TIPR) retrieves target persons using natural language descriptions. However, existing methods largely overlook resolution variance in real-world surveillance. They characterize cross-resolution TIPR through two coupled failure modes: Evidence Reliability Collapse (ERC), where degraded visual tokens become unreliable for grounding fine-grained text, and Ranking Distribution Drift (RDD), where mixed-resolution galleries distort similarity neighborhoods and destabilize retrieval rankings. To address this challenge, we propose Cross-Resolution Semantic Transfer (CRST), a CLIP-style framework with three modules: resolution-conditioned reasoning, text-guided refinement and CR-RDA. Resolution-conditioned reasoning estimates token reliability to suppress corrupted evidence. Text-guided refinement injects semantic priors to recover discriminative cues. CR-RDA transfers HR neighborhood geometry to stabilize LR ranking under mixed resolutions. Experiments on CUHK-PEDES, ICFG-PEDES, and RSTPReid show that CRST improves ultra-low-resolution Rank-1 and mAP on average by 5.7% and 5.3%, while stabilizing mixed-resolution retrieval without sacrificing high-resolution accuracy. The code is available at \hrefhttps://github.com/AKADOUQ/CRST-Cross-Resolution-Semantic-Transfer-for-Robust-Text-to-Image-Person-Retrieval \textcolorlinkbluehttps://github.com/AKADOUQ/CRST

Abstract:
Visual Place Recognition (VPR) demands representations robust to drastic environmental and viewpoint shifts. Existing aggregation paradigms either depend on extensive supervised training or rely on first-order pooling, often struggling to preserve structural correlations under extreme shifts or incurring high adaptation costs. In this work, we propose Riemannian Invariant Aggregation (RIA), a unified geometric framework that explicitly models second-order scene structure on the Symmetric Positive Definite (SPD) manifold. By modeling feature-space perturbations as congruence transformations, RIA uses geometry-aware mappings to embed covariance descriptors in a linearized Euclidean space while suppressing weak correlations. Across nine benchmarks, the training-free variant is competitive with supervised methods and obtains the highest Recall@1 among the compared methods on all four unstructured benchmarks. With fine-tuning, RIA obtains the highest Recall@1 among the compared methods on all five structured benchmarks. The source code is available in \hrefhttps://github.com/LeviiveLl/Beyond-First-Order-Learning-Riemannian-Geometries-for-Invariant-Visual-Place-Recognition\textcolorblueour GitHub repository.

Abstract:
Automatically generating interactive 3D indoor scenes from natural language is crucial for virtual reality, gaming, and embodied AI. However, existing LLM-based approaches often suffer from spatial errors and collisions, in part because common scene representationsraw coordinates or verbose codeare difficult for models to reason about 3D spatial relationships and physical constraints. We propose SpatialGrammar, a domain-specific language that represents gravity-aligned indoor layouts as BEV grid placements with deterministic compilation to valid 3D geometry, enabling verifiable constraint checking. Building on this representation, we develop (1) SG-Agent, a closed-loop system that uses compiler feedback to iteratively refine scenes and enforce collision constraints, and (2) SG-Mini, a 104M-parameter model trained on compiler-validated synthetic data. Across 159 test scenes spanning five scenarios of different complexity, SG-Agent improves spatial fidelity and physical plausibility over prior methods, while SG-Mini performs competitively against larger LLM-based baselines on single-shot generation scenarios. Our project page is available at \urlhttps://xgrids-3d.github.io/SpatialGrammar/.

Abstract:
Reinforcement learning (RL) with verifiable rewards (RLVR) has demonstrated the great potential of enhancing the reasoning abilities in multimodal large language models (MLLMs). However, the reliance on language-centric priors and expensive manual annotations prevents MLLMs' intrinsic visual understanding and scalable reward designs. In this work, we introduce SSL-R1, a generic self-supervised RL framework that derives verifiable rewards directly from images. To this end, we revisit self-supervised learning (SSL) in visual domains and reformulate widely-used SSL tasks into a set of verifiable visual puzzles for RL post-training, requiring neither human nor external model supervision. Training MLLMs on these tasks substantially improves their performance on multimodal understanding and reasoning benchmarks, highlighting the potential of leveraging vision-centric self-supervised tasks for MLLM post-training. We think this work will provide useful experience in devising effective self-supervised verifiable rewards to enable RL at scale. Project page: \urlhttps://github.com/Jiahao000/SSL-R1.

Abstract:
Camera-Incremental object Re-IDentification (CI-ReID) is challenging owing to the shortage of cross-camera annotations and the absence of historical features for joint optimization. To handle these challenges, this paper proposes a new CI-ReID framework named Hierarchical Optimization with bidirectionaL Association (HOLA). Rather than a monolithic memory bank in current works, HOLA introduces Camera-Specific Memory banks (CSMs) for every camera to preserve view-specific knowledge with negligible extra overhead. To acquire cross-camera labels, HOLA leverages the Bidirectional identity Association (BA) that dynamically aggregates historical identity prototypes in CSMs and performs bidirectional matching between aggregated prototypes and current prototypes. Compared to existing unidirectional strategy, BA selects prototype pairs with mutual consistency, thus providing cleaner labels for the proposed Hierarchical feature Optimization (HO). HO is designed to train the CI-ReID model by jointly exploiting incremental samples and CSMs. Instead of adopting conventional ReID training paradigms, HO simultaneously optimizes features across three progressive levels, i.e., inter-instance, intra-camera, and inter-camera. By explicitly learning both camera-specific discriminability and the global identity alignment, HO yields more effective optimization and better performance for the CI-ReID task. Extensive experiments on four public ReID datasets, i.e., Market-1501, VeRi-776, MSMT17, and SYSU-MM01, demonstrate the state-of-the-art performance of HOLA, e.g., achieving final mAP of 59.3% on Market-1501, significantly surpassing recent methods by 10+%. Code is released at: \urlhttps://github.com/NanH5837/HOLA.

Abstract:
Despite remarkable progress in visual understanding, Multimodal Large Language Models (MLLMs) remain prone to hallucinations when reasoning about spatial relationships, often producing judgments that contradict the true 3D structure of the scene. Though several existing works have proposed to mitigate hallucinations, our analysis indicates that they show limited effectiveness in spatial reasoning, as they fail to bridge the fundamental gap between 2D visual representations and 3D spatial reality. Based on this finding, we define hallucinations arising from insufficient spatial structure modeling as spatial reasoning hallucination, a subcategory of relation hallucination that existing mitigation methods fail to address. We further identify three typical scenarios where such hallucinations frequently occur: perspective effects, object orientation, and viewpoint changes. To this end, we propose Geo3R, a training-free, plug-and-play framework that incorporates geometric evidence and structured 3D reasoning to mitigate spatial reasoning hallucination. Experiments on three benchmarks, covering 18 tasks across all three scenarios, show that Geo3R substantially reduces spatial reasoning hallucination across diverse MLLMs without additional training, outperforming existing models and methods. Code is available at https://github.com/hiMyee/Geo3R-Mitigating-Spatial-Reasoning-Hallucination.

Abstract:
Urban socioeconomic indicator estimation from multimodal visual data is an important task for urban analytics. Existing approaches typically rely on supervised learning with large volumes of labeled data, often complemented by additional metadata, handcrafted features, or representations derived from pretrained models. Recent multimodal large language models have made zero-shot estimation from satellite and street view imagery increasingly feasible. While these models can produce free-form rationales, they are typically not explicitly grounded in structured evidence, internally verified for consistency, or designed to integrate complementary and conflicting information across multiple urban views. To address these limitations, we propose Graph-of-Evidence (GoE), a task-aware and verification-aware framework for zero-shot urban indicator estimation. Specifically, GoE follows a three-stage \emphGround--Verify--Graph paradigm: the framework first grounds satellite and street view observations into structured evidence, then performs verification-aware claim assessment to validate and repair evidence--claim relations, and finally constructs an evidence graph that organizes evidence items, claims, and latent urban factors for graph-based calibration. Extensive experiments on multiple urban indicator estimation tasks demonstrate that GoE achieves competitive zero-shot performance while substantially improving interpretability over direct end-to-end prediction. These results suggest that decoupling evidence construction from final estimation via verification-aware evidence graphs is a promising direction for reliable and explainable multimodal urban intelligence. The code and appendix are available at \urlhttps://github.com/axin1301/Graph-of-Evidence-for-socioeconomic-inference.

Abstract:
Audio-visual instance segmentation (AVIS) requires accurately identifying and tracking individual sounding objects with pixel-level masks. Existing methods struggle to match overlapping acoustic events with visual instances and handle asynchronous audio-visual dynamics. Therefore, two critical questions arise: how can a model establish precise correspondence between overlapping sound sources and visual instances, and how can a model maintain robust tracking when audio and visual signals are temporally misaligned?This paper proposes Hear to See (H2S), addressing these challenges through two mechanisms. The Acoustic-Semantic Projector (ASP) disentangles mixed audio and establishes hierarchical correspondence from semantic to spatial domains. The Asynchronous Dynamics Modulator (ADM) adaptively adjusts state transitions via audio-modulated Mamba, prioritizing current information during dynamic variations and maintaining continuity in stable periods.Experiments on AVISeg show H2S achieves SOTA performance, attaining 48.54 mAP with a COCO pretrained ResNet50 and surpassing the previous by 7.8%. The code will be open-sourced once the paper is accepted. The source code will be publicly available at https://github.com/leiyeliu/H2S.

Abstract:
Diffusion-based style transfer often suffers from semantic leakage and structural distortion due to indiscriminate style aggregation during denoising. We attribute this to dense, noisy matching between all content and style tokens. To address this, we propose One-to-One Token Distillation (OOTD), a simple yet effective training-free framework that formulates style transfer as a precise token distillation task. The core of OOTD is the establishment of a strict one-to-one sparse correspondence between content and style tokens. Specifically, we first compute the similarity between each content token and all style tokens to determine the best-matched style feature for each content position. Based on this correspondence, we directly manipulate style features to form a target representation aligned with the content structure. Subsequently, we introduce two complementary objectives: (1) the Token Distillation (TD) loss, which enforces feature-level consistency between generated and manipulated style features; and (2) the Correspondence-Aligned Gram (CAG) loss, which constrains generated features to align with the precise composition of manipulated style features, thereby ensuring superior style fidelity. Extensive experiments demonstrate that OOTD significantly outperforms state-of-the-art training-free methods, achieving superior content preservation and style fidelity by eliminating interference from semantically irrelevant style cues. Furthermore, the core idea can be extended to existing dense attention distillation frameworks, significantly boosting their structural preservation capability. Our code is available at \urlhttps://github.com/IIP-Lab-XDU/OOTD.

Abstract:
Video virtual try-on (VVT) aims to generate realistic videos of a person wearing a target garment. Recent methods leverage a keyframe-driven video generation paradigm to improve in-the-wild performance, yet they still rely on masks to localize try-on regions, making them vulnerable to large motions and severe occlusions. Although mask-free image-based try-on methods have shown promising results by leveraging large-scale pseudo data, extending this paradigm to videos remains difficult, as constructing video-level pseudo data is prohibitively expensive. Furthermore, coarse keyframe sampling and the scarcity of multi-view try-on data limit existing keyframe-driven methods in maintaining garment consistency and handling diverse try-on tasks. To address these challenges, we propose BooM-VVT, a mask-free VVT framework built upon the keyframe-driven paradigm. To achieve mask-free VVT, we introduce a multi-stage training strategy that leverages image-level pseudo data for mask-free localization learning, substantially reducing the need for costly video-level pseudo data. To improve garment consistency, we propose Garment-Sensitive Keyframe Sampling, which selects keyframes based on garment-relevant body regions to better capture garment appearance. We further introduce Frame-Shared 3D-RoPE to establish spatiotemporal correspondences between keyframes and target video frames for accurate garment-detail transfer. Finally, we construct OmniView, a large-scale multi-view try-on dataset to support reliable try-on video generation under complex camera viewpoints and diverse try-on tasks. Extensive experiments demonstrate that BooM-VVT achieves superior temporal consistency and garment fidelity over existing methods. Project page: \urlhttps://boomvvt.github.io/boomvvt.

Abstract:
Video Moment Retrieval (VMR) aims to localize temporal segments in videos that correspond to a natural language query, but typically assumes only a single matching moment for each query. This assumption does not always hold in real-world scenarios, where queries may correspond to multiple or no moments. Thus, we formulate Generalized Moment Retrieval (GMR), a unified setting that requires retrieving the complete set of relevant moments or predicting an empty set. To enable systematic study of GMR, we introduce Soccer-GMR, a large-scale benchmark of 22.1K query-moment pairs over 5.5K clips from 139 diverse matches, built on challenging soccer videos that reflect general GMR scenarios, with realistic negative and positive queries. The benchmark is constructed via a duration-flexible semi-automated pipeline with human verification, enabling scalable data generation while maintaining high annotation quality. We further design a unified evaluation protocol with complementary metrics tailored for null-set rejection, positive-query localization, and end-to-end GMR performance. Finally, we establish strong baselines across two modeling paradigms: a lightweight plug-and-play GMR adapter for discriminative VMR models, and a GMR-tailored GRPO reward for fine-tuning multimodal large language models (MLLMs). Extensive experiments show consistent gains across all metrics and expose key limitations of current methods, positioning GMR as a more realistic and challenging benchmark for video-language understanding.

Abstract:
All-in-one image restoration aims to recover clean images degraded by multiple corruption types using a single unified model. Existing methods typically rely on image-level prompts or shared guidance to handle diverse degradations. However, such a paradigm becomes inadequate when degradations are spatially heterogeneous or even coexist in mixed forms within a single image. Yet spatially adaptive guidance alone is not sufficient, since accurate restoration also requires each spatial query to reliably aggregate complementary information from local neighborhoods and global contexts. To this end, we propose QuReC, a unified framework for all-in-one image restoration. QuReC consists of a Degradation-Guided Query Reconstruction Module (DQRM) and a Local-Global Response Calibration Module (LGRCM). Specifically, DQRM matches each spatial query against a degradation prototype space to reconstruct a query-specific degradation-aware representation, thereby providing fine-grained spatially adaptive restoration guidance. To further stabilize this query-wise matching process, we introduce a weakly supervised prototype matching learning strategy to improve optimization stability and degradation semantic consistency. Meanwhile, LGRCM performs local-global dual-branch aggregation and calibrates the aggregated responses with learnable priors, improving the reliability of feature aggregation and the coordination between local detail modeling and global context modeling. Extensive experiments demonstrate that QuReC achieves superior performance on multiple all-in-one image restoration benchmarks. The code is released at \textcolorpurple\urlhttps://github.com/zhoushen1/QuReC.

Abstract:
3D vision-language models (3D VLMs) enable spatial reasoning over multi-view scenes but suffer from substantial token redundancy due to duplicated observations and large uninformative regions, leading to high computational cost. Although visual token compression has shown promise in accelerating 2D VLMs, it fails to capture the structured nature of 3D scenes and leads to incomplete spatial coverage and loss of fine-grained details. In this paper, we propose HiSC, a training-free framework for hierarchical spatial clustering token compression in 3D VLMs. HiSC lifts token compression from token-level selection to cluster-level processing by organizing tokens into spatially grounded clusters using joint geometric and semantic cues. Specifically, we first introduce a spatial graph-based merging (SGraM) strategy that models cross-view redundancy as spatial connectivity and consolidates physically consistent regions, effectively merging extremely similar redundant tokens prior to LLM inference. We then propose a spatial clustering-based pruning (SCluP) paradigm within LLM inference, which performs hierarchical compression across clusters and within clusters, preserving object instance completeness while retaining fine-grained details for important regions. Extensive experiments on diverse 3D reasoning benchmarks show validate the effectiveness of HiSC, particularly under high visual token pruning ratios. Besides, HiSC achieves over 90% token reduction with minimal performance degradation. Code is accessible at \urlhttps://github.com/elecreak/HiSC.

Abstract:
Graphic design generation demands a delicate balance between high visual fidelity and fine-grained structural editability. However, existing approaches typically bifurcate into either non-editable raster image synthesis or abstract layout generation devoid of visual content. Recent combinations of these two approaches attempt to bridge this gap but often suffer from rigid composition schemas and unresolvable visual dissonances (e.g., text-background conflicts) due to their inexpressive representation and open-loop nature. To address these challenges, we propose DesignAsCode, a novel framework that reimagines graphic design as a programmatic synthesis task using HTML/CSS. Specifically, we introduce a Plan-Implement-Reflect pipeline, incorporating a Semantic Planner to construct dynamic, variable-depth element hierarchies and a Visual-Aware Reflection mechanism that optimizes the code to rectify rendering artifacts. Extensive experiments demonstrate that DesignAsCode significantly outperforms baselines in both structural validity and aesthetic quality. Furthermore, our code-native representation unlocks advanced capabilities, including automatic layout retargeting, complex document generation (e.g., resumes), and CSS-based animation. Our project page is available at \urlhttps://liuziyuan1109.github.io/design-as-code/.

Abstract:
Personalized Image Aesthetic Assessment (PIAA) aims to predict aesthetic ratings of images that vary across individuals. The aesthetic preferences manifest to different extents across distinct visual stimuli and exhibit cohort-specific patterns. Motivated by the above fact, this paper presents a Multimodal Large Language Model (MLLM)-based approach, which models individual aesthetic preferences by Preference-Rich sample mining and Aesthetically-resonant Cohort merging (PRAC). Specifically, PRAC first identifies preference-rich samples by analyzing both Collective Controversy and Personalized Deviation of images, maximizing the utility of limited user data. Based upon the preference-rich samples, cross-user preference similarities are measured by comparing preference embeddings. Then, a cohort-based model merging strategy, is proposed by aggregating preference patterns from aesthetically-resonant users, which further enhances the personalization for the target individual. Extensive experiments and comparisons demonstrate the superiority of the proposed method. The project is at \hrefhttps://github.com/yzc-ippl/PRAC\textcolorblueWebsite.

Abstract:
Vision-Language-Action (VLA) models map visual observations and natural-language instructions to robot actions; however, hierarchical and autoregressive paradigms often incur architectural overhead, accumulate long-horizon errors, and require auxiliary modules to capture environment dynamics. To this end, we present MMaDA-VLA, a fully native, pretrained discrete diffusion VLA that unifies multi-modal understanding and generation. Specifically, MMaDA-VLA uses a shared discrete token space to jointly denoise a future goal observation and an action chunk, grounding actions in predicted visual outcomes without an auxiliary world model. In this way, parallel, order-free refinement improves long-horizon consistency. Extensive experiments and comprehensive analyses demonstrate that MMaDA-VLA achieves an average success rate of 98.0% on LIBERO and an average successful sequence length of 4.78 on CALVIN, while performing strongly in real-world settings. The project page is available at \urlhttps://yliu-cs.github.io/MMaDA-VLA.

Abstract:
Diffusion models have demonstrated remarkable effectiveness in image restoration tasks. However, when guiding image reconstruction, existing Diffusion Model-based Image Restoration (DMIR) methods typically rely on fixed data constraints and uniform step sizes, thereby overlooking the dynamic nature of the generative process. Such rigid designs render the models vulnerable to spatially non-uniform degradations, thus resulting in structural distortions and loss of fine details. Meanwhile, uniform step sizes introduce computational redundancy, whereas nave step reduction strategies tend to accumulate approximation errors. To address these limitations, we propose a Local Epistemic Uncertainty Guided Active Sampling framework (LEADer). In the spatial domain, LEADer leverages pixel-wise uncertainty to dynamically modulate the prior strength within the null space, which effectively balances detail preservation and artifact suppression. In the temporal domain, it quantifies sampling stability via the uncertainty trace to enable adaptive trajectory pruning, thereby accelerating convergence. Theoretical proofs demonstrate that our framework achieves strict data consistency, while the trajectory pruning strategy admits a deterministic error bound, thereby guaranteeing stable convergence under skip sampling. Notably, our plug-and-play method can be seamlessly integrated into various DMIR baselines. Extensive experiments show that LEADer improves the performance of multiple state-of-the-art DMIR methods, while significantly reducing sampling time with negligible memory overhead. Code is available at \hrefhttps://github.com/JiaqiZhang-Sengoku/LEADerhere.

Abstract:
Few-shot class-incremental learning (FSCIL) aims to learn novel classes from limited annotations while preserving prior knowledge. Existing methods typically assume a sufficiently large base session, but this assumption fails when both base and incremental data are scarce, leading to weak initial representations, semantic drift, and unstable boundaries. We study this underexplored yet realistic setting, termed Generalizable FSCIL (G-FSCIL), where the base session itself contains only a few classes. Although synthetic data can alleviate supervision scarcity, naively mixing generated samples often introduces semantic noise and exacerbates oldnew boundary conflicts. To address this, we propose a framework that curates trustworthy synthetic knowledge for stable G-FSCIL. Specifically, we first construct class-specific synthetic candidate pools using a frozen latent diffusion model, where class inversion is performed at the first observation and the resulting condition embeddings are reused for on-demand generation. Building on these candidates, we learn a knowledge curation strategy that selects samples with both semantic consistency and visual diversity, and distill this process into a transferable selection policy during the base session, which is then reused without further optimization. Leveraging the curated synthetic data, we further design a boundary-stable incremental adaptation scheme, including synthetic-informed prototype initialization and bidirectional boundary calibration to mitigate oldnew conflicts. Extensive experiments demonstrate that our method consistently outperforms existing FSCIL baselines, with reduced forgetting and improved balance between old and new classes. Code is available at \urlhttps://github.com/NiHaoWoJiaoYYC/G-FSCIL.

Abstract:
Multilingual Text-Based Person Retrieval (TBPR) aims to retrieve individuals with texts in multiple languages. Despite its practical importance, it has been overlooked by existing research, which mainly focuses on English-only retrieval. The only prior attempt at multilingual TBPR suffers from (1) the absence of a large-scale, multilingual pretraining dataset and (2) a suboptimal model architecture for multilingual text encoding and cross-modal alignment, which employs a shared encoder that overlooks linguistic heterogeneity and relies on English as a pivot for cross-modal alignment, introducing English-centric bias. For this, we contribute MultiPedes, the first large-scale, pedestrian-centric multilingual pretraining dataset. It contains 10 million imagetext pairs across four languages, constructed by a multilingual quality-aware data construction pipeline. We also propose ViSeLA (Vision-grounded Semantic Learning and Alignment), a novel framework featuring: (1) a semantic-aware multilingual learning module that extends the mixture-of-experts architecture by routing language-invariant semantics to specialized experts for improved multilingual feature learning; and (2) a vision-grounded cross-modal alignment module that employs a semantic probe to capture fine-grained semantics across modalities, enabling direct cross-modal alignment without relying on a pivot language. Extensive experiments on both multilingual and English-only TBPR datasets demonstrate the effectiveness of ViSeLA. The dataset and code are available at https://github.com/Flame-Chasers/ViSeLA.

Abstract:
Real-world multimodal clinical prediction must operate under variable input reliability: medical images may be degraded by acquisition artifacts, physiological signals by noise, and entire modalities may be unavailable at inference time. Existing multimodal fusion methods typically handle quality degradation and missing modalities with separate mechanisms, and learned fusion gates do not explicitly preserve the ordering of interpretable reliability cues in the resulting allocation weights. To address this gap, we propose HyRAF, a reliability-aware multimodal fusion framework in hyperbolic space. HyRAF leverages two geometric cues on the Lorentz model, namely the radial distance from the origin and the angular deviation from a modality-specific exponential moving average (EMA) anchor, to estimate sample-level modality reliability without quality-specific supervision. At its core, Geometric Reliability Weighting (GeoWeight) enforces a monotonic mapping from these cues to per-modality allocation weights and uses cross-modal agreement to estimate overall prediction confidence. HyRAF unifies corrupted and missing modalities through reliability-aware logit fusion with availability masking, together with a confidence-gated hyperbolic prototype branch. Across three diverse clinical multimodal benchmarks, HyRAF achieves strong clean-condition performance and consistently improves robustness to modality corruption and absence, with up to a 30% reduction in average degradation under corruption relative to the strongest baseline. Code is available at \urlhttps://github.com/Cat-Aotian/HyRAF.

Authors: Tianxin Xie (The Hong Kong University of Science and Technology (Guangzhou)), Wentao Lei (The Hong Kong University of Science and Technology (Guangzhou)), Kai Jiang (The Hong Kong University of Science and Technology (Guangzhou)), Guanjie Huang (The Hong Kong University of Science and Technology (Guangzhou)), Pengfei Zhang (The Hong Kong University of Science and Technology (Guangzhou)), Chunhui Zhang (The Hong Kong University of Science and Technology (Guangzhou)), Fengji Ma (The Hong Kong University of Science and Technology (Guangzhou)), Haoyu He (The Hong Kong University of Science and Technology (Guangzhou)), Han Zhang (The Hong Kong University of Science and Technology (Guangzhou)), Jiangshan He (The Hong Kong University of Science and Technology (Guangzhou)), Jinting Wang (The Hong Kong University of Science and Technology (Guangzhou)), Linghan Fang (The Hong Kong University of Science and Technology (Guangzhou)), Lufei Gao (The Hong Kong University of Science and Technology (Guangzhou)), Orkesh Ablet (The Hong Kong University of Science and Technology (Guangzhou)), Peihua Zhang (Tencent), Ruolin Hu (The Hong Kong University of Science and Technology (Guangzhou)), Shengyu Li (The Hong Kong University of Science and Technology (Guangzhou)), Weilin Lin (The Hong Kong University of Science and Technology (Guangzhou)), Xiaoyang Feng (The Hong Kong University of Science and Technology (Guangzhou)), Xinyue Yang (The Hong Kong University of Science and Technology (Guangzhou)), Yan Rong (The Hong Kong University of Science and Technology (Guangzhou)), Yanyun Wang (The Hong Kong University of Science and Technology (Guangzhou)), Zihang Shao (The Hong Kong University of Science and Technology (Guangzhou)), Zelin Zhao (The Hong Kong University of Science and Technology (Guangzhou)), Chenxing Li (Tencent), Shan Yang (Tencent), Wenfu Wang (Tencent), Meng Yu (Tencent), Dong Yu (Tencent), Li Liu (The Hong Kong University of Science and Technology (Guangzhou))

Abstract:
Text-to-audio-video (T2AV) generation is central to applications such as filmmaking and world modeling. However, current models often fail to produce physically plausible sounds. Previous benchmarks in this area primarily focus on audio-video temporal synchronization, while largely overlooking explicit evaluation of audio-physics grounding, thereby limiting the study of physically plausible audio-visual generation. To address this issue, we present PhyAVBench, the first benchmark designed to systematically evaluate the audio-physics grounding capabilities of T2AV, image-to-audio-video (I2AV), and video-to-audio (V2A) models. PhyAVBench offers PhyAV-Sound-11K, a new dataset of 25.5 hours of 11,605 audible videos collected from 184 participants to ensure diversity and avoid data leakage. It contains 337 paired-prompt groups with controlled physical variations that drive sound differences, each grounded with an average of 17 videos and spanning 6 audio-physics dimensions and 41 fine-grained test points, from basic phenomena (e.g., collision) to complex effects (e.g., Helmholtz resonance). Each video includes step-by-step audio-physics reasoning, and each prompt pair is annotated with the physical factors underlying their acoustic differences. Importantly, unlike prior benchmarks that cannot measure sensitivity to underlying acoustic conditions, PhyAVBench leverages paired text prompts to evaluate this capability. We term this evaluation paradigm the Audio-Physics Sensitivity Test (APST) and introduce a novel metric, the Contrastive Physical Response Score (CPRS), which quantifies the acoustic consistency between generated videos and their real-world counterparts. We conduct a comprehensive evaluation of 17 state-of-the-art (SOTA) models across T2AV, I2AV, and V2A tasks, along with human studies involving 74 participants, which show a strong positive correlation with the CPRS metric. Our results reveal that even leading commercial models struggle with fundamental audio-physical phenomena, exposing a critical gap beyond audio-visual synchronization and pointing to future research directions. We hope PhyAVBench will serve as a foundation for advancing physically grounded audio-visual generation. Code and datasets are available at https://github.com/imxtx/PhyAVBench.

Abstract:
Propagation structures provide crucial evidence for fake news detection, yet existing approaches primarily rely on supervised GNN-based models, which require substantial labeled data and exhibit limited generalization. Although large language models (LLMs) exhibit strong reasoning capabilities, directly feeding them raw propagation graphs creates a significant modality mismatch and severe information overload, making structure-aware reasoning unreliable in zero-shot and few-shot settings. To bridge this gap, we propose MAGER, a multi-agent genetic evolution framework that automatically discovers meta-paths optimized for LLM reasoning. By compressing complex propagation graphs into informative subgraphs, the evolved meta-paths alleviate both information overload and modality mismatch, enabling frozen LLMs to perform structure-aware veracity reasoning. We further introduce a graph in-context learning strategy that retrieves semantically and structurally similar demonstrations to strengthen classification and reasoning. Extensive experiments show that MAGER substantially improves frozen LLMs as standalone fake news detectors in data-efficient settings. Our code is available at \urlhttps://github.com/SenticNet/MAGER.

Abstract:
World models provide a principled framework for modeling latent dynamics in embodied agents. However, existing approaches typically rely on a unified latent representation that assumes single temporal transition, which may be insufficient for target-centric embodied settings where heterogeneous dynamics coexist. In such scenarios, agent motion induces rapid global scene changes, while task-relevant entities require temporally consistent representations over time. We propose a bipartite world modeling framework that structures the latent state into an egocentric component modeling scene evolution and an allocentric component maintaining temporally coherent target representation. These two states are jointly inferred and evolve through differentiated transition mechanisms within a unified world model. We evaluate our approach on embodied target tracking benchmarks, a representative setting where such heterogeneous dynamics are prominent. Experiments on three benchmarks demonstrate state-of-the-art performance, along with effective real-world deployment on a quadruped robot. Code is available at https://github.com/JahaoZhang/Bipartite-World-Modeling-for-EVT.

Abstract:
Thermal infrared (TIR) image super-resolution is indispensable for robust perception under adverse conditions, yet its efficacy is fundamentally bottlenecked by hardware-intrinsic distortions. While recent generative models leverage large-scale priors for TIR image super-resolution, they often overlook the physical characteristics and degradation mechanisms of TIR imaging. In particular, they do not explicitly model sensor-dependent degradation well, making it difficult to disentangle scene-independent non-uniformity from scene radiance, while their reliance on heuristic, scene-dependent conditioning limits their ability to handle the complex degradations associated with physical sensor states. To address this issue, we propose a physics-informed generative model for TIR image super-resolution. This model includes an imaging-inspired disentanglement scheme that uses auxiliary short-term observations to decouple scene radiance prior together with sensor-state maps that characterize multiplicative and additive non-uniformity. These estimates are then used to initialize generative sampling and guide adaptive feature modulation during fine-tuning. We further introduce a sensor-routed spatial-spectral mixture of adapters, which adaptively adjusts feature modulation according to the different degradation strength. Experiments on multiple benchmark datasets under different degradation levels show that our method consistently outperforms state-of-the-art approaches and effectively boosts performance on downstream TIR perception tasks. The code is available at \urlhttps://github.com/bzHunter/DECO.

Abstract:
Existing audio-to-score (A2S) systems primarily focus on classical music, and the application to popular music remains underexplored. This paper first presents the new SheetSage-A2S Dataset, which includes 61 hours of audio with \textttkern score encodings for 9,468 clips originating from 6,066 unique songs, the first of its kind to facilitate A2S research for popular music. Additionally, we improve on existing A2S approaches by using data augmentation and MuQ, a pretrained feature-extraction model for music audio, to enhance generalisation abilities and extract meaningful audio features. Our results show that the proposed A2S model achieves 4.98% symbol error rate (SER) on the Quartets collection for classical music, which significantly outperforms the 15.3% SER from the existing state-of-the-art \citealfaro-contrerasTransformer2024. Additionally, our model achieves 20.92% SER on the SheetSage-A2S dataset for popular music, serving as a strong benchmark for future research. The dataset, model, and code are made publicly available at: \urlhttps://github.com/Multimodal-Music-Research-Lab/SheetSage2Kern_model.

Abstract:
Recent advances in large language models and multimodal models have pushed remote sensing (RS) processing from simple perception models to agentic systems designed to tackle complex, long-horizon RS tasks. However, existing systems often rely on monolithic decision-making frameworks, which fail to accommodate the multi-stage, interdependent nature of RS tasks. This centralized approach leads to challenges such as unstable task execution, incorrect tool usage, and error propagation across stages. To address these issues, we propose HiRS-Agent, a hierarchical multi-agent system for long-horizon RS task solving. HiRS-Agent adopts a two-level collaborative architecture: the Manager Layer handles dynamic routing, step-level verification, replanning, and termination control, while the Specialist Layer organizes domain-specific tools according to the RS workflow and is responsible for subtask reasoning and tool execution. To further enhance the system's capability, we introduce a two-stage supervised tuning strategy and a verification-guided hierarchical reinforcement learning stage to jointly optimize coordination and tool-use policies. Experiments on Earth-Agent Benchmark and ThinkGeo show that HiRS-Agent substantially improves long-horizon tool-use capability and final-task correctness, demonstrating the effectiveness of structured multi-agent collaboration for reliable RS agents. The code is publicly available at https://github.com/IntelliSensing/HiRS-Agent.

Abstract:
Trusted multi-view classification typically relies on a view-wise evidential fusion process: each view independently produces class evidence and uncertainty, and the final prediction is obtained by aggregating these independent opinions. While this design is modular and uncertainty-aware, it implicitly assumes that evidence from different views is numerically comparable. In practice, however, this assumption is fragile. Different views often differ in feature space, noise level, and semantic granularity, while independently trained branches are optimized only for prediction correctness, without any constraint enforcing cross-view consistency in evidence strength. As a result, the uncertainty used for fusion can be dominated by branch-specific scale bias rather than true sample-level reliability. To address this issue, we propose \underlineTrusted \underlineMulti-view learning with \underlineUnified \underlineRouting (TMUR), which decouples view-specific evidence extraction from fusion arbitration. TMUR uses view-private experts and one collaborative expert, and employs a unified router that observes the global multi-view context to generate sample-level expert weights. Soft load-balancing and diversity regularization further encourage balanced expert utilization and more discriminative expert specialization. We also provide theoretical analysis showing why independent evidential supervision does not identify a common cross-view evidence scale. Extensive experiments on 14 datasets and comparisons with 15 recent baselines demonstrate that TMUR consistently improves both classification performance and reliability. Code is available at \hrefhttps://github.com/YilinZhang107/TMUR\textcolordarkbluehere.

Abstract:
Infrared and visible image fusion (IVIF) combines robust infrared sensing with the rich structural details of visible imagery. Existing IVIF methods are predominantly developed in 2D image space, limiting them to fixed viewpoints and hindering explicit modeling of modality discrepancies at the scene level. Although recent works introduce multimodal cues into 3D representations, they focus primarily on cross-modality rendering conflicts or modality-dominant reconstruction rather than fusion-oriented selective preservation and suppression. We therefore propose Discrepancy-Aware Fusion with 3D Gaussian Splatting (DAF-GS), which learns robust 3D Gaussian representations through discrepancy-aware infrared and visible modality fusion. Specifically, a discrepancy-aware modality-guided cue mechanism estimates modality dominance from current-view responses and long-term modality dependency statistics, and incorporates the resulting cue into pixel-wise color rendering for explicit discrepancy-aware optimization. A primitive-wise detail compensation module further restores complementary fine-grained details weakened during selective suppression. Extensive experiments demonstrate the robustness of the learned representations and clear advantages over 2D fusion methods for novel-view synthesis across multiple datasets. The code is available at \urlhttps://github.com/RollingPlain/DAF-GS.

Abstract:
Audiovisual synchronization aims to accurately determine the temporal alignment between lip motions and the corresponding speech signal. However, existing approaches often struggle to achieve an optimal balance between in-sync and out-of-sync detection accuracy. In this paper, we provide a formal analysis of this performance bottleneck and rigorously show that both clip-level and frame-level synchronization models are inherently constrained by the alignmentuniformity trade-off on the representation hypersphere. To overcome this fundamental limitation, we propose HybridSyncNet, an efficient hybrid-level synchronization paradigm that leverages adaptive gating and multi-level contrastive learning to fuse the complementary strengths of clip-level and frame-level information in alignment and uniformity. Crucially, we theoretically demonstrate that, compared with single-level models, the proposed HybridSyncNet simultaneously achieves alignment improvement and uniformity gain. Extensive experiments validate the effectiveness of our approach. With only 2.46M parameters, HybridSyncNet outperforms state-of-the-art audiovisual synchronization models on the CREMA-D and HDTF benchmarks, establishing a new paradigm for audiovisual synchronization. The code is available at~\urlhttps://github.com/tlemangen/HybridSyncNet.

Abstract:
In K-12 educational scenarios, handwritten Chinese text recognition should not only transcribe student writing, but also detect faked characters. However, existing recognition models are usually confined to a predefined set of normal characters and therefore cannot explicitly identify faked characters. Existing detection methods exhibit complementary limitations: character-level methods provide interpretable structural evidence but suffer from low efficiency, whereas line-level methods are efficient but rely heavily on confidence scores, making them prone to missed detections and lacking explicit structural evidence. Thus, the key challenge is to preserve character-structural evidence independent of contextual inference while maintaining line-level efficiency. To this end, we propose DTRNet, a dual Text-Radical decoding framework for line-level faked character detection. DTRNet decouples context-aware text recognition from character-wise structural verification, where the text branch performs line-level transcription and the radical branch predicts legal Ideographic Description Sequences (IDS) for lexicon-based faked character judgment. We further introduce IDS-Guided Confidence Adjustment (IGCA) to refine text predictions using structural evidence during inference. Experimental results demonstrate that DTRNet effectively detects faked characters while maintaining strong recognition performance and providing interpretable radical-level evidence. Code, checkpoints, and the processed dataset are publicly available at https://github.com/BNU-ERC-ITEA/DTRNet.

Abstract:
In long-context decoding for LLMs and LMMs, attention becomes increasingly memory-bound because each decoding step must load a large amount of KV-cache data from GPU memory. Existing acceleration strategies often trade efficiency for accuracy by relying on heuristic pruning that may discard useful information. At a deeper level, they also tend to indiscriminately preserve all high-scoring tokens, treat early tokens as indispensable anchors, or rely on heuristic head routing, reflecting an insufficient mechanistic understanding of the attention sink phenomenon. In this paper, we show that attention sinks correspond to a model-encoded initial-position state with distinctive key geometry and low-impact value updates. Counterfactual anchor and head-output interventions further indicate that detected sink computations can be skipped with negligible output distortion. Based on this insight, we propose SinkRouter, a training-free selective routing framework that detects the sink signal and skips computations that would otherwise produce near-zero output. To translate this mechanism into real-world acceleration, we develop a hardware-aware Triton kernel with block-level branching and Split-K parallelism. We conduct extensive evaluations on a diverse suite of long-context benchmarks, including \textscLongBench, \textscInfiniteBench, \textscCVBench, \textscMileBench, and \textscMMVP, using both text-only and multimodal backbones such as Llama-3.1-8B, Llama-3.1-70B, Yi-9B-200K, LLaVA-1.5-7B, and LLaVA-1.5-13B. Across these settings, SinkRouter consistently improves decoding efficiency while maintaining competitive accuracy, and reaches 2.03× speedup with a 512K context. Code is available at \urlhttps://github.com/streamer-AP/SinkRouter.

Abstract:
Multi-shot video generation models can now produce visually coherent sequences, yet often fail to execute prescribed entity entrances and exits, revealing a fundamental consistencydynamics trade-off: mechanisms that preserve cross-shot appearance can resist the localized state changes required by dynamic narratives. Current evaluation paradigms emphasize visual quality and identity consistency, leaving entity-level narrative transitions under-measured by reducing them to frame-wise similarities. We present the Span (State Progression and Algebraic Narrative) framework, which represents multi-entity, multi-shot scripts as state transition matrices encoding entity entrances, exits, and holds. We prove that two generating patterns (Relay and Split) span the canonical integer transition-matrix space under character role swapping and integer linear combination. This establishes representation-level algebraic completeness, rather than empirical enumeration of every physically realizable narrative. To evaluate generated videos, we introduce Stem (State Transition Error Matching). Stem constructs an error matrix by comparing observed transitions with the prescribed script, then applies a sparsity-aware decomposition that reduces sensitivity to static-cell detector noise, yielding Prescribed Transition Match (PTM) for transition fidelity and Entity Presence Accuracy (EPA) for cast compliance. On MEMS-1K, a basis-derived benchmark of 1,000 scenarios across seven interpretable patterns, PTM discriminates four representative generation systems and yields an inverse ordering relative to Identity Consistency. This ordering is preserved under multi-frame aggregation and an alternative VLM detector, while absolute scores remain detector-dependent. The error matrix further induces a four-type taxonomy of entity-transition failures, providing structured diagnostics complementary to visual-quality, identity-consistency, and human-preference evaluation. https://merrybabyxmas.github.io/The-Algebra-of-Storytelling/

Abstract:
While Multimodal Large Language Models (MLLMs) are increasingly deployed in safety-critical domains, their reliability is threatened by multimodal implicit risks. Unlike explicit threats, these hazards emerge when individually benign text and neutral visual entities logically converge to induce unsafe outputs. Current detection methods fail to address this because they overlook the underlying risk activation mechanisms that govern cross-modal risk activation, leading to single-modality shortcut learning and hallucinated rationalizations. To bridge this gap, we first construct TriggerBench, the first dataset explicitly modeling risk compositionality (5,600 instances). By formally isolating Key Elements and Trigger Elements to build counterfactual contrastive pairs, TriggerBench eliminates risk residues and forces models to perform genuine logical deduction rather than superficial pattern matching, which provides a rigorous foundation for both large-scale training and fine-grained evaluation. Building on this, we propose a Step-Supervised Structured Reasoning training framework and employ it to train ThinkingGuard, a specialized guard model. Inspired by Situation Awareness theory, we decouple implicit risk identification into progressive cognitive stages, and utilize a step-reward Monte Carlo Tree Search algorithm to explore optimal reasoning trajectories, which are then distilled into the model through Dual-Constraint Preference Alignment. Extensive experiments across both standard and implicit safety benchmarks demonstrate that ThinkingGuard achieves strong performance. Project resources are available at https://github.com/FroggyChen/ThinkingGuard.

Abstract:
Multimodal Large Language Models (MLLMs) are increasingly deployed in safety-critical domains, yet their susceptibility to backdoor attacks remains a formidable challenge. While existing defenses have evolved from visual-centric patch cues to unsupervised internal attention statistics, they are fundamentally constrained by a reliance on static snapshots and attack-specific heuristics. Given the intricate interplay between dual modalities, the vast design space for multimodal triggers allows diverse threats, such as textual injections, sub-perceptual noises, and coordinated cross-modal triggers, to easily circumvent such rigid priors, particularly when attack types are unknown in real-world applications. In this paper, we rethink backdoor detection by shifting the diagnostic paradigm from static observations to dynamic output evolution. We observe that a successful attack inevitably hijacks the model's decoding trajectory, skewing internal modal contributions during generation while maintaining an abnormal output invariance despite input perturbations. To capture this behavior, we propose EVO-Detect, a unified detection framework that monitors the dynamic evolution of internal signals during early decoding steps through two complementary channels: \emphAttention Flow and \emphLogits Flow. The attention branch measures how image-text attention allocation evolves during early generation, whereas the logits branch quantifies the sensitivity of the output distribution to modal perturbations through interventional ablations. We then convert these trajectories into compact sample-level metrics and normalize them against negative-control signatures so that natural perturbations are less likely to be mistaken for backdoor evidence. Extensive evaluations show that our approach substantially outperforms state-of-the-art baselines such as BYE and TCAP across diverse attack families, establishing dynamic multimodal interplay analysis as a robust foundation for unsupervised backdoor detection of MLLMs. The code is publicly available at \urlhttps://github.com/kaoru0822-kitauji/EVO_Detect.

Abstract:
Multimodal Emotion Reasoning (MER) requires perceiving and integrating emotion cues from multiple modalities. Recent advances increasingly leverage Multimodal Large Language Models (MLLMs) to enhance reasoning capacity. However, due to the inherent heterogeneity in conveying emotions across modalities, these models often suffer from modality imbalance and confront flawed emotion understanding. In this paper, we conduct an in-depth analysis on MER models and uncover two key empirical insights. (i) In MER, modality imbalance manifests two distinct failure modes---unimodal over-reliance and insufficient cross-modal utilization, which call for different intervention strategies. (ii) MLLM layers exhibit implicit functional hierarchies in multimodal emotion processing, playing separable roles in stabilizing unimodal representations and integrating cross-modal information. Motivated by these observations, we propose a \textttLayer-\textttAware \textttModality \textttRebalancing (\textttLAMR) framework by explicitly aligning intervention strategies with layer roles. Specifically, we first introduce two quantized metrics for adaptive layer role profiling. In each layer, Modality Entropy (ME) measures the dispersion of modality contributions, and Cross-modal Intensity (CMI) captures the strength of cross-modal interactions. Then, we respectively enhance modality equilibrium in stabilization layers and cross-modal propagation in integration layers. Extensive experiments on two comprehensive MER benchmarks demonstrate that \textttLAMR effectively mitigates modality imbalance and improves emotion reasoning across diverse MLLM backbones, without additional training. Project url: \urlhttps://pengyuhuang1.github.io/LAMR/

Abstract:
In demanding professional environments and meeting review scenarios, lengthy text often imposes a high cognitive load. To facilitate efficient information communication, transforming verbose text into logically clear diagrams is essential. Scalable Vector Graphics (SVG) provide an effective representation for this purpose due to their editability and resolution independence. However, current research on Text-to-SVG generation remains hindered by three major challenges: (1) the scarcity of datasets for complex, logic-rich diagrams; (2) the absence of explicit layout priors, which leads to chaotic spatial arrangements; and (3) the lack of fine-grained visual feedback to validate rendered outputs and correct aesthetic defects. To address these challenges, at the data level, we introduce DocMeetSVG-100K, a large-scale SVG dataset tailored for document authoring and meeting review scenarios. At the model level, we propose GVR-Coder, a novel framework designed to generate high-quality logical diagrams from lengthy professional texts. Specifically, we adopt a curriculum-driven rejection sampling fine-tuning to progressively enhance the model's capability in modeling complex structures, while explicitly incorporating layout constraint knowledge during training. In addition, we introduce reinforcement learning from dual rendering feedback, a mechanism that provides implicit feedback through reward signals to jointly optimize structural complexity and visual aesthetics. Furthermore, we design a generate-verify-repair agent loop, which improves generation quality through explicit, fine-grained feedback and targeted refinement. Extensive experiments demonstrate that GVR-Coder outperforms competitive baselines and reliably produces logically coherent and visually appealing diagrams. Code, data, and the extended version are available at https://github.com/CurryaNa/GVR-Coder.

Abstract:
\beginabstract Multimodal Large Language Models (MLLMs) show strong visual perception, yet remain limited in reasoning about space under changing viewpoints. We study this challenge as Perspective-Conditioned Spatial Reasoning (PCSR) in 360^\circ omnidirectional images, where broad scene coverage reduces ambiguity from partial observations without eliminating the need for viewpoint-dependent inference. To assess this capability, we introduce PCSR-Bench, a diagnostic benchmark of 84,373 QA pairs from 2,600 omnidirectional images across 26 indoor environments, organized into eight tasks under three cognitive groups--- \emphPerception, \emphSpatial, and \emphadvanced PCSR. We evaluate 14 representative MLLMs and observe a substantial perception--reasoning gap: accuracy reaches 57.59% on Limited Field-of-View Reasoning (T7) but drops to 13.49%, 7.13%, and 0.64% on Relative Direction (T2), Egocentric Rotation (T4), and open-ended Compositional Directional Chains (T3), respectively. To probe the plasticity of this gap, we conduct an RL-based diagnostic study on a 7B-scale model. Reward shaping improves a matched 7B baseline from 31.10% to 60.06% under a controlled setting, suggesting that PCSR exhibits partial plasticity rather than being fully immutable. Still, these gains are task-selective, sensitive to reward design, and partially dependent on the evaluation protocol. These results position PCSR as a key bottleneck in current MLLMs and highlight meaningful yet bounded room for recovery under targeted optimization. Details and access are available at \urlhttps://github.com/Caleb-ychen/PCSR-Benchmark. \endabstract

Abstract:
Recent advances in visionlanguage models (VLMs) have enabled robots to follow instructions and demonstrate impressive commonsense reasoning. However, current visionlanguageaction (VLA) frameworks primarily rely on static representations and limited temporal context, restricting agents to short-horizon, reactive behaviors and hindering robust generalization in dynamic embodied environments. Inspired by cognitive neuroscience theories of episodic memory, we propose a functional episodic world model that integrates temporally extended interaction context, enabling embodied robots to integrate temporally extended interaction context and predict future dynamics. As an instantiation of this concept, our unified TriVLA realizes the episodic world model through a triple-system architecture: integrating multimodal grounding from a pretrained VLM (System 2) and temporally rich dynamics perception from a video diffusion model (System 3). This enables the agent to accumulate sequential experiences, interpret current contexts, and condition action generation on predicted future environmental evolution. Guided by episodic representations that span both the past and anticipated future, the downstream policy (System 1) generates coherent, context-aware action sequences through flow-matching and cross-modal attention mechanisms. Experimental results show that TriVLA operates efficiently at ~36 Hz and outperforms baseline models on standard benchmarks and real-world manipulation tasks. It demonstrates strong long-horizon manipulation and open-ended intent understanding, showcasing the advantages of episodic world model-inspired representations for generalizable robot control.

Abstract:
Physics-driven 4D dynamic simulation from static 3D scenes remains constrained by an overlooked contradiction: reliable motion supervision often relies on online video diffusion or optical-flow pipelines whose computational cost exceeds that of the simulator itself. Existing methods further simplify inverse physical modeling by optimizing only partial material parameters, limiting realism in scenes with complex materials and dynamics. We present Resonance4D, a physics-driven 4D dynamic simulation framework that couples 3D Gaussian Splatting with the Material Point Method through lightweight yet physically expressive supervision. Our key insight is that dynamic consistency can be enforced without dense temporal generation by jointly constraining motion in complementary domains. To this end, we introduce Dual-domain Motion Supervision (DMS), which combines spatial structural consistency for local deformation with frequency-domain spectral consistency for oscillatory and global dynamic patterns, substantially reducing training cost and memory overhead while preserving physically meaningful motion cues. To enable stable full-parameter physical recovery, we further combine zero-shot text-prompted segmentation with simulation-guided initialization to automatically decompose Gaussians into object-part-level regions and support joint optimization of full material parameters. Experiments on both synthetic and real scenes show that Resonance4D achieves strong physical fidelity and motion consistency while reducing peak GPU memory from over 35\,GB to around 20\,GB, enabling high-fidelity physics-driven 4D simulation on a single consumer-grade GPU.Code is available at https://github.com/CsZhang-Research/Resonance4D.

Abstract:
Existing Multimodal Large Language Models (MLLMs) for image forgery detection and localization predominantly operate under a text-centric Chain-of-Thought (CoT) paradigm. However, forcing these models to textually characterize imperceptible low-level tampering traces inevitably leads to hallucinations, as linguistic modalities are insufficient to capture such fine-grained pixel-level inconsistencies. To overcome this, we propose ForgeryVCR, a framework that incorporates a forensic toolbox to materialize imperceptible traces into explicit visual intermediates via Visual-Centric Reasoning. To enable efficient tool utilization, we introduce a Strategic Tool Learning post-training paradigm, encompassing gain-driven trajectory construction for Supervised Fine-Tuning (SFT) and subsequent Reinforcement Learning (RL) optimization guided by a tool utility reward. This paradigm empowers the MLLM to act as a proactive decision-maker, learning to spontaneously invoke multi-view reasoning paths including local zoom-in for fine-grained inspection and the analysis of invisible inconsistencies in compression history, noise residuals, and frequency domains. Extensive experiments reveal that ForgeryVCR achieves state-of-the-art (SOTA) performance in both detection and localization tasks, demonstrating superior generalization with minimal tool redundancy. The code is available at https://github.com/youqiwong/ForgeryVCR.

Abstract:
In multimodal crack segmentation for industrial facilities, the key challenge is preventing missing modalities from degrading pixel-level performance while maintaining low computational cost. Existing methods struggle to address semantic degradation caused by missing modalities. We propose Compass, a lightweight network for robust crack segmentation under arbitrary missing modalities. Compass comprises Degradation Simulation Distillation (DSD), Needle Block, and Evidential Topology-Preserving Fusion (ETPF). DSD constructs a degradation simulation stream that mimics more severe missing conditions and performs reciprocal distillation with the original stream, decoupling complete perception from degradation adaptation. Within DSD, Feature-Aware Prototype Transmitter (FAPT) performs modality agnostic prototype-guided feature completion to maintain semantic integrity under incomplete modality conditions. As a lightweight backbone, Needle injects crack-direction cues into WKV modulation and combines connectivity-aware gating with anisotropic context probing for structure-aware modeling. ETPF fuses multimodal features via Dempster-Shafer evidential combination with uncertainty-gated decoding, preserving crack topology while suppressing unreliable features. Experiments on three datasets demonstrate state-of-the-art (SOTA) performance under diverse missing modality scenarios. Even with 90% depth modality missing on CrackDepth, Compass achieves F1 of 0.8216 and mIoU of 0.8434 with only 2.58M parameters. The code is available at https://github.com/Karl1109/Compass.

Abstract:
Autoregressive vision-language models (VLMs) deliver strong multimodal capability, but their token-by-token decoding imposes a fundamental inference bottleneck. Diffusion VLMs offer a more parallel decoding paradigm, yet directly converting a pretrained autoregressive VLM into a large-block diffusion VLM (dVLM) often leads to substantial quality degradation. In this work, we present BARD, a simple and effective bridging framework that converts a pretrained autoregressive VLM into a same-architecture, decoding-efficient dVLM. Our approach combines progressive supervised block merging, which gradually enlarges the decoding block size, with stage-wise intra-dVLM distillation from a fixed small-block diffusion anchor to recover performance lost at larger blocks. We further incorporate a mixed-noise scheduler to improve robustness and token revision during denoising, and memory-friendly training for long multimodal sequences. A key empirical finding is that direct autoregressive-to-diffusion distillation is poorly aligned and can even hurt performance, whereas distillation within the diffusion regime is consistently effective. Experimental results show that, using at most 4.4M training samples, BARD-VL transfers strong multimodal capability from Qwen3-VL to a large-block dVLM. BARD-VL achieves state-of-the-art performance among comparable-scale open dVLMs on our evaluation suite at both 4B and 8B scales, while delivering up to 3× decoding throughput speedup compared with the source model. Code is available at: https://github.com/fudan-generative-vision/Bard-VL.

Abstract:
While multimodal large language models (MLLMs) enable a wide range of image-text reasoning tasks, recent incidents indicate that they are vulnerable to illicit deployment and unauthorized distillation. Existing solutions for model provenance are typically confounded by shared language backbones in MLLMs and struggle to detect violations of distillation. To bridge this gap and safeguard model ownership, we present the first study on multimodal model fingerprinting. Inspired by recent findings that self-attention acts as a low-pass filter and that its low-frequency components are informative, we develop AttnPrint for white-box provenance. Specifically, we extract cross-modal attention distributions and isolate their low-frequency components to serve as model fingerprints. To facilitate black-box auditing, we further introduce DistillTrace, which employs hypothesis testing of MLLM outputs to identify potential model infringement. We conduct extensive experiments on 154 model instances across 19 multimodal architectures. Notably, AttnPrint achieves strong derivative-model detection performance while remaining robust to five downstream modification techniques. DistillTrace also provides evidence of distillation relationships under three parameter-independent techniques.

Abstract:
Large Multimodal Models (LMMs) have achieved remarkable progress in generating photorealistic and prompt-aligned images, but they often produce outputs that contradict verifiable knowledge, especially when prompts involve fine-grained attributes or time-sensitive events. Conventional retrieval-augmented approaches attempt to address this issue by introducing external information, yet they are fundamentally incapable of grounding generation in accurate and evolving knowledge due to their reliance on static sources and shallow evidence integration. To bridge this gap, we introduce ORIG, an agentic open multimodal retrieval-augmented framework for Factual Image Generation (FIG), a new task that requires both visual realism and factual grounding. ORIG iteratively retrieves and filters multimodal evidence from the web and incrementally integrates the refined knowledge into enriched prompts to guide generation. To support systematic evaluation, we build FIG-Eval, a benchmark spanning ten entity classes across perceptual, compositional, and temporal dimensions. Experiments demonstrate that ORIG substantially improves factual consistency over strong baselines, highlighting the potential of open multimodal retrieval for factual image generation.

Abstract:
Text-driven 3D human motion generation models face significant challenges in responding to diverse and unconstrained textual prompts, primarily due to the limited availability of 3D motion training data. To address this, we introduce MoVT, a novel framework that effectively leverages the extensive range of human action videos to enhance text-to-motion generation. At the core of our approach is the cross-modal augmented motion tokenizer, which projects discrete 3D motion tokens into the 2D domain. This projection allows us to enrich the motion codebook with complex, real-world motion patterns derived from videos. The enriched discrete tokens are then mapped back to the 3D domain, resulting in aligned 3D and 2D codebooks with an enhanced capacity to represent intricate motions. These enhanced codebooks are integrated into a generative masked transformer, which predicts masked motion token indices in a modality-agnostic manner. This enables the use of text-index pairs, generated from the 2D codebook and annotated motion videos, to further enhance the generator. Extensive empirical evaluations show that MoVT performs favorably against prior state-of-the-art methods across multiple key metrics.

Abstract:
Video emotion recognition aims to infer human affect from dynamic facial videos, yet existing methods are challenged by two coupled factors: identity-affect entanglement and imperfect observability caused by occlusion, pose variation, and unreliable local evidence. To address these issues, we formulate dynamic facial expression recognition as learning kinematics-inspired affective dynamics over anchor-relative facial states, and propose PDADNet. First, we introduce a Self-supervised Deviation-to-Prototype Reconstruction module (SDPR), which learns anchor-referenced deviations and regularizes them with a part-wise prototype dictionary, suppressing identity-dominant appearance bias while inducing an expression-structured latent space. Second, we propose a Kinematics-inspired Spatio-Temporal Dynamics module (KSTD), which combines mask-guided spatial interaction, a Visibility-Aware Temporal Dynamics (VATD) operator, and bidirectional temporal modeling to capture structurally coherent affect dynamics under imperfect observability. A reliability-calibrated Temporal Reversibility Regularization (TRR) further constrains bidirectional temporal consistency on valid observations, reducing blind-update-induced drift and improving robustness to jitter, flicker, and partial occlusion. Experimental results on DFEW, FERV39k, and MAFW demonstrate that our PDADNet achieves state-of-the-art performance for video emotion recognition. The source code is in the Supplementary Material.

Abstract:
Despite recent advances in unified multimodal models for multi-reference image generation, existing benchmarks remain organized around predefined task types (e.g., ``subject composition''), which are ill-suited to this combinatorial setting and lead to fragmented coverage, uncontrolled complexity, and little diagnostic value. Recognizing that diverse multi-reference tasks share a common set of atomic operations, we adopt a capability-oriented perspective and formalize four operators: Anchor~(f), Disentangle~(g), Apply~(\oplus), and Compose~(C). Any multi-reference prompt can then be represented as a compositional formula over these operators, whose structural complexity is quantified by the number of operator slots. Building on this formulation, we construct TRACE-Bench, comprising approximately 1,600 evaluation cases across slot counts 1--8, built from 631 formula templates and around 4,000 reference images spanning diverse artistic styles and real-world subjects. The formula structure directly drives an operator-aligned evaluation protocol for per-capability scoring and a diagnostic tree analysis for recursive failure localization. Evaluating 9 leading models reveals insights invisible to holistic scoring: the primary bottleneck lies in disentanglement~(g) and attribute binding~(\oplus) rather than scene-level composition~(C), with even the best model scoring only 0.74 on attribute fidelity.

Abstract:
Multimodal Agentic Retrieval (MAR) requires agents to solve complex information-seeking tasks by iteratively invoking external tools. The prevailing frameworks maintain raw multimodal inputs and the accumulating interaction history in a single, ever-growing context, leading to the context explosion problem. While existing methods alleviate this issue by compressing redundant text, effective strategies for managing token-intensive visual content remain largely underexplored. To address this gap, we first conduct a systematic empirical study of approximately 10,000 trajectories. The results show that as visual cues are progressively extracted through external tools and textualized into the context, raw images become increasingly redundant. Continued image retention is associated with higher output entropy and can even degrade task accuracy. Motivated by these findings, we propose MM-ContextFold, a training-free framework that loads raw images only when needed. It maintains a persistent, text-only main context for high-level planning and spawns ephemeral branch contexts for image-dependent subtasks. Within each branch, the agent loads the relevant images, completes the subtask, and folds the result back into the main context as a concise textual summary; the images and branch trace are then discarded. Experiments on seven MAR benchmarks across five backbone models show that MM-ContextFold improves average accuracy by 6.3 percentage points over ReAct while reducing the working context length by 27.5%.

Abstract:
Connecting two photo-realistic reconstructed spaces is largely unexplored yet opens new opportunities in spatial dataset creation and immersive telepresence. While existing works focus on single room-wise generation, we present TranSpace, a method that synthesizes transition geometry to seamlessly connect two distinct reconstructed spaces via video generation models. This setting introduces a fundamental challenge: the transition region requires plausible hallucination, while the pre-existing geometry must be strictly preserved. Because video diffusion operates in pixel space rather than metric space, this asymmetry exposes the 3D inconsistency of video diffusionthe tendency to gradually distort existing geometry or hallucinate artifacts that degrade downstream novel view synthesis. To address this, we introduce Progressive Anchoring, an iterative scheme that uses camera pose estimation as a quality gate for generation. Two mechanisms are central: volumetric blending fuses generated frames with pre-existing geometry to preserve trusted regions, and bidirectional PnP localization estimates camera poses against blended references to bound error accumulation from both endpoints. We evaluate against state-of-the-art camera-conditioned video generation and pose estimation baselines, demonstrating improvements across PSNR, SSIM, and LPIPS metrics. A user study further confirms superior perceptual quality over text-to-3D scene generation and 3D inpainting methods.

Abstract:
Human-centric multimodal understanding (HCMU) aims to infer subjective human states from heterogeneous multimodal signals.Existing multimodal fusion methods typically operate at a coarse modality level, overlooking intra-modal diversity (e.g., semantic, affective, and acoustic cues), which limits their ability to capture fine-grained cross-modal consistency.To address this limitation, we propose Decoupled Multimodal Subspace Clustering (DMSC), a framework that reformulates multimodal fusion as a subspace-level structured assignment problem under a unified structural constraint. DMSC decomposes each modality into multiple latent subspaces and performs cross-modal clustering to selectively aggregate semantically consistent factors while preserving distinctive ones. A cluster-aware fusion mechanism with structural regularization further enforces intra-cluster consistency and inter-cluster separability, enabling principled and fine-grained multimodal fusion.Extensive experiments on six datasets across sentiment analysis, emotion recognition, and humor detection demonstrate that DMSC achieves state-of-the-art performance and remains robust under noisy and missing-modality settings.

Abstract:
Spotting micro-expressions (MEs) in naturalistic conversations is hindered by the ``Speech-Artifact Dilemma,'' in which speech-related facial kinematics mask genuine emotional cues. To overcome the limitations of semantic-dependent approaches, we propose AV-SyncME, a physics-driven framework that uses label-free acoustic--visual coherence to guide supervised ME spotting without speaking-state, speech/silence, or language annotations. Grounded in the source--filter model of speech production, AV-SyncME employs a non-semantic gating mechanism operating within the 2--7~Hz theta band to reduce the influence of speech-induced facial motion. In addition, an acoustic-anchored boundary loss explicitly couples temporal boundary regression to audio-energy transients, while a physiology-driven auxiliary task enhances sensitivity to subtle facial dynamics. Experiments show that AV-SyncME achieves a state-of-the-art F1_\mathrmDR of 34.12 on ConvoInsight-DB, outperforming existing methods by 7.5%, and achieves consistent cross-dataset improvements on three laboratory benchmarks.

Abstract:
Spiking Neural Networks (SNNs) have emerged as a promising energy-efficient paradigm for multimedia processing. Within SNNs, temporal coding conveys information through spike firing time, leading to sparser spike activity than in rate coding SNNs. While this sparsity improves computational efficiency, it also limits the representational capacity of temporal coding SNNs. In this work, we propose the Last-Spike-Modulated Spike Response Model (LSM-SRM), a self-connected spiking neuron for temporal coding SNNs that integrates neuron's last spike firing time inspired by autapse in biology. This model adaptively modulates excitability and improves the utilization of historical information, consequently enhancing neuronal dynamics and networks representation. To address the history-dependent optimization challenge posed by autapse, we introduce a counting spike-timing-dependent plasticity rule as an event-driven backward approximation by decomposing post-synaptic current into the spike count term and average temporal decay factor. Extensive experiments on both static and neuromorphic tasks demonstrate the effectiveness and efficiency of our method. For example, on CIFAR100 and DVSGesture datasets, our method increases sparse spike activity by 1.27% and 2.48% while improving performance by 2.06% and 1.04%, respectively. Additionally, we are the first to report the performance of directly trained temporal coding SNNs on larger-scale Tiny ImageNet dataset, highlighting the potential of efficient temporal coding SNNs.

Abstract:
\beginabstract Human-to-robot motion retargeting can be viewed as a structured cross-modal motion translation problem, where the semantic intent and temporal dynamics of human motion must be preserved, while respecting embodiment-specific topology, joint constraints, and physical geometry. Existing optimization-based methods can provide geometrically plausible initial mappings, but they are typically per-sequence and non-amortized, and often suffer from self-collisions, unsafe near-contacts, and local temporal artifacts. In this paper, we propose a two-stage framework for motion retargeting across diverse humanoid embodiments. In the first stage, we introduce a minimum-change refinement procedure that repairs collision-prone local kinematic chains through discrete local search. This produces physically improved supervisory trajectories while preserving motion semantics and temporal continuity. We then propose MaMoRetarget, a morphology-aware neural retargeting network that learns an end-to-end feed-forward mapping from structured human motion and robot specifications to executable robot motion. MaMoRetarget integrates a morphology-conditioned human motion encoder, a kinematic tree-based robot specification encoder, and a cross-attention decoder with explicit joint-limit handling, enabling joint modeling of human dynamics and robot-specific embodiment constraints. Experiments on a MoCap-based multi-robot benchmark show that our framework improves physical plausibility and temporal smoothness, supports downstream policy learning, and generalizes well across diverse humanoid embodiments. These results demonstrate that MaMoRetarget offers an efficient and scalable paradigm for high-quality humanoid motion retargeting. \endabstract

Abstract:
A primary challenge in deep-sea 3D reconstruction is the strong coupling between active illumination and the participating water medium. Existing underwater 3D Gaussian Splatting (3DGS) methods typically model backscattering along pixel rays as a fixed constant, neglecting the distance-dependent attenuation of active light sources. In terms of optimization, the densification process often confounds view-space positional gradients with medium noise signals, leading to systematic errors in radiance measurements and redundant primitives. To address these challenges, we introduce AbyssSplat, the first 3D reconstruction framework specifically designed for complex deep-sea optical environments. AbyssSplat incorporates two core innovations: a Hybrid Light Field Illumination Attenuation Model, which explicitly models active light source scattering to achieve physical-domain decoupling between non-uniform illumination and scene-intrinsic radiance; and a Medium-Guided Densification Modulation Mechanism, which integrates medium transmittance and Gaussian attributes to construct a geometric confidence, applying adaptive gradient amplitude modulation to low-confidence regions to suppress overfitting of medium noise by Gaussian primitives. Evaluations on the Eiffel Tower deep-sea and SeaThru-NeRF shallow-water datasets demonstrate that AbyssSplat achieves consistent state-of-the-art performance across all scenes, with average PSNR improvements of 0.67 dB and 0.39 dB respectively over the best competing methods, alongside up to 49.5% LPIPS reductionwhile requiring fewer Gaussian primitives and less training time.

Abstract:
Monocular metric depth estimation is often hindered by scale ambiguity. Lacking macroscopic physical priors, mainstream models struggle with complex scenes and tend to produce over-smoothed predictions. To address this, we propose MQ-Depth, a decoupled generative framework that extracts physical and semantic priors from Multimodal Large Language Models (MLLMs) to anchor the macroscopic metric scale and layout skeleton, thereby driving a Diffusion Transformer (DiT) to generate high-quality metric depth maps. To ensure robust cross-modal knowledge transfer, we introduce a "Blind-Querying" cross-modal representation filter. During the first training stage, the visual input to the DiT is masked. This forces learnable Meta-Queries to extract global metric scales from an MLLM, which then serve as semantic prompts for the DiT. In the second stage, visual features are restored to carve high-frequency geometric edges. Furthermore, to adapt to the perspective variations and asymmetric resizing common in diverse training data, we design Camera Perspective Modulation to inject camera intrinsic parameters and modulate feature distributions via Adaptive Layer Normalization (AdaLN). Experiments demonstrate that MQ-Depth achieves competitive zero-shot metric accuracy across multiple benchmarks while preserving sharp geometric boundaries.

Abstract:
In smart cabins, in-cabin behavior understanding typically relies on multimodal large language models (MLLMs) to perform reasoning over synchronized multi-view camera inputs. While such inputs provide complementary cues, real-world cabin monitoring is often affected by view occlusion and sensor failure. Therefore, robust reasoning should selectively exploit high-quality views rather than treating all views equally. However, existing MLLM-based approaches and reinforcement learning (RL) methods lack (i) explicit modeling of cross-view evidential roles and logical relations, and (ii) domain-grounded multi-view multimodal datasets that support both supervised fine-tuning (SFT) and RL-based optimization. To address these limitations, we introduce Driver-R1, a multi-view multimodal reasoning framework that employs Masked Multi-View GRPO (MVPO) to regularize reasoning by optimizing policies under randomly masked view subsets, encouraging evidence-based view selection and cross-view verification. To support this paradigm, we construct a multi-view Driver-Reasoning (DR) dataset with DR-6.4k for SFT and DR-16k for RL. Experiments on multiple benchmarks demonstrate that Driver-R1 consistently outperforms prior single-view, multi-view fusion, and standard GRPO baselines.

Abstract:
Video-based visible-infrared person re-identification (VVI-ReID) aims to match person identities across visible and infrared video sequences. Existing methods usually rely on fixed cross-modal alignment and insufficient temporal modeling, making them less effective in handling sequence-dependent visible-infrared discrepancies and temporal instability caused by transient frame-level disturbances in realistic video surveillance scenarios. To address these challenges, we propose a Dynamic Modality-Temporal Adaptation (DMTA) framework for VVI-ReID, which unifies modality-adaptive perception and temporal relational modeling within a complementary and coherent framework. Specifically, we design a Modality-Prompted Dynamic Perception (MPDP) method, which performs adaptive feature learning via dynamic expert routing conditioned on input sequence features and modality prompts, while enhancing cross-modality identity consistency with a modality compatibility constraint. We further develop a Sequential Graph Relational Modeling (SGRM) strategy, which captures temporal continuity among neighboring frames through relational propagation and aggregates global sequence semantics for more robust and stable sequence representation learning. Extensive experiments on the HITSZ-VCM and BUPTCampus benchmarks demonstrate that DMTA consistently outperforms existing state-of-the-art methods.

Authors: Yuankang Pan (School of Computing and Artificial Intelligence, Southwest Jiaotong University), Zhaoquan Yuan (School of Computing and Artificial Intelligence, Southwest Jiaotong University; Engineering Research Center of Sustainable Urban Intelligent Transportation, Ministry of Education), Fangying Xiong (School of Computing and Artificial Intelligence, Southwest Jiaotong University), Ao Luo (School of Computing and Artificial Intelligence, Southwest Jiaotong University), Wei Li (School of Computing and Artificial Intelligence, Southwest Jiaotong University), Xiao Wu (School of Computing and Artificial Intelligence, Southwest Jiaotong University)

Abstract:
Class-Incremental Object Detection (CIOD) aims to continuously extend object detectors to novel categories without catastrophic forgetting. Although LoRA has shown promise in class-incremental learning, directly applying it to CIOD remains suboptimal due to overlooked optimization dynamics induced by sequential low-rank updates in detection. In this paper, we propose LAMDA (Low-rAnk Magnitude-Direction Adaptation), a novel framework that explicitly regulates continual adaptation from a unified magnitudedirection decoupling perspective. Specifically, magnitude characterizes the extent of parameter drift on old tasks after learning new ones, while direction captures the gradient orientation induced by the new task. To regulate LoRA update magnitude, we introduce a Low-Rank Pyramid (LoRP) that progressively shrinks the rank based on task correlation and tightens the bound on cumulative interference. To control LoRA update direction, we propose to construct a Foreground-Centric Gradient Space (FCGS) that encapsulates the object-relevant historical gradient subspace and mitigates background-dominated bias. Furthermore, based on FCGS, we develop a Relaxed Gradient Projection (RGP) mechanism that adaptively preserves beneficial historical gradient components under object co-occurrence, enabling positive knowledge transfer. Extensive experiments on PASCAL VOC and MS COCO demonstrate that the proposed LAMDA consistently outperforms state-of-the-art methods, especially under long-sequence incremental settings.

Abstract:
Robust image matching across multi-modal and multi-domain conditions is a critical bottleneck for deploying vision systems in the real world, where autonomous agents must operate reliably across day and night, adverse weather conditions, and heterogeneous sensors (e.g., RGB, infrared, and event). In such scenarios, severe appearance variations and modality gaps often lead to unstable feature correspondences, significantly degrading matching accuracy. In this paper, we propose MMmatch, a plug-and-play solution in terms of architecture that enhances existing semi-dense and dense matching methods. To refine feature representations under multi-modal and multi-domain scenarios, we introduce a multi-scale latent diffusion framework. Since the backbone features of the existing matcher are not Gaussian-distributed and thus incompatible with diffusion, we adopt a two-stage training strategy. In the first stage, backbone features are mapped into a Gaussian latent space via a multi-scale encoder to align the latent distribution with the Gaussian assumption of the diffusion model. In the second stage, a lightweight multi-scale denoiser performs efficient latent-space refinement from multi-modal and multi-domain inputs by leveraging clean-image priors. MMmatch can be seamlessly integrated into existing pipelines and significantly improves matching robustness and pose estimation accuracy across diverse multi-modal and multi-domain scenarios, as validated by extensive experiments on both synthetic and real-world datasets.

Abstract:
Text-driven 3D scene editing faces a bottleneck: optimization-based methods often require modifications within diffusion architectures, precluding the use of powerful black-box Multimodal Large Language Models (MLLMs), whereas atlas-based methods support arbitrary MLLMs but suffer from geometric distortion and optical flow failures in wide-baseline scenes. To address these limitations, we propose Multi-Atlas Collaborative Editing (MACE), a novel framework for zero-shot 3D Gaussian Splatting manipulation. MACE adaptively partitions the viewing sphere via local geometric entropy to construct localized, low-distortion neural atlases. These decoupled atlases are then assembled into a unified mosaic, empowering off-the-shelf MLLMs to perform context-aware, globally coherent modifications. To ensure robust 3D reconstruction and multi-view consistency, we introduce a FastVGGT-driven geometry-guided consistency loss that replaces fragile flow heuristics, alongside a direction-aware selection mechanism for final 3DGS fine-tuning. Evaluations indicate that MACE effectively mitigates topological distortion, achieving robust multi-view consistency and semantic alignment with significantly reduced per-prompt editing latency.

Abstract:
Recovering a simulation-ready 3D garment asset from a single in-the-wild photograph remains a long-standing challenge. Prior methods address topology estimation and texture normalization in isolation and without closed-loop quality control. We present UniGarment, a closed-loop framework that couples the two stages through vision-language verification: a topology stage predicts a GarmentCode sewing pattern, a critic checks panel-boundary alignment against the photograph, the verified pattern localizes which image regions are fabric, and a final critic judges the draped result and re-triggers whichever stage failed. The recovered sewing pattern is what makes the output editable and re-drapeable rather than a static mesh. We evaluate pattern accuracy against genuine ground truth by rendering GarmentCode specifications through the same cloth simulator used at inference, so every evaluation image carries the exact pattern it was produced from. On this 156-garment benchmark the system reaches Panel IoU 0.357 and Seam Accuracy 0.594, and on in-the-wild photographs all garments drape successfully. Our analysis isolates two concrete bottlenecks. Panel count is recovered exactly for only 17% of garments, and seam accuracy degrades from 0.723 to 0.478 as garment complexity grows. It also shows that closed-loop behaviour is governed less by architecture than by how discriminative the critic's score is, a dependency we make explicit.

Abstract:
Appearance-based gaze estimation is fundamentally limited by the scarcity of training data that are both diverse and reliably labeled. Graphics-based synthesis methods provide geometric labels but often suffer from limited appearance diversity, whereas generative synthesis methods improve visual quality but usually model gaze implicitly, making it difficult to guarantee physically consistent eye motion and reliable gaze labels. To address these limitations, we propose a novel 3D-aware framework for controllable and label-faithful gaze data synthesis, with support for few-shot cross-domain adaptation. The framework decomposes eye appearance into a static field and a motion field, where gaze-dependent variation is modeled through explicit rigid eyeball rotation in a canonical 3D space. To improve label fidelity, we introduce a geometry-aware discriminator conditioned on projected pupil landmark masks, which explicitly enforces consistency between synthesized eye appearance and gaze geometry. We further develop an inversion-based few-shot adaptation strategy that recovers target-domain appearance factors in the latent space of the proposed controllable generator and recombines them with arbitrary gaze directions and camera parameters, enabling label-preserving target-style gaze synthesis from a small set of unlabeled target images. Experiments on MPIIGaze, Columbia, and EVE show that the proposed synthetic data achieve the best average downstream performance, obtaining the lowest average angular error across both MnistNet and ResNet-18, and that the proposed adaptation strategy remains effective with only a limited number of unlabeled target images. These results demonstrate that the proposed framework provides an effective solution for controllable gaze data synthesis, enabling diverse sample generation while preserving label fidelity for appearance-based gaze estimation.

Abstract:
Conventional image forgery detection methods produce only binary scores or pixel-level masks without interpretable evidence, while recent multimodal large language model (MLLM)-based approaches generate textual explanations that are merely post-hoc verbalizations of pre-determined classification results rather than products of genuine reasoning. Inspired by the forensic workflow of human judicial experts---``experimental analysis -- logical reasoning -- scientific evidence''---we propose an Agentic Tool-Augmented Reasoning (ATAR) framework for explainable image forgery detection, integrating 22 specialized forensic tools spanning seven complementary domains to autonomously detect, localize, and explain image forgeries through multi-turn reasoning. Concretely, we propose a Dual-Stream Forensic Reasoning paradigm to emulate the experimental analysis process of forensic experts: (1) a high-level semantic anomaly path, which magnifies suspicious regions for fine-grained inspection; and (2) a low-level forgery artifact path, which invokes forensic tools to extract objective artifactual evidence. To this end, we further propose a Forensics Curriculum Learning training strategy. First, during the General Experience SFT stage, an automated teacher--student mentoring pipeline is designed to synthesize multi-turn tool-usage reasoning trajectories. Subsequently, during the Forensic Scene RL stage, a Tool Prior Curriculum is introduced to guide early tool exploration and progressively transfer control to the agent, while a Structured Evidence Reward provides fine-grained process-level supervision. Experiments across IMDL, Deepfake detection, DMDL, and AIGC detection show that ATAR achieves 78.5% average image-level F1 on six zero-shot IMDL benchmarks, surpassing the strongest MLLM baseline by 11.8 percentage points, and remains competitive with specialized detectors on other tasks, while producing substantially more faithful and grounded explanations.

Abstract:
The rapid advancement of multi-agent systems has catalyzed increasing interest in Multi-Agent Communication (MAC). However, exchanging massive task-oriented multimodal data introduces a critical bandwidth bottleneck. While existing coding standards target human perception or point-to-point machine vision, coding strategies for MAC remain largely unexplored. To address this challenge, we formulate coding for multi-agent communication as a new problem and propose a coding paradigm motivated by the characteristics of MAC systems. Specifically, this paradigm is built upon two key principles. First, we introduce a Transform-before-Coding principle with a Characteristic-Aware Latent Rectification mechanism, which maps irregular multimodal representations into an entropy-friendly latent manifold for efficient coding. Second, we propose a Communication-Guided Coding principle that leverages inter-agent interactions to align the coding process with collaborative objectives. Furthermore, we instantiate the proposed paradigm for multi-agent collaborative perception in autonomous driving by developing an efficient Birds-Eye-View (BEV) feature compression scheme equipped with Sparse Mask-Guided Attention and Learnable Mask Modulator. Extensive experiments demonstrate that our approach seamlessly integrates with existing collaborative perception systems, reducing communication bandwidth by up to several hundred times while preserving competitive collaborative performance, thereby validating the effectiveness of the proposed coding paradigm.

Abstract:
In tri-modal visual question answering (VQA), auxiliary text is commonly used to complement visual and textual inputs, yet its reliability is often uncontrolled. While prior work studies modality conflicts in general, it remains unclear how different types of unreliable auxiliary text affect answer selection under fixed imagequestionoption contexts. In this work, we show that the most harmful auxiliary text is not necessarily the most factually incorrect, but the one that aligns with the question while contradicting the image and favoring a specific distractor, leading to systematic redirection of model predictions. To isolate this effect, we introduce the Textual Reliability Ladder, a controlled diagnostic protocol that decomposes auxiliary text along three axes: image consistency, question relevance, and option support. Across multiple datasets (ScienceQA, VCR, A-OKVQA, Causal-VidQA) and recent VLMs, we find that such distractor-supporting text induces the largest accuracy drops (up to 53.1%) and concentrates errors on specific incorrect options. To mitigate this failure mode, we propose a training-free inference-time intervention that explicitly counteracts this redirection effect via noise-stability steering and dynamic grounding, reducing redirected errors while largely preserving performance under faithful text. Our results highlight that auxiliary-text reliability must be understood at the decision level, rather than solely through factual correctness, and provide a practical pathway toward more robust tri-modal reasoning.

Abstract:
Class-incremental facial expression recognition (CI-FER), which learns new expression categories continuously without forgetting old expression categories, plays a crucial role for real-world affective computing. However, existing class-incremental learning methods suffer from severe catastrophic forgetting when applied to CI-FER. This performance drop mainly stems from their inability to learn robust semantic topological structures, due to semantic isolation among expressions and structure interference between old and new expressions during incremental learning. To address these challenges, we propose an adaptive semantic topological alignment (ASTA) method that leverages the powerful generalization capabilities of the large-scale vision-language model (CLIP) for CI-FER. ASTA adaptively aligns and calibrates the semantic topological structure of the expression space through two key components: (1) an AU-guided semantic anchoring (ASA) module, which overcomes semantic isolation by integrating a weighted action unit (AU) prompt ensemble block and an AU-based soft labeling mechanism to effectively establish fine-grained expression semantic anchors; and (2) an elastic subspace calibration (ESC) module, which acts as a structural safeguard against structure interference. By leveraging a multimodal confusion metric matrix that integrates visual-textual semantics with anatomical priors, the ESC module adaptively calibrates decision boundaries to carve out mutually non-intersecting feature subspaces for highly overlapping expressions. Extensive experiments across multiple benchmark datasets demonstrate that our ASTA method significantly mitigates catastrophic forgetting, substantially outperforming state-of-the-art methods.

Abstract:
3D vision-language understanding is central to multimedia research. Representative tasks such as visual grounding and text-to-shape retrieval require aligning 3D geometry with natural language that often depends on fine-grained geometric cues. Most existing 3D-language models build on generic backbones with limited geometric inductive bias, leaving them brittle to spatial transformations. Conversely, strictly invariant encoders risk collapsing the very spatial distinctions that geometry-sensitive queries depend on. We introduce Equivariant Covering Networks on E(3) (E3CoverNet), a 3D backbone designed to close this gap. Rather than directly modeling the full E(3) transformation space, E3CoverNet progressively builds equivariant representations along the subgroup chain E(1) ? E(2) ? E(3). At each stage, a learned covering set discretizes the corresponding symmetry group, and a geometry-aware attention layer synchronously updates invariant point features and equivariant coordinates through a dual-path design. We plug E3CoverNet into standard 3D-language pipelines and evaluate it on text-to-shape retrieval and 3D visual grounding benchmarks, where it outperforms competitive baselines across both tasks. A geometry-sensitive diagnostic further shows that E3CoverNet substantially improves robustness to spatial perturbations without sacrificing the semantic discrimination required for complex spatial queries.

Abstract:
Music emotion recognition is fundamental to affective computing and intelligent music systems. Contrastive audio-language pre-training models, designed for broad music understanding, offer a natural path to training-free recognition through cross-modal similarity matching. However, similarity matching is limited to discrete emotion words, making precise emotional value prediction infeasible without supervision. We propose MUSE (Music emotion recognition Using Semantic Embeddings), a training-free framework for static music emotion recognition that predicts song-level emotional values. We prompt a large language model to generate fine-grained descriptions conditioned on each node of a structured emotional coordinate grid, grounding each description in objective acoustic attributes. These descriptions form semantic acoustic anchors with precise coordinate bindings. At inference, any audio embedding is mapped to emotional coordinates using kernel regression over these anchors. MUSE surpasses training-free baselines built on contrastive pre-trained models and approaches fully supervised performance on multiple benchmarks. To the best of our knowledge, MUSE is the first training-free approach to formulate static audio-based music emotion recognition as a direct continuous Valence-Arousal regression, without relying on any discrete label as an intermediate output.

Abstract:
Recent progress in Design-to-Code (D2C) has shown that large multimodal models (MLLMs) can translate visual mockups into front-end code. However, existing benchmarks emphasize static visual fidelity while ignoring dynamic behaviors, including user interactions, state transitions, and data flow, that define real-world web applications. We introduce DynoGen, the first benchmark dedicated to dynamic front-end code generation. DynoGen provides a DSL-driven automatic data generation pipeline and a comprehensive evaluation suite that jointly measure visual accuracy, functional integrity, and executable correctness. Each DynoGen instance couples structured design intent (via DSL), multi-fidelity visual inputs, fully executable front-end code with dynamic logic, simulated backend APIs, and rich interaction traces. Evaluations of state-of-the-art MLLMs (e.g., GPT-5, Gemini 2.5-Flash) show that DynoGen exposes subtle but critical deficiencies in interactive reasoning and state management. By grounding assessments in actual execution, DynoGen establishes a scalable foundation for advancing practical, end-to-end web application generation.

Abstract:
Transparency-aware generation requires modeling not only RGB appearance but also alpha-based opacity and cross-layer composition, which are essential for tasks such as image matting, object removal, layer decomposition, and multi-layer content creation. However, existing RGBA-related methods remain largely fragmented, with separate pipelines designed for individual tasks. While a unified model is desirable, supervised fine-tuning alone is insufficient, as localized regression objectives cannot directly optimize the compositional fidelity, alpha-boundary precision, and structural consistency required for high-quality RGBA generation. To address this, we propose OmniAlpha, a unified multi-task reinforcement learning framework for transparency-aware generation and manipulation. OmniAlpha combines an end-to-end alpha-aware VAE and a sequence-to-sequence Diffusion Transformer, with a bi-directional layer axis in positional encoding to jointly model multiple RGBA inputs and outputs within a single forward pass. Built on a multi-task SFT cold start, it further performs GRPO-style post-training with layer-aware rewards defined on decoded RGBA outputs, enabling direct optimization of cross-layer coherence and fine transparency details. Experiments across five categories of transparency-aware tasks show that OmniAlpha consistently outperforms its unified SFT baseline and achieves strong performance against specialized expert models, including a 9.07% relative reduction in RGB L1 on layer decomposition and 74%/68% improvements over conventional matting tools on SAD/Grad for automatic matting.

Abstract:
Omni-modal Large Language Models (OmniLLMs) process visual, auditory, and textual signals within a unified autoregressive framework, yet their internal cross-modal information processing mechanisms remain underexplored, and the massive multi-modal token sequences challenge inference efficiency. Through layer-wise causal intervention with 12 attention knockout strategies, we reveal a three-stage processing pipeline: (1) Intra-modal Integration in shallow layers, where video and audio independently build coherent representations with negligible mutual interaction; (2) Cross-modal Injection in middle layers, where audiovisual information is written into text query tokens within a concentrated window of approximately 5 layers; (3) Textual Propagation in deep layers, where the model relies solely on fused text representations for decoding. Based on these findings, we propose OmniIF, a training-free, FlashAttention-compatible three-stage token pruning framework driven by a unified information flow capacity formula that combines directed transmission power with multi-head channel independence. The formula naturally specializes across stages: full log-det form for intra-modal deduplication, linear form with Joint Budget Competition for cross-modal pruning, and zero capacity for complete audiovisual removal. On Daily-Omni, OmniIF achieves 1.97× prefill speedup with only 1.17% accuracy drop; on WorldSense, 1.92× prefill speedup with only 0.47% drop.

Abstract:
Recent hybrid methods combining 3D Gaussians and parametric templates have shown progress in monocular human avatars. However, fixed topology and limited pose space often prevent accurate reconstruction of fine-grained dynamic details. We present E3-Avatar, a unified framework for building Efficient, Expressive, and Easy-to-use human avatars that achieve high fidelity with minimal training cost. E3-Avatar introduces a feature-based Gaussian initialization guided by the generative priors, which no longer relies on the topology of template meshes and significantly accelerates convergence. We further design a region-adaptive density control strategy that dynamically regulates Gaussian density and distribution across body regions with semantic attributes, ensuring global consistency while enhancing local details in expressive parts like the face and hands. Finally, a motion-aware spatial refinement network decouples pose-induced deformations from intrinsic appearance via cross-attention, enabling the modeling of dynamic high-frequency details under novel motions. E3-Avatar can reconstruct a photorealistic human avatar in 200 seconds with rendering at 70 FPS simply from a short monocular sequence, which also supports easy-to-use deployment and customization. Extensive experiments demonstrate its superior reconstruction quality, robust motion generalization, and efficient inference compared to state-of-the-art baselines.

Abstract:
Multi-view multi-label classification (MvMLC) is fundamental to multimedia applications yet suffers from missing views and incomplete annotations that undermine semantic correlation modeling. Existing methods tackle these issues through label completion, either by learning label correlations or imputing missing labels with pseudo-labels. However, both rely on heuristic metrics like cosine similarity that conflate genuine semantic relationships with random co-occurrences, severely limiting generalization under sparse supervision. To address this limitation, we propose SLD-Flow, a statistically grounded framework that reformulates label correlation learning as a problem of significance-aware dependency inference. At its core lies Statistical Label Dependency (SLD), which models label co-occurrence via the hypergeometric distribution, employing expectation to establish a baseline for random co-occurrence and variance to quantify its fluctuation. This formulation enables principled differentiation between meaningful dependencies and random noise, yielding a robust label relation graph. Building upon SLD, we further construct a label-structure-guided learning paradigm that integrates two key components: (i) manifold-based label completion guided by statistically significant dependencies to mitigate noise propagation, and (ii) distribution consistency regularization that injects SLD priors into representation learning to enforce structural alignment between feature and label spaces. Through the synergy of completion and regularization, this unified design effectively bridges label space structure with feature space learning, substantially enhancing robustness against missing views and label sparsity. Extensive experiments show that SLD-Flow consistently outperforms SOTA methods.

Abstract:
Graph-based federated multi-view clustering (FMVC) enables privacy-preserving collaborative analysis, yet traditional methods rely on full-sample graphs and suffer from prohibitive O(N^2) complexities, especially on large-scale data. Recent advances have introduced anchor graphs into FMVC, effectively reducing overhead to linear complexity. Despite efficiency, existing anchor-based FMVC methods exhibit two critical limitations: 1) They strictly rely on first-order similarities, failing to mine the latent high-order nonlinear semantics of complex data. 2) Due to the independent construction of local anchor spaces, there is a severe lack of correlation and geometric alignment among them. To tackle these issues, we propose a novel framework termed Tensor Coupled Federated Multi-View Clustering with Multi-order Anchor Graph (TFedMAG). Specifically, on the client, TFedMAG employs a multi-order anchor graph learning strategy to capture high-order correlations, while utilizing virtual anchors for privacy protection, and further produces a unified anchor graph. On the server, TFedMAG receives and splices fused graphs uploaded by each client into a third-order tensor to capture global multi-client features. Then, a tensor low-rank constraint is applied to enhance consistency and complementarity among clients. Additionally, extensive experiments are conducted on several real-world datasets, and the corresponding results demonstrate that the TFedMAG method achieves state-of-the-art clustering performance compared to existing FMVC methods.

Abstract:
Multimodal chain-of-thought methods produce textual reasoning steps, but the visual grounding underlying each step is never made explicit. When a model attends to the wrong region, nothing in the chain records this: the error is unverifiable during inference, unlocalizable after the fact, and uncorrectable without regenerating the entire chain. We propose \textscPerception Chains, which treats each reasoning step as a grounding commitment: before drawing any conclusion, the model must name a region, extract an attribute from that region, and assign it to a reasoning slot. With this structure in place, every grounding decision is independently checkable and a wrong answer is traceable to the step where the binding failed. Correct output format, however, is not sufficient. A model satisfying the format can still leave its hidden states at anchor positions in a non-attenuating drift regime. Across our measured settings, task-loss gradients are nearly orthogonal to the stability objective at these positions, and standard fine-tuning does not eliminate the drift. A stability loss directly constrains this geometry and enforces transversal contraction. High-quality reference trajectories are constructed without human annotation by retaining only chains where independent samples from a frozen 72B VLM converge, filtering at the anchor level rather than by answer correctness alone. At inference time, a consistency check identifies anchors whose evidence conflicts with the provisional answer; only those anchors are re-localized before the final answer is committed. On eight benchmarks against six baselines, \textscPerception Chains achieves the highest average accuracy, with the largest gains on tasks requiring precise region-attribute binding. On DIOR-RSVG, grounding accuracy rises from 61.7% to 72.6% over the strongest spatial baseline. The stability loss, not the anchor format, is the transferable component: gains carry to seven held-out benchmarks unseen during training.

Abstract:
\beginabstract Vision-Text Compression (VTC) renders long texts into images and encodes them through the vision encoder (ViT), compressing thousands of text tokens into far fewer visual tokens. However, since the ViT is pretrained predominantly on natural images, it captures visual attributes (glyphs, font sizes, layout) rather than linguistic semantics, causing rendered-image representations to diverge from native-text representations. We term this \emphcross-path inconsistency and show, via rendering perturbation experiments, that it is a critical yet overlooked bottleneck of VTC. We propose SPIRAL (Self-improving Path Integration and Re\-alignment), a self-supervised alignment framework that closes this gap using only the model's own text-path behavior as supervision, requiring no external teachers or additional annotations. SPIRAL operates at two complementary granularities: token-level \emphon-policy distillation (OPD) for local faithfulness, and sequence-level \emphpreference optimization (DPO) for global coherence. On VTCBench, SPIRAL improves the overall score of Qwen3-VL-8B from 35.10 to 54.02, approaching the native text-input performance (55.60) and outperforming models up to 30× larger. The two granularities exhibit complementary strengths: OPD excels at retrieval and is sample-efficient, while DPO is stronger on reasoning and memory and scales better with data. SPIRAL's benefits also generalize to out-of-domain benchmarks, confirming that effective VTC hinges on aligning rendered-image representations back to native-text semantics. \endabstract

Abstract:
Recent portrait video generation methods rely on either hand-crafted motion priors or heavyweight pretrained video generators. Prior-based descriptors often miss subtle dynamics, while pretrained models incur high inference costs. We address this trade-off with Semantic Latent Motion (SeMo), a self-supervised framework that learns a compact, generation-oriented portrait motion latent. SeMo follows three steps: abstraction, reasoning, and generation. A masked motion encoder compresses each frame into a small latent token set, an audio-conditioned generator predicts motion in this space, and a rectified-flow decoder renders target frames from the reference image and target motion. This design keeps the reasoning space compact while leaving high-frequency portrait details to the decoder. On portrait benchmarks, SeMo achieves strong reconstruction and audio-driven generation quality, handles challenging cases such as partial faces and head turns, and runs at 28.47 FPS with one latent token per frame at 256×256. In a user study, SeMo obtains an 81% win rate in perceived realism over prior methods. These results indicate that SeMo provides an effective compact motion latent for efficient, high-quality portrait animation.

Abstract:
Tampered text detection is increasingly important for digital forensics, media integrity assessment, and document security, since manipulated text can substantially alter image semantics while remaining visually plausible. However, existing methods remain insufficient in two key aspects. First, they provide limited evidence-grounded interpretability, making it difficult to relate predictions to forged text content, low-level visual artifacts, and high-level semantic inconsistencies. Second, they rely primarily on visual appearance and therefore struggle to capture subtle forensic traces and precisely delineate small, elongated, and irregular tampered text regions. To address these challenges, we propose SMELT, a Semantic-guided Multi-Evidence Localization framework for explainable Tampered text detection. SMELT consists of a semantic-guided evidence prompting branch and a multi-evidence localizer. The former extracts tamper-relevant semantic cues from a vision-language model and converts them into localization prompts. Subsequently, the multi-evidence localizer jointly exploits visual appearance, compression-aware frequency evidence, and boundary-aware high-resolution decoding to produce precise tamper masks. To facilitate this task, we further construct VISTA-Text, an interpretable tampered text dataset with diverse domains, tampering types, and structured annotations covering authenticity judgment, manipulated text content, low-level visual artifacts, and high-level semantic inconsistencies. Extensive experiments on VISTA-Text and additional benchmarks demonstrate that SMELT consistently improves tampered text localization performance while supporting evidence-grounded interpretation.

Abstract:
Large pre-trained Vision-Language Models (VLMs) excel in zero-shot generalization but face challenges in efficient few-shot adaptation. Existing parameter-efficient methods apply uniform updates across all tokens, overlooking the inherent sparsity of cross-modal interactions and exacerbating overfitting. We propose SFASA: Semantic Factor-Aware Sparse Adaptation (SFASA), an approach that leverages multimodal contrastive learning to recover invariant semantic factors and dynamically routes only semantically salient tokens into lightweight low-rank adapters. By reusing frozen VLM projection heads as zero-parameter modality mappers, SFASA preserves pre-trained cross-modal priors while enabling targeted task-specific refinement. Extensive experiments on 11 datasets across four evaluation protocols show that SFASA consistently outperforms state-of-the-art methods in base-to-novel generalization, cross-dataset transfer, domain generalization, and few-shot learning.

Abstract:
Remote sensing object detection is fundamentally challenged by scale variations and background clutter. Existing discriminative methods often lack the generative comprehension needed to interpret ambiguous objects from incomplete visual cues. While adapting text-to-image diffusion models offers rich semantic priors to bridge this gap, their application is limited by a dual dissonance: semantic manifold misalignment and generative temporal trajectory underutilization. In this paper, we propose rectified consolidated latent diffusion (RCLD), a novel generative-perception framework. Firstly, we present a semantic manifold rectification to purify the semantic space. It uses a learnable latent probe to dynamically filter out irrelevant text embeddings, thereby restricting multi-modal interactions to the valid scene subspace. Then we propose a temporal trajectory consolidation to integrate the generative history. This module aligns the detailed textures of late diffusion stages with the stable categorical certainty established in early stages. Extensive experiments on four challenging benchmarks demonstrate that RCLD achieves state-of-the-art performance, exhibiting superior stability against complex degradations in aerial imagery.

Abstract:
As modern video-based human mesh regressors produce increasingly smooth motions, test-time refinement must optimize 2D alignment without destroying this inherent temporal consistency. However, current refinement paradigms, both classical per-frame optimization and recent score-based diffusion samplers, typically rely on frame-wise updates, which can reintroduce high-frequency jitter. To address this, we propose SplineHMR, a representation-level temporal prior that constrains pose and camera updates to a low-dimensional B-spline trajectory subspace. We instantiate this prior in two refinement settings: (1) Spline-Opt, a control-space LBFGS optimizer for fast (>140 FPS) and temporally stable sequence refinement; and (2) Spline-Diff, a B-spline projection operator embedded into ScoreHMR's DDIM loop to improve the alignment-stability trade-off during diffusion-based refinement. Across both instantiations, the proposed subspace prior consistently reduces temporal inconsistency, reflected most clearly by acceleration error, while maintaining good 2D alignment and competitive 3D fidelity.

Abstract:
While Multimodal Large Language Models (MLLMs) excel in short video understanding, extending them to long videos remains challenging due to memory bottlenecks and temporal hallucinations. Existing uniform sampling and discrete pruning methods often disrupt temporal boundaries and distort physical time scales. To address this, we propose ELASTIC-V, an efficient, high-fidelity framework for long video temporal grounding. It contains two key components. To break memory bottlenecks, we propose Saliency-aware Parallel Weight Aggregation (SPWA), a query-driven compression module that adaptively suppresses redundant background content while preserving temporally salient anchors under a fixed budget. To mitigate temporal distortions, we propose Elastic Temporal Attribute-aware Positional Encoding (ETAPE), which reconstructs temporal triplets via visual-gated Fourier mapping to restore causal consistency. To support evaluation, we further establish CE-Bench, a multi-task benchmark evaluating hierarchical and causal complex events. Experiments demonstrate that ELASTIC-V outperforms state-of-the-art MLLMs, achieving 48.5% mIoU and an 11.5s MAE-T on CE-Bench, alongside a notable 5.3-point accuracy boost on the public Video-MME benchmark. By simultaneously reducing peak VRAM by 27.3% with competitive inference latency, ELASTIC-V offers a robust, hardware-efficient paradigm for long-form video reasoning. Our code and dataset will be made publicly available.

Abstract:
Considering efficiency, ultra-high-definition (UHD) low-light image restoration is extremely challenging. Existing methods based on Transformer architectures or high-dimensional complex convolutional neural networks often suffer from the ``memory wall" bottleneck, failing to achieve millisecond-level inference on edge devices. To address this issue, we propose a novel real-time UHD low-light enhancement network that fuses geometric features using Clifford algebra in 2D Euclidean space. First, we construct a four-layer feature pyramid with gradually increasing resolution, which decomposes input images into low-frequency and high-frequency structural components via a Gaussian blur kernel, and adopts a lightweight U-Net based on depthwise separable convolution for dual-branch feature extraction. Second, to resolve structural information loss and artifacts from traditional high-low frequency feature fusion, we introduce spatially aware Clifford algebra, which maps feature tensors to a multivector space (scalars, vectors, bivectors) and uses Clifford similarity to aggregate features while suppressing noise and preserving textures. In the reconstruction stage, the network outputs adaptive Gamma and Gain maps that perform physically constrained non-linear brightness adjustment according to Retinex theory. Integrated with FP16 mixed-precision computation and dynamic operator fusion, our method achieves millisecond-level inference for 4K/8K images on a single consumer-grade device, while outperforming state-of-the-art (SOTA) models on several restoration metrics.

Abstract:
Skeleton-based human activity recognition (HAR) with deep learning typically requires large-scale labeled datasets, yet collecting such data in real environments is costly. Recent advances in generative models, such as text-to-video and image-conditioned video generation, make synthetic activity videos increasingly accessible, providing auxiliary data that can be converted into skeleton sequences for training. However, directly applying generated data to real-world HAR remains challenging because generated videos often fail to faithfully capture fine-grained pose patterns and temporal dynamics of real actions, leading to discrepancies in pose structure, execution speed, and local motion rhythm. These discrepancies become especially problematic when only limited labeled real skeleton data is available. To address this challenge, we propose MultiWin-RealAnchorMix, a generated-data-assisted sim-to-real skeleton-based HAR framework under limited real supervision. Our method constructs multi-window views of generated and real sequences to capture local motion dynamics at different temporal scales. For each generated view, we retrieve same-class real anchors from the limited labeled real set and perform dynamics-aware dual-space mixing to reduce the mismatch while preserving motion diversity. Extensive experiments on OpenPack, NTU60, and NTU120 demonstrate consistent improvements over strong baselines, with particularly large gains on the challenging industrial benchmark OpenPack.

Abstract:
Long-term video understanding requires interpreting complex temporal events and reasoning over procedural activities. While instructional video corpora, like HowTo100M, offer rich resources for model training, they present significant challenges, including noisy ASR transcripts and inconsistent temporal alignments between narration and visual content. In this work, we introduce an automated, training-free pipeline to extract high-quality procedural annotations from in-the-wild instructional videos. Our approach segments videos into coherent shots, filters poorly aligned content, and leverages state-of-the-art multimodal and large language models (Qwen2.5-VL and DeepSeek-R1) to generate structured, temporally grounded procedural steps. This pipeline yields DenseStep2M, a large-scale dataset comprising approximately 100K videos and 2M detailed instructional steps, designed to support comprehensive long-form video understanding. To rigorously evaluate our pipeline, we curate DenseCaption100, a benchmark of high-quality, human-written captions. Evaluations demonstrate strong alignment between our auto-generated steps and human annotations. Furthermore, we validate the utility of DenseStep2M across three core downstream tasks: dense video captioning, procedural step grounding, and cross-modal retrieval. Models fine-tuned on DenseStep2M achieve substantial gains in captioning quality and temporal localization, while exhibiting robust zero-shot generalization across egocentric, exocentric, and mixed-perspective domains. These results underscore the effectiveness of DenseStep2M in facilitating advanced multimodal alignment and long-term activity reasoning. Our dataset is available at https://huggingface.co/datasets/mingjige/DenseStep2M.

Abstract:
Large Vision-Language Models (LVLMs) have achieved remarkable performance on a wide range of multimodal tasks, yet they still suffer from hallucination, generating content that is inconsistent with the input image. Existing inference-time mitigation methods typically enhance visual attention or suppress language priors, but they mainly focus on attention allocation while overlooking the reliability of the attended visual representations. In this paper, we show that effective hallucination mitigation depends not only on where the model attends, but also on whether the attended visual evidence is internally consistent. To this end, we propose VCI, a training-free inference-time intervention framework that adaptively regulates visual enhancement according to both cross-source visual-language discrepancy and visual consistency. Specifically, VCI first estimates the discrepancy between visual evidence and textual prior knowledge at the current decoding step to identify potential deviation from visual grounding. It then measures the consistency of the attended visual evidence by computing the attention-weighted dispersion of visual representations. Based on these two signals, VCI adaptively adjusts visual attention enhancement and fuses the corrected prediction with the original decoding distribution, thereby reducing hallucinations while preserving generation fluency. Extensive experiments on multiple LVLM backbones and hallucination benchmarks, including POPE, CHAIR, and MME, demonstrate that VCI consistently outperforms existing inference-time baselines.

Abstract:
Spoken language models (SLMs) enable natural human-computer interaction, but their reasoning ability still lags behind that of text-based large language models, especially on spoken mathematical question answering tasks. One important reason is that SLMs reason over purely verbalized mathematical expressions, which are harder to interpret than symbolic text. However, directly transferring text-based reasoning to SLMs is nontrivial due to architectural constraints and the additional computational requirements. To address this challenge, we propose Efficient Chain-of-Modality Reasoning (ECoM Reasoning), the first framework to introduce compressed reasoning into SLMs. By compressing the textual component so that it jointly serves as speech guidance and reasoning representation, ECoM Reasoning improves reasoning accuracy while using a smaller token budget than the standard Chain-of-Modality (CoM) architecture, which generates intermediate text before speech. To train this capability, we further propose Progressive Compression, a curriculum-based strategy that gradually trains the model from full-form reasoning to compressed reasoning. Experiments on spoken mathematical question answering benchmarks show that ECoM Reasoning improves accuracy by 21% over standard CoM without explicit reasoning, and by 3% over CoM with full reasoning traces while using only 40% of the text tokens, demonstrating that it enhances SLM reasoning while remaining inference-efficient.

Abstract:
The Cross-View Geo-Localization (CVGL) task aims to match images of the same geographic location captured from different viewpoints. For this task, Spiking Neural Networks (SNNs) offer a promising, energy-efficient solution for deployment on power-constrained edge devices. However, their limited information representation capacity restricts their performance. This can be attributed to two key factors: 1) severe spike degradation weakens the discriminability of spike features, and 2) fixed partitioning strategies disrupt the spatial continuity of spike representations. To address these issues, we propose SpikeGeo, the first SNN-based framework for CVGL. For the first challenge, we introduce Multi-scale and Descriptor information Guided Attention (MDGA), which employs dual guidance mechanisms to enhance the expressiveness. For the second, we design an Adaptive Partitioning Module (APM) driven by spike activation intensity to adaptively determine partition centers and boundaries. Extensive experiments on the University-1652 and SUES-200 benchmarks demonstrate that SpikeGeo achieves State-Of-The-Art (SOTA) performance with over 8.68 lower energy consumption than prior best-performing methods.

Abstract:
Multimodal large language models (MLLMs) have shown growing potential in digital pathology, yet their reasoning often remains weakly grounded in the actual visual evidence within images. Existing supervised fine-tuning and reinforcement learning approaches primarily optimize final answer correctness, resulting in a persistent disconnect between visual perception, evidence localization, and reasoning, thereby leaving models unconstrained to rely on linguistic priors. We propose Counterfactual Grounding Policy Optimization (CGPO), a reinforcement learning algorithm that explicitly promotes Evidence-Sensitive Reasoning (ESR) by modeling the causal dependence between model predictions and localized visual evidence under weak supervision. CGPO constructs counterfactual images by masking predicted evidence regions and leverages token-level distributional changes to derive ESR-aware response scaling and fine-grained token-level advantage modulation. A grounding consistency regularization further constrains evidence localization to prevent reward hacking during training. Experiments across multiple pathology benchmarks demonstrate that CGPO consistently improves both task performance and visual evidence localization, achieving performance exceeding strong closed-source models and approaching human experts.

Abstract:
Cross-modal hashing enables efficient large-scale retrieval with compact storage, but most methods assume complete paired data and supervised settings. In practice, data are often incomplete and unlabeled, where existing methods rely on GANs or VAEs to reconstruct missing modalities or discard incomplete samples. However, even improved reconstructions are underutilized due to suboptimal geometric optimization and weak cross-modal alignment, resulting in unreliable similarity preservation under high missing rates. To overcome these challenges, we propose Unsupervised Ge\-nerative Manifold Alignment Hashing (UGMAH), which leverages foundation models for high-quality imputation and introduces two algorithmic innovations. First, dynamically coupled manifold reconstruction optimizes latent topology through polynomial-based learnable activations to preserve multi-scale structural properties. Second, hierarchical semantic preserving hashing discovers fine-grained structures and establishes robust alignment via graph convolutional propagation and Fused Gromov-Wasserstein optimal tr\-ansport. Extensive experiments on three benchmarks demonstrate that UGMAH consistently outperforms state-of-the-art methods, achieving superior retrieval accuracy even under severe incompleteness.

Abstract:
Multi-view unsupervised feature selection (MUFS), which selects informative features from multi-view unlabeled data, has attracted increasing research interest in recent years. Although great efforts have been devoted to MUFS, several challenges remain: 1) existing methods for incomplete multi-view data are limited to handling missing views and are unable to address the more general scenario of missing variables, where some features have missing values in certain views; 2) most methods address incomplete data by first imputing missing values and then performing feature selection, treating these two processes independently and overlooking their interactions; 3) missing data can result in an inaccurate similarity graph, which reduces the performance of feature selection. To solve this dilemma, we propose a novel MUFS method for incomplete multi-view data with missing variables, termed Tensorized Reliable UnSupervised mulTi-view Feature Selection (TRUST-FS). TRUST-FS introduces a new adaptive-weighted CP decomposition that simultaneously performs feature selection, missing-variable imputation, and view weight learning within a unified tensor factorization framework. By utilizing Subjective Logic to acquire trustworthy cross-view similarity information, TRUST-FS facilitates learning a reliable similarity graph, which subsequently guides feature selection and imputation. Comprehensive experimental results demonstrate the effectiveness and superiority of our method over state-of-the-art methods.

Authors: Bin Chen (College of Computer and Data Science, Fuzhou University), Jia-Li Yin (College of Computer and Data Science, Fuzhou University), Zhaozhe Hu (College of Computer and Data Science, Fuzhou University), Genna Ba (Fuzhou University), Xiaolei Liu (National Interdisciplinary Research Center of Engineering Physics, China Academy of Engineering Physics), Wei Lin (College of Mathematics and Computer Science, Fujian University of Technology), Yuanlong Yu (College of Computer and Data Science, Fuzhou University), Shouling Ji (College of Computer Science and Technology, Zhejiang University)

Abstract:
Deep neural networks (DNNs) are known to be vulnerable to adversarial examples, raising serious security concerns in real-world applications. In black-box settings, transfer-based attacks exploit the transferability of adversarial examples crafted on a surrogate model to fool unknown target models, making transferability a central issue in practical adversarial attacks. Recent feature-level transfer attacks improve transferability by perturbing intermediate representations that are more likely to be shared across architectures. However, their effectiveness is still limited by unreliable feature-importance estimation and an over-reliance on single-layer features, which together hinder the accurate localization and suppression of transferable intrinsic features. In particular, existing methods often estimate feature importance from a single input or heuristic transformations, making the resulting indicators easily contaminated by unstable non-robust responses. We observe that, as the diffusion forward process progressively perturbs an input, unstable non-robust responses are gradually attenuated, while semantically relevant robust responses become more concentrated. Inspired by this observation, we propose the Diffusion-based Importance-aware Feature Pyramid Attack (DIFPA), a novel feature-level attack that improves transferability from both importance estimation and hierarchical feature modeling. Specifically, we introduce a Diffusion-Guided Importance Estimator (DGIE), which exploits the diffusion forward marginal as a principled and controllable neighborhood sampler to probe local feature stability, suppress unstable non-robust responses, and produce more reliable feature-importance indicators. We further propose a Hierarchical Feature Pyramid Attack (HFPA), which systematically aggregates important features across multiple layers and exploits the hierarchical semantic diversity of DNNs to suppress transferable intrinsic features in a pyramid manner rather than at a single layer. Extensive experiments demonstrate that DIFPA consistently outperforms state-of-the-art methods, achieving average improvements of 8.17% against normally trained models and 7.92% against defended models.

Abstract:
Coverless image-in-image steganography (CIIS) has attracted increasing attention due to its high hiding capacity and stronger resistance to steganalysis. However, our investigation reveals that current schemes for general image scenarios suffer from significant information leakage under unauthorized recovery, limited diversity of stego content, and constraints on secret image categories. We attribute these limitations to the inherent correlations within the secret images. Motivated by the need for an effective CIIS scheme that can transform arbitrary secret images into suspicion-free and diverse stego images, we propose a scheme with enhanced security for general image scenarios. We first analyze the compatibility between steganography and the mainstream image generation process, based on which we develop a more general and reasonable paradigm for CIIS. Under this paradigm, we further propose an implementation in which a general key-driven latent transformation is used to break the internal correlations of secret images, thereby reducing secret leakage and improving content diversity, while semantically unrelated auxiliary information is transmitted through a text modal to support reliable recovery. In addition, a compensation strategy is incorporated to improve recovery accuracy. Experimental results on diverse image categories demonstrate that our scheme outperforms existing schemes in security, flexibility, and recovery accuracy, paving the way for secure and practical CIIS.

Abstract:
\beginabstract Learned image compression (LIC) has achieved impressive rate-distortion performance. However, existing methods remain highly vulnerable to packet loss, a common challenge in satellite and emergency communications. This vulnerability stems from non-uniform information distribution at the packetization stage and sequential decoding dependencies at the entropy coding stage. We propose an end-to-end loss-resilient image compression scheme that addresses both. Before packetization, we introduce an Inter-Channel Redistribution (ICR) mechanism to redistribute channel energy, preventing critical information concentrating in a small subset of channels. Then, an Interleaved Channel Grouping (ICG) strategy partitions latent channels in a strided manner to disperse information across packets, with each packet kept within constrained sizes. To limit cascading errors from lost packets, we adopt a two-layer dual-branch autoregressive structure to shorten the dependency chain. Extensive experiments demonstrate that our method consistently outperforms existing approaches in both reconstruction quality and stability. At 20% packet loss, it achieves an average PSNR gain of 1.84 dB over LossResilientLIC while reducing PSNR variance by an order of magnitude. Notably, trained under uniform random loss only, our model generalizes to bursty loss modeled by the GilbertElliott channel, outperforming methods explicitly trained for such conditions. \endabstract

Abstract:
Multimodal signals, such as visual, audio, and tactile data, are increasingly maintained as persistent digital assets in immersive communication systems and digital twins. In these settings, the same multimodal content is repeatedly accessed by heterogeneous receivers with varying modality and bandwidth requirements. Existing compression and Joint Source-Channel Coding (JSCC) methods typically follow a per-request encoding paradigm, resulting in redundant computation and low efficiency during repeated access. To address this issue, we propose ParaJSCC, a multimodal JSCC framework designed for reusable representation serving. ParaJSCC converts each multimodal sample offline at the cloud/content server into a compact, quantized parameter package, which is then stored at the edge serving node for low-latency access. During serving, only the subset required by the current request is transmitted over the wireless channel, followed by lightweight decoding at the receiver. The framework employs a progressive shared-private parameterization to support modality-selective transmission and scalable reconstruction under varying bandwidth constraints. Experiments on multimodal datasets show that ParaJSCC significantly reduces online latency (e.g., from 17.18 ms to 4.34 ms for image-only requests and from 43.96 ms to 11.21 ms for full multimodal requests) and transmission rate (by 47.8%--51.2% for selective requests), while maintaining strong reconstruction quality under noisy channels.

Abstract:
Real-time volumetric media streaming is emerging as a key enabler for immersive communication, yet the perceptual impact of transport-layer impairments on point cloud delivery over QUIC remains largely unexplored. Existing studies either evaluate codecs through objective metrics alone, rely on simulated network traces, or omit subjective validation in immersive environments, leaving a gap between measured network performance and the quality actually perceived by users wearing VR headsets. We present VQUEST, a cross-layer measurement framework that bridges this gap through an end-to-end pipeline spanning Draco-encoded point cloud delivery over QUIC, 5G edge rendering on a private standalone network, and subjective quality evaluation on a Meta Quest 3 HMD following ITU-T P.910 guidelines. Through controlled perceptual evaluation of Draco and transport-layer analysis, we identify quality thresholds for latency and packet loss, revealing that latency and packet loss degrade delivery through distinct mechanisms, a broadly uniform shift of the delivery-time distribution versus expansion of its tail. Our cross-layer analysis connects QUIC transport behavior to the observed quality differences, providing actionable insights for transport protocol design in volumetric streaming systems.

Abstract:
Generating natural-language coaching feedback on motor skills can accelerate learning, yet expert coaches are scarce and expensive. Existing reference-based methods require expert demonstrations at both training and inference time, limiting practical deployment. We propose AIDE (Automated Instruction via Distilled Expertise), a framework that exploits expert references only during training and generates feedback from a learner's pose sequence alone at inference. A teacher model first learns to generate feedback from paired learner--expert poses via a frozen language model, producing separate learner tokens and difference tokens that encode the learner--expert difference. A student model then inherits the teacher's encoder and weight initialization, replacing the explicit expert comparison with an auxiliary module that produces complementary tokens from the learner's pose alone. On the ExpertAF dataset, AIDE outperforms reference-free baselines on most metrics and performs comparably to methods requiring expert demonstrations at both training and inference, with LLM-based evaluation supporting these findings.

Abstract:
Multimodal large language models (MLLMs) provide semantic reasoning but lack the physical grounding required for embodied tasks. Conversely, vision-language-action models (VLAs) excel at low-level robot control yet struggle with complex multimodal reasoning, and embodied LLMs (ELLMs) remain constrained to digital space, limiting their ability to generalize to the physical world for direct robot control. Thus, unified large models that operate seamlessly across digital and physical spaces while generalizing across tasks and embodiments require further exploration. In this paper, we introduce Boundless Large Model (BLM-1), a multimodal spatial foundation model that preserves instruction-following and reasoning, incorporates embodied knowledge, and supports cross-embodiment control. BLM-1 integrates three key capabilities (cross-space transfer, cross-task learning, and cross-embodiment generalization) via a two-stage training paradigm. Stage I injects embodied knowledge into the MLLM through curated digital corpora while maintaining instruction-following ability. Stage II trains a cross-embodiment policy through an intent-bridging interface that extracts high-level semantics from the MLLM to guide control, without fine-tuning the MLLM backbone. Evaluations across digital and physical benchmarks show that a single BLM-1 instance outperforms baselines from four model families (MLLMs, ELLMs, VLAs, and GMLMs).

Abstract:
EEG event detection is crucial for braincomputer interfaces (BCIs), where reliable control and user safety depend on precise temporal localization of cognitive state transitions. In affective computing and neural decoding, such temporally localized events provide the key link between brain dynamics and human behavior. However, weakly supervised EEG event detection remains challenging in scalable BCIs, as conventional multi-instance learning (MIL) typically compresses complex temporal dynamics into coarse bag-level predictions. This limitation is especially severe for cross-boundary bags containing both event and non-event states, leading to ambiguous representations and limited interpretability. To this end, we propose Boundary-Aware Multi-Instance Learning (BAMIL), a unified framework that reconstructs temporal boundaries within bags and converts weak bag-level supervision into reliable instance-level event inference. Specifically, BAMIL introduces an Instance Tri-state-Aware (ITA) module, modeling continuous yet non-stationary cognitive transitions, to identify boundary instances and extract event-related evidence. An Evidence-Assisted Aggregation (EAA) then produces more discriminative bag-level decisions via attention mechanism. For finer temporal detection, a Self-consistent Structural Inference (SSI) module enforces consistency between bag-level predictions and instance-level latent trajectories, enabling instance-level event localization with improved temporal precision and interpretability. Extensive experiments on several public EEG datasets covering both active and passive BCIs demonstrate that BAMIL achieves state-of-the-art (SOTA) performance, providing a unified and interpretable solution across various BCI scenarios.

Abstract:
Diffusion models have driven remarkable advancements in fashion image generation, yet prior works usually treat garment generation and virtual dressing as separate problems, limiting their flexibility in real-world fashion workflows. Moreover, fashion image synthesis under multi-source heterogeneous conditions remains challenging, as existing methods typically rely on simple feature concatenation or static layer-wise injection, which often causes attribute entanglement and semantic interference. To address these issues, we propose VersaVogue, a unified framework for multi-condition controllable fashion synthesis that jointly supports garment generation and virtual dressing, corresponding to the design and showcase stages of the fashion lifecycle. Specifically, we introduce a trait-routing attention (TA) module that leverages a mixture-of-experts mechanism to dynamically route condition features to the most compatible experts and generative layers, enabling disentangled injection of visual attributes such as texture, shape, and color. To further improve realism and controllability, we develop an automated multi-perspective preference optimization (MPO) pipeline that constructs preference data without human annotation or task-specific reward models. By combining evaluators of content fidelity, textual alignment, and perceptual quality, MPO identifies reliable preference pairs, which are then used to optimize the model via direct preference optimization (DPO). Extensive experiments on both garment generation and virtual dressing benchmarks demonstrate that VersaVogue consistently outperforms existing methods in visual fidelity, semantic consistency, and fine-grained controllability.

Abstract:
Recently, state space models (SSMs) with strong global modeling capabilities and linear computational complexity have attracted significant attention from the community. Existing SSM-based methods for point-cloud analysis typically flatten 3D data into 1D sequences by following manually designed scanning orders. However, this data-agnostic strategy inherently constrains adaptability to complex geometric variations in point clouds, thereby precluding effective exploitation of explicit spatial structures during feature extraction. To address this issue, we proposed a novel SSM-based architecture termed as Point Deformable Mamba (PDMamba). The key insight of our approach is to exploit a deformable greedy scanning (DGS), that dynamically adjusts the scanning order in a data-driven manner rather than manually designed to enhance the capture and processing of relevant input features, significantly improving the ability and robustness to learn point structures. PDMamba significantly surpasses SOTA MLP?based method DeLA, Transformer?based method PTv3, and Mamba?based method PCM in downstream tasks including object classification, part segmentation and semantic segmentation.

Abstract:
Embodied failures in human-robot interaction rarely appear as isolated mistakes. They emerge as multimodal risk trajectories involving delayed responses, misunderstandings, interruptions, unreliable sensing, and escalating user frustration. Existing work often treats anticipation, intervention timing, action selection, and explanation as separate tasks, which fragments the logic behind proactive recovery. We present PRISM, \emphProactive Recovery via Interpretable State Modeling, a framework for embodied multimedia agents that formulates proactive recovery as multimedia state modeling followed by a structured risk--recoverability decision readout. PRISM estimates near-future hazard under non-intervention, organizes interaction states into explicit failure prototypes, ranks recovery actions through prototype-conditioned recoverability, combines risk reduction and action cost into an offline calibrated defer/intervene readout, and attaches retrieval-grounded evidence to the resulting decision. On HRI-SENSE, PRISM supports a complete pipeline spanning hazard prediction, prototype learning, recoverability ranking, intervention, and explanation. On the public Handover Failure Detection benchmark, a shared-feature bridge with lightweight calibration and modest target-side adaptation supports practical cross-dataset adaptation rather than zero-shot semantic transfer. In a preliminary blinded human study, retrieval-grounded explanations are preferred for clarity, trust, and especially debugging usefulness. These results suggest that proactive recovery in embodied multimedia is better framed as interpretable state modeling and evidence-grounded decision making than as disconnected prediction heads or a direct policy classifier.

Abstract:
Knowledge distillation with vision-language models (VLMs) offers a promising solution for training lightweight models under data-scarce conditions. However, most existing approaches typically assume that the knowledge learned by pretrained large VLMs can be effectively transferred to capacity-constrained student models. In practice, these representations are often high-dimensional and semantically entangled, making them difficult to transfer under limited data. In this paper, we propose SPARK, a two-stage distillation framework that explicitly addresses this challenge through a feature-space evolution paradigm. Instead of directly distilling the original VLM representations, SPARK first restructures the teachers semantic space into compact and student friendly representations via learnable context parameterization and an entropy-aware adaptive distillation mechanism. It then enables the student to further adapt this semantic space conditioned on its own visual features, while preserving the structure of the textual feature space through a subspace-level constraint. Extensive experiments across multiple datasets and architectures demonstrate that SPARK consistently improves student performance in low-data regimes, achieving significant gains over existing distillation approaches.

Abstract:
Human cognition is fundamentally driven by a bilateral synergy loop, where multimodal understanding (MMU) and generation interleave and reinforce each other. This observation motivates Unified Multimodal Models (UMMs) to foster bilateral synergy across multiple dimensions, such as data and architecture. However, existing methods typically explore these dimensions in isolation and remain confined to limited training stages, leaving the hierarchical role of interleaving in advancing holistic bilateral synergy largely underexplored. To this end, we propose SAIL-Unify, a new UMM that systematically explores the synergistic gains of hierarchical interleaving across data, architectural design, learning strategies, and shared policy optimization throughout the full lifecycle of pre-training and post-training. Specifically, we first unify data formats and establish dual-track pipelines for rigorous filtering and new data construction, yielding over 179 million samples. Architecturally, we introduce the Cross-Level Adaptive Modulation (CLAM) Component within the native UMM paradigm to better harmonize bilateral representational flows. To stabilize task synergy, we further propose a curriculum-like data arrangement strategy coupled with a progressive training recipe, aligning task difficulty with model capacity for robust convergence. Moreover, we develop a new shared policy optimization framework established through a reciprocal causal loop, which mitigates gradient mismatch in the unified RL space and more fully unlocks the potential of pre-training. Extensive quantitative and qualitative experiments across nine benchmarks demonstrate the effectiveness of proposed SAIL-Unify in advancing both MMU and visual generation. Ablation studies further shed light on the crucial factors behind its performance.

Abstract:
Remote sensing visual grounding (RSVG) aims to localize objects in remote sensing imagery according to natural language expressions. Previous methods typically rely on sentence-level vision-language alignment, which struggles to exploit fine-grained linguistic cues, such as spatial relations and object attributes, that are crucial for distinguishing objects with similar characteristics. Importantly, these cues play distinct roles across different grounding stages and should be leveraged accordingly to provide more explicit guidance. In this work, we propose ProVG, a novel RSVG framework that improves localization accuracy by decoupling language expressions into global context, spatial relations, and object attributes. To integrate these linguistic cues, ProVG employs a simple yet effective progressive cross-modal modulator, which dynamically modulates visual attention through a survey-locate-verify scheme, enabling coarse-to-fine vision-language alignment. In addition, ProVG incorporates a cross-scale fusion module to mitigate the large-scale variations in remote sensing imagery, along with a language-guided calibration decoder to refine cross-modal alignment during prediction. A unified multi-task head further enables ProVG to support both referring expression comprehension and segmentation tasks. Extensive experiments on two benchmarks, i.e., RRSIS-D and RISBench, demonstrate that ProVG consistently outperforms existing methods, achieving new state-of-the-art performance.

Abstract:
Large-model-based point-to-CAD generation holds immense potential for advancing industrial design and enhancing 3D modeling efficiency. However, most existing methods approach the problem as a general point-cloud encoding and token prediction task, neglecting the tokenization and supervision specifically for CAD-related primitives. As a result, these methods often struggle to accurately reconstruct the intricate primitive structures. To address this limitation, we propose PrimitiveCAD, a novel multi-stage paradigm for point-to-CAD reconstruction that enhances the geometric accuracy of generated CAD models while better preserving critical geometric features. First, we introduce a primitive-aware point cloud tokenization model, enabling the system to learn more robust geometric representations from CAD point clouds. Next, we perform supervised finetuning on a large language model (LLM) and introduce an operation alignment loss to align key CAD operation frequencies, thereby improving the preservation of global shape features. Finally, we incorporate reinforcement learning (RL) and introduce a feature-line alignment reward to further reduce stochasticity and enhance the fine-grained preservation of geometric features. Experiments on the DeepCAD and Fusion360 datasets show that our method achieves state-of-the-art performance in code validity, geometric accuracy, and geometric feature preservation.

Abstract:
This paper presents SurfelSoup, an end-to-end learned surface-based framework for point cloud geometry compression, with surface-structured primitives for representation. It proposes a probabilistic surface representation, pSurfel, which models local point occupancies using a bounded generalized Gaussian distribution. In addition, the pSurfels are organized into an octree-like hierarchy, pSurfelTree, with a Tree Decision module that adaptively terminates the tree subdivision for rate-distortion optimal Surfel granularity selection. This formulation avoids redundant point-wise compression in smooth regions and produces compact yet smooth surface reconstructions. Experimental results show consistent gain on geometry compression over voxel-based baselines and MPEG standard G-PCC-GesTM-TriSoup across diverse datasets, while providing visually superior reconstructions with smooth and coherent surface structures.

Abstract:
Incomplete multi-view clustering methods recover missing information to restore inherent structure relationships between data, gaining satisfactory performance on pattern mining. However, there still exist two issues that inevitably induce biased imputations of missing information. Those methods learn cross-view generation mappings solely guided by inductive biases on available complete samples for information recovery, which may cause imputation data to deviate from the true distribution of incomplete data. Moreover, most of them leverage sample-level structure alignments to calibrate semantics of imputation data, ignoring inter-view fruitful structures in neighborhoods and clusters. Therefore, a bottom-up consistent structure guided flow generation scheme (TRUE) is proposed via defining an ordinary differential equation of information recovery to underpin end-to-end incomplete data clustering. Specifically, TRUE establishes a cross-view flow matching architecture that leverages a time-dependent vector field to construct the inter-distribution optimal transport path of information recovery, which steers data-driven recovery processes with inter-view continuous flow mappings to realize unbiased imputations faithful to the true distribution of incomplete data. Meanwhile, a multi-hierarchical structure constraint is proposed via defining a bottom-up semantic calibration mechanism to purify the optimal transport path of information recovery from the perspectives of samples, neighborhoods, and clusters, capturing intra-sample and inter-sample structure consistencies to endow imputation data with accurate semantics. Experiments on 5 benchmarks confirm the superiority of TRUE against 9 baselines.

Abstract:
Continual Multi-view Clustering (CMVC) tackles the realistic challenge where data views sequentially arrive and historical data becomes inaccessible. However, existing methods struggle to balance the Stability-Plasticity Dilemma, typically failing in two aspects: (1) Poor Adaptability. They struggle to learn discriminative knowledge from new views with significant domain shifts; (2) Catastrophic Forgetting. They rely on short-sighted knowledge transfer (e.g., from the immediate previous view), leading to error accumulation and loss of global structure. To address these issues, we propose a novel Deep Dual-Stream Distillation framework for Continual multi-view Clustering (D3C). Specifically, we design an expert model equipped with Kolmogorov-Arnold Networks (KAN) for promoting view adaptability. Guided by a history prior-similarity contrastive learning mechanism, this expert aggressively explores the intrinsic structure of the current view. To mitigate knowledge forgetting, we introduce a global consensus voting mechanism. Instead of relying on the noisy knowledge from the last view, we maintain a knowledge bank to aggregate voting from all historical views, generating a robust consensus matrix. Finally, a stable base model absorbs new knowledge from the expert via margin-aware distillation loss while aligning itself to the global consensus through cross-view structure alignment. Extensive experiments on multi-view, cross-domain, and multi-modal datasets demonstrate that D3C achieves state-of-the-art performance, effectively adapting to new views while retaining historical knowledge.

Abstract:
Large Vision-Language Models (VLMs) are increasingly deployed in open-ended environments, where ensuring reliable safety under multimodal inputs is critical. However, existing evaluations remain largely instruction-centric, focusing on explicit malicious queries while overlooking a more realistic and underexplored risk: whether safety alignment remains robust under harmful contextual exposure. This limitation is particularly important for multimodal systems, where visual inputs can substantially steer model behavior and render text-only auditing insufficient. In this work, we study multimodal safety auditing under harmful contextual exposure, asking whether VLMs can maintain safe behavior when partial toxic text is paired with visual context. To enable systematic auditing, we propose RedDiffuser (RedDiff), a reinforcement-based framework that leverages diffusion models to generate semantically coherent visual inputs for black-box safety testing. By combining greedy prompt search with reinforcement optimization, RedDiffuser uncovers high-risk multimodal inputs that expose latent safety failures. Extensive experiments on both open-source and commercial VLMs show that such context-conditioned failures are widespread. On LLaVA, RedDiffuser increases unsafe response rates by up to 10.69% on the original set and 8.91% on a hold-out set, with strong transferability to Gemini and LLaMA-Vision. These vulnerabilities persist even under external safety guardrails, suggesting that current system-level safety mechanisms remain insufficient for realistic multimodal risks. Our findings reveal a critical blind spot in existing safety evaluations and establish context-aware multimodal auditing as an essential paradigm for diagnosing hidden vulnerabilities in modern VLM systems.

Abstract:
We present a novel energy-based continuous repulsion field network, CRFNet, to address the challenging issues in simulating the physical dynamics realism of 3D clothing with temporal and large body movements. A fully differentiable physical loss is proposed to capture the garment dynamics following Newton's Second Law of dynamics and maintain energy consistency, generating physical dynamics realism and enforcing physical interpretability, which has been overlooked by the state-of-the-art 3D clothing dynamics methods. The novel energy-based continuous repulsion field is designed to handle penetrations by penalising material interference between layers via elastic energy. For non-interpenetration regions, the repulsion serves as a preventive mechanism by introducing a distance-aware repulsion force that smoothly increases as the gap decreases. CRFNet is designed under physically-based dynamic modelling principles. As a result, it is not only capable of capturing garment dynamics in a physically realistic and interpretable manner through an elastic energy formulation, but also of generating real-world garments and garments with diverse material properties. Our experiments show that CRFNet reduces interpenetration and deep penetration severity in contact-rich sequences while producing garment dynamics that are physically realistic and remain competitive in deformation quality and temporal smoothness. It also has strong performance for unseen 3D real-world garments that are not trained on.

Abstract:
Large Vision-Language Models (LVLMs) have reshaped visual understanding, yet they remain brittle for Context-Based Emotion Recognition (CBER), where fine-grained facial evidence can be overwhelmed by salient scene context. Facial Action Units (AUs) offer objective micro-expression cues, but static AU injection ignores that emotion categories depend on facial evidence unevenly and can introduce visual noise. We propose Adaptive Action Unit Injection (AAUI), a training-free framework that couples coarse scene understanding with filtered facial priors. AAUI first builds an Emotion-Dependent Strategy Map (?) on the validation set to select category-specific AU thresholds, then performs online coarse-to-fine inference: an LVLM hypothesis triggers strictly filtered AU descriptions to refine the final prediction. On EMOTIC, AAUI reaches 40.32% mAP in a zero-shot setting, improving long-tail categories such as Pain and Embarrassment by mitigating context bias and hallucination.

Abstract:
Multimodal Large Language Models (MLLMs) provide strong semantic reasoning capabilities for embodied agents, yet often struggle to ground language in 3D scene structure and spatial relationships. This limitation affects both Vision-and-Language Navigation (VLN), which requires action decisions from partial observations, and memory-based Embodied Question Answering (EQA), which requires aggregating spatial evidence across trajectories. We propose Hierarchical Dual-Teacher Distillation (HDTD), a local-to-global spatial distillation framework that transfers geometric priors from a frozen 3D vision foundation model into an image-based MLLM. HDTD derives complementary supervision from the frame-wise and global attention blocks of the same frozen VGGT backbone. The local branch aligns fine-grained frame-level geometry, while the global branch preserves cross-view relational structure, enabling the student to capture both immediate spatial cues and trajectory-level relations. Experiments on R2R-CE, RxR-CE, and Open-EQA show consistent improvements over strong MLLM baselines across navigation and embodied question answering tasks, demonstrating that hierarchical local-to-global spatial distillation strengthens embodied spatial reasoning.

Abstract:
Pose-free dynamic 3D Gaussian Splatting suffers from a coupled explanation ambiguity: under photometric supervision, camera motion and non-rigid deformation compete to explain the same image variation, causing rigid motion to be spuriously absorbed into the deformation field. This failure stems less from insufficient representational capacity than from unstable joint optimization induced by weak geometric bias and strong temporal parameter coupling. To address this issue, we introduce a framework that first biases optimization toward geometry-consistent rigid motion through identity-centered pose learning, and then models residual non-rigid dynamics using a piecewise orthogonal temporal representation with reduced coefficient coupling. We further adopt a rigid-to-deformable curriculum that reduces early-stage ambiguity in joint optimization. Together, these designs suppress trajectory jitter and deformation leakage, leading to state-of-the-art average reconstruction performance in pose-free settings on the NVIDIA Dynamic Scenes benchmark.

Abstract:
With the rapid development of 3D Gaussian Splatting (3DGS), physical world scenes can now be virtualized with highly photorealistic quality, enabling promising interactive virtual walkthrough for immersive interaction. Yet, unlike existing object- or courtyard-scale volumetric streaming, networking support for large-scale 3DGS scenes remains largely unexplored and challenging. In this work, we conduct a comprehensive analysis of the issues and opportunities associated with large-scale 3DGS scene streaming. We accordingly present Toruk, an end-to-end spatially adaptive 3DGS streaming system for photorealistic interactive walkthrough. To address key challenges in highly non-atomic Gaussian primitives streaming, content prefetching under low viewport prediction accuracy, and spatially adaptive streaming with Schrodinger's buffer, we propose a localglobal crossed 3DGS segmentation and layering representation, a scene-aware viewport prediction pipeline, and a cross-temporal and inter-spatial causal sequence model. We further explore practical considerations for system development and deployment, and implement a working prototype to validate the superior performance of Toruk, achieving a 2.2 dB improvement in PSNR, a 67.5% reduction in download timeout penalty, and a 36.3% improvement in user experience on average.

Abstract:
Enabling artificial intelligence to transcend surface-level visual perception and achieve deep cultural-semantic understanding of Chinese figure painting is a core open challenge in Cultural Heritage Computing. Although multimodal large language models (MLLMs) can generate fluent art commentary, they often produce ungrounded cultural interpretations during cross-level cognitive transitions, including iconographic-level confusion and contextual misattribution. We use cultural hallucination as an operational umbrella encompassing these documented failure patterns. Existing retrieval-augmented generation (RAG) approaches, moreover, lack stage-structured reasoning constraints and domain-specific ontological support. We therefore present TCFP-Sight (Traditional Chinese Figure Painting Sight), an ontology-gated multi-stage reasoning framework for auditable appreciation of Chinese figure painting. The system is grounded in a structured prior knowledge space integrating Panofskian iconology with Chinese painting theory into a computable semantic schema, instantiated across a knowledge graph of 1,323 annotated figure paintings spanning the Tang through Qing dynasties. An Ontology Gating mechanism prevents premature activation of high-level cultural meaning before low-level evidential foundations are secured, mandating that all output claims carry explicit evidence links and confidence scores. Evaluated on a newly constructed benchmark of 100 traditional Chinese figure paintings split between historical In-KG and modern/traditional Out-of-KG works, TCFP-Sight achieves consistent gains over generic MLLMs and RAG baselines across appreciation quality and hallucination suppression dimensions, while improving evidence traceability. Ablation studies further confirm that Ontology Gating is the most critical anti-hallucination component, and that the staged reasoning pipeline is an indispensable complement to domain knowledge.

Abstract:
\beginabstract Camouflage image generation (CIG) focuses on generating visually concealed objects that seamlessly blend into their backgrounds. Existing methods typically follow either background-guided paradigms that adapt object appearance via style transfer, or foreground-guided strategies that outpaint surrounding regions conditioned on object features. However, they still suffer from appearance discrepancy and background artifacts. We attribute these limitations to cross-context representation leakage, where object and background cues are entangled in a coupled conditional space, resulting in ambiguous control and degraded camouflage fidelity. To tackle this, we propose a new context-decoupled generative paradigm, termed CamoDreamer, which aims to isolate contextual conditional guidance and explicitly decouple latent camouflage features into coordinated object and background control streams. First, a Contrast-aware Contextual Bridge is designed to model cross-context discrepancies and construct contrast-aware dual conditional guidance. Second, Context-Decoupled Assimilation Streams are employed to separate generative interactions conditioned on the dual guidance, while facilitating background rendering with target-aware cues in the latent space. Finally, a Frequency-Adaptive Contextual Blend module integrates complementary high-frequency textures and low-frequency structures from decoupled features to improve holistic coherence. Extensive experiments demonstrate that CamoDreamer consistently outperforms existing methods with a substantial margin, while maintaining a relatively lightweight design. \endabstract

Abstract:
Functional magnetic resonance imaging (fMRI) and electroencephalography (EEG) provide complementary views of brain function and can be linked physiologically via neurovascular coupling (NVC), which can be viewed as a multi-stage dynamical process through which neural activity drives metabolic demand, regulates local blood flow, and produces blood-oxygen-level-dependent (BOLD) signal changes. This process exhibits long-range temporal and spatial/network dependencies. Existing graph-based approaches to multimodal neuroimaging are ill-suited to this setting: message-passing GNNs struggle with oversquashing and oversmoothing when modeling long-range dependencies. To address these limitations, we propose AFE, an NVC-inspired fMRI-EEG modeling method with a dynamical system in a graph learning framework. AFE embeds a learnable, selective Autoregressive Moving Average (ARMA) system into a graph neural network, treating message passing as a dynamical process over brain-graph sequences to capture long-window convolutional effects such as the hemodynamic response function and multi-stage lags. By aggregating historical states and residuals, AFE enhances spatial long-range coupling. Experiments on BrainLat and EMBARC datasets show that AFE achieves competitive performance compared with state-of-the-art multimodal brain imaging methods.

Abstract:
Reference-based object composition becomes particularly challenging in cross-domain scenarios, where models must preserve the reference object's identity while harmonizing it to match stylized environments. This under-explored problem is currently split between practical "blenders" that lack generative fidelity and "generators" that require impractical, per-subject online finetuning. In this work, we introduce Insert In Style, the first zero-shot generative framework that is both practical and high-fidelity. Our core contribution is a unified framework with three key innovations: (i) a novel multi-stage training protocol that disentangles representations for identity, style, and composition; (ii) a specialized masked-attention architecture that surgically enforces this disentanglement during generation; and (iii) complementary prior-preservation mechanisms that keep learned identity and style priors intact. By design, this approach mitigates concept interference typical in unified-attention architectures while ensuring robust generalization across diverse references and styles. To power this framework, we further introduce a scalable data-generation pipeline that couples large-scale, multi-method generation with iterative, human-in-the-loop filtering, yielding a 115k-sample cross-domain corpus calibrated on human annotations for both identity preservation and style coherence. On public cross-domain benchmarks and our new Insert In Style Bench, Insert In Style consistently leads in joint identity and style performance, among existing methods in quantitative, qualitative, and human evaluations, while remaining competitive on in-domain, photorealistic composition.

Abstract:
Deepfake detection is a widely researched topic that is crucial for combating the spread of malicious content, with existing methods mainly modeling the problem as a classification task or a localization task. The rapid advancements in generative models impose new demands on Deepfake detection. In this paper, we propose multimodal alignment and reinforcement for explainable Deepfake detection via vision-language models, termed MARE, which aims to enhance the accuracy and reliability of large vision-language models (VLMs) in Deepfake detection and reasoning. Specifically, MARE designs comprehensive reward functions, incorporating reinforcement learning from human feedback (RLHF), to incentivize the generation of text-spatially aligned reasoning content that adheres to human preferences. Besides, MARE introduces a forgery disentanglement module to capture intrinsic forgery traces from high-level facial semantics, thereby improving its authenticity detection capability. We conduct thorough evaluations on the reasoning content generated by MARE. Both quantitative and qualitative experimental results demonstrate that MARE achieves state-of-the-art performance in terms of accuracy and reliability.

Abstract:
Existing Security-oriented Video Understanding (SVU) systems assume a closed world, i.e., static category sets, abundant labels, and the premise that all event types are known upfront. Real-world security-critical events break these assumptions: they follow long-tailed distributions, new types emerge continuously, and critical security events may offer only a few samples. We formalize this gap as Lifelong Evolving Task for Long-Tailed Security-Critical Events in Videos (L-SCE), a new task that requires VLMs to continually classify and temporally localize newly emerging security-critical events from scarce samples without forgetting previously learned events. Furthermore, L-SCE reveals two critical challenges: (1) Intra-Event Scarcity, where extreme data scarcity may weaken both classification and temporal localization for new events, and (2) Inter-Event Interference, where cross-event feature entanglement and representation drift may strengthen catastrophic forgetting. On this basis, we propose EvoSherlock, a causal-enhanced approach orchestrated end-to-end by an Agentic Controller with self-reflective closed-loop control, which includes two core modules: the Intra-Event Causal Video Generation module (CVG) and the Inter-Event Causal Decoupling and Alignment module (CDA), to address the above two challenges, respectively. Especially, this paper constructs a L-SCE dataset to simulate real-world incremental conditions. Extensive experiments on our benchmark demonstrate the advantages of EvoSherlock over several advanced baselines. These justify the importance of the proposed L-SCE and the effectiveness of EvoSherlock in classifying and temporally localizing emerging security-critical events from scarce samples.

Abstract:
Recent advances in sign language (SL) research have shown a trend toward unifying multiple sign language understanding (SLU) subtasks, such as isolated sign language recognition (ISLR), continuous sign language recognition (CSLR), and sign language translation (SLT), within a single framework, leading to substantial progress. Meanwhile, sign language production (SLP), which generates sign sequences from text, has also attracted growing attention. This naturally raises an important question: can sign language understanding and production be unified within a single framework? Compared with unifying SLU subtasks, this problem is substantially more challenging. Existing SLU tasks largely share the same direction of mapping, namely from sign inputs to linguistic outputs, whereas SLT and SLP lie in opposite directions of sign-text mapping. A unified framework must therefore address two key challenges: (1) bridging the modality gap between continuous sign motions and discrete text tokens through a shared sign tokenizer that supports both linguistic abstraction and motion reconstruction; and (2) learning a single conditional autoregressive model that can take either sign or text as input and generate the corresponding target sequence in the opposite modality. To this end, we propose Uni-SLTP, a unified framework for SLT and SLP with two key components: (1) a shared sign tokenizer that converts sign sequences into discrete tokens and latent representations, capturing both semantic and reconstructive information; and (2) a unified autoregressive generation model that formulates both tasks as conditional sequence generation. Experiments on widely used public datasets show that Uni-SLTP achieves superior motion accuracy for SLP while maintaining competitive SLT performance.

Abstract:
As AI agents are increasingly deployed in real-world applications and take on social roles, dynamic social intelligence, the ability to continuously perceive multimodal cues, infer latent cognitive states, intervene through embodied actions, and ultimately achieve social goals, has become essential. However, existing evaluations largely fragment social intelligence into text-based abstractions, static reasoning tasks, or embodied simulators that lack social dynamics. These approaches fail to capture the multimodal, continuous, and interpersonal complexity of authentic human interactions. To bridge this gap, we introduce ALIVE-World, an embodied simulation environment supporting dynamic multimodal interactions encompassing egocentric vision, language, and non-verbal action. To provide socially grounded interactive feedback, we populate this environment with counterpart agents driven by a profile-conditioned, dual-system (intuitive-affective and logical-cognitive) architecture. Building upon this platform, we propose ALIVE-Bench to transform the assessment of social intelligence into a trackable quantitative framework. By deeply leveraging underlying environmental states and agents' internal variables, ALIVE-Bench systematically evaluates Vision-Language Model (VLM) agents across four critical dimensions: perception, understanding, influence, and outcome. Extensive evaluations reveal that dynamic social intelligence is fundamentally multi-dimensional and scenario-dependent. We identify a critical bottleneck between upstream perception and downstream social intervention, exposing VLM vulnerabilities in fine-grained proxemic awareness, adaptive intervention, and multi-party coordination. By uncovering these limitations and the diverse interaction strategies they produce, this benchmark provides actionable insights for the future behavioral alignment of multimodal agents.

Abstract:
Camera motion often reflects directorial intent and requires professional equipment, making it a high value form of intellectual property. However, generative video models can imitate such high value camera motions with simple prompts, while existing similarity detection methods mainly operate on visual content and fail to capture deeper motion similarity. This is mainly because their training data entangles camera motion with visual content. Moreover, traditional optical flow is insufficient to represent complex camera motions. We therefore build the first benchmark for camera motion analysis, including a motion dataset with 11 motion styles and evaluation protocols. Furthermore, we propose a motion representation that augments optical flow with vorticity cues from fluid dynamics, thereby better capturing motions. Experiments show that our detector achieves a 3.02 improvement in plagiarism detection over the strongest baseline and remains effective on generative videos. We believe our work extends copyright protection beyond static content to dynamic camera motion.

Abstract:
Recent advances in video generation have enabled the synthesis of videos with strong temporal consistency and impressive visual quality. As these perceptual capabilities continue to improve, the next critical challenge toward vision foundation models is to equip models with more robust reasoning abilities for generating coherent and logically consistent video dynamics. Recently, several benchmarks have been proposed to assess the reasoning abilities of video generation models, most of which primarily focus on visual transformations and physics-aware reasoning (e.g., gravity, fluid dynamics), while largely lacking a human-centric perspective in their evaluations. To address this gap, we introduce RULER-Bench, a benchmark grounded in a cognitive rule-based formulation. Unlike prior work, we characterize reasoning as the ability of models to infer and reliably execute human-centric logic across diverse domains. Built upon both text-to-video and image-to-video paradigms, RULER-Bench encompasses 25 representative tasks spanning five rule categories with 519 high-quality annotated instances. For robust assessment, we propose a multi-dimensional evaluation checklist covering four metrics: Rule Coherence, Instruction Following, Visual Consistency, and Visual Fidelity. By leveraging GPT-o3-based scoring for each question, our evaluation protocol achieves a PLCC score of 0.8042 with human judgments. Extensive experiments reveal a stark "reasoning gap": the state-of-the-art model achieves only 43.78% on the Rule Coherence metric, highlighting significant room for improvement in the human-centric reasoning capability of next-generation video models. We expect that the insights obtained from RULER-Bench will facilitate further development of reasoning-aware video generation, advancing video generation models toward vision foundation intelligence.

Abstract:
Current one-pass 3D scene synthesis methods often suffer from spatial hallucinations, such as collisions, due to a lack of deliberative reasoning. To bridge this gap, we introduce SceneReVis, a vision-grounded self-reflection framework that employs an iterative ``diagnose-and-act'' loop to explicitly intercept and resolve spatial conflicts using multi-modal feedback. To support this step-wise paradigm, we construct SceneChain-12k, a large-scale dataset of causal construction trajectories derived through a novel reverse engineering pipeline. We further propose a two-stage training recipe that transitions from Supervised Fine-Tuning to Agentic Reinforcement Learning, evolving the model into an active spatial planner. Extensive experiments demonstrate that SceneReVis achieves state-of-the-art performance in high-fidelity generation and goal-oriented optimization, with improved transfer to underrepresented room categories in the dataset and promising potential for extension to out-of-distribution room types.

Abstract:
Cross-modal retrieval with noisy labels has drawn significant research interest due to inevitable annotation noise during data collection. Several approaches have been proposed to tackle this challenge and have achieved promising success. However, they still suffer from the unreliability issue. On the one hand, noisy labels can mislead the model into overemphasizing non-discriminative features, resulting in significant errors in estimating noisy class posteriors. On the other hand, the label correction often relies on model predictions obtained from the softmax function, which can be overconfident and unreliable due to the translation invariance property of the conventional softmax. To tackle these challenges, we propose a framework entitled Dynamic Masking Regularization meets Evidential Calibration (DMREC) for cross-modal retrieval with noisy labels. Specifically, a Reliability-guided Dynamic feature Masking (RDM) scheme is designed to explicitly impose regularization, preventing the model from overfitting noisy samples. RDM adaptively regularizes the input data by masking the wrongly high-activated features and modulates the label simultaneously. In addition, an Evidential Prediction Calibration (EPC) module based on the Dirichlet distribution is introduced to break softmaxs translation invariance. Extensive experiments demonstrate the significant superiority of our DMREC.

Abstract:
Composed image retrieval (CIR) enables expressive multimodal search: given a reference image and a text modifier, the goal is to retrieve images that match the visual content as altered by the text. Building effective CIR systems has historically required expensive triplet-annotated training data, limiting scalability. Training-free CIR sidesteps this bottleneck by repurposing pre-trained vision-language models to compose queries without any task-specific supervision. However, existing training-free methods commit to a single, globally fixed projection granularity when aligning visual and textual features. This creates an inherent trade-off: coarse granularities capture broad categorical distinctions but lose fine-grained detail, while fine granularities recover instance-level attributes but amplify noise. The problem is especially acute in instance-level CIR, where per-query databases vary from fewer than 10 to over 10,000 images, making any fixed setting suboptimal across the evaluation. We therefore propose a unified retrieval method with two complementary components that fuse scores across multiple granularities to capture complementary evidence and yield more robust rankings. MACE (Multi-granularity Adaptive Composition with Expansion) constructs a spectral ensemble from multiple nested truncation levels and fuses their scores without any additional eigendecomposition or database modification. CGCQE (Cross-Granularity Consensus Query Expansion) integrates into MACE to strengthen pseudo-relevance feedback, selecting expansion candidates that rank consistently well across all granularities and preventing granularity-specific false positives from corrupting the expanded query. Experiments on five CIR benchmarks, including the challenging instance-level i-CIR, show that MACE with CGCQE achieves state-of-the-art performance, outperforming both training-free and supervised zero-shot methods.

Abstract:
Medical imaging models are often deployed without the demographic, acquisition, and quality metadata needed for subgroup auditing. Once those metadata disappear, clinically critical failure modes can be masked by strong aggregate performance, and many robust-learning methods lose the group structure they rely on. We present CAPRA, a calibrated proxy-axis framework for hidden subgroup analysis under missing metadata. CAPRA predicts image-derived semantic axes, calibrates axis posteriors on a small metadata-labeled split via patient-level cross-fitting, and organizes those posteriors into a calibrated subgroup interface that supports both deployment-time failure analysis and downstream robust learning without requiring subgroup labels at deployment. Across fundus, dermoscopy, and chest radiography, CAPRA reveals disparity patterns missed by metadata-only slicing, remains informative under dataset shift, and produces subgroup partitions that align more closely with explicit failure axes than image-only or latent-slice baselines. The same interface can also be reused by downstream robust learners, although those gains are domain-dependent. Overall, CAPRA turns hidden subgroup analysis under missing metadata into a calibrated, interpretable, and reusable subgroup interface for deployment-time analysis and robust transfer.

Abstract:
Category-level 6D object pose and size estimation remains challenging due to severe intra-class geometric variation, which makes direct absolute prediction unreliable for unseen instances. In this paper, we reformulate the task as a similarity-driven retrieve-and-refine problem, where pose estimation is grounded in geometry-semantic analogies rather than one-shot regression. To this end, we propose a geometry-language retrieval stage that first retrieves a small set of pose-valid and semantically consistent exemplars, and then performs candidate-conditioned residual refinement for final prediction. Specifically, we introduce a Siamese Geometry Retrieval Network (SGRNet) with a pose-and-shape-aware metric objective to learn retrieval embeddings aligned with downstream pose validity. We further incorporate language descriptors derived from a multimodal foundation model to regularize retrieval under ambiguous geometry and partial observations. Given the retrieved candidates, a Residual Pose Estimation Network (RPENet) predicts pose and size residuals relative to each candidate, converting the original ill-posed global estimation problem into a set of better-conditioned local refinement problems. Across both stages, we design a bidirectional depth-point fusion module to couple view-dependent depth cues with global 3D shape structure. Extensive experiments on REAL275, CAMERA25, and Omni6DPose demonstrate that the proposed formulation achieves strong generalization and state-of-the-art performance, with additional transferability to instance-level estimation on YCB-Video.

Abstract:
Reliable robotic mapping in dynamic real-world environments is essential for embodied intelligence and long-term autonomy. However, existing methods often involve a trade-off: lightweight online mapping is efficient but captures limited semantic context, while dense mapping provides richer semantics at higher computational cost and is more vulnerable to distractors. They also struggle to relocalize objects after scene changes. To address these challenges, we propose the Online Anchor-centric Scene Inference System (OASIS), a unified system for online mapping, complex open-vocabulary retrieval, and navigation in dynamic environments. Motivated by the observation that indoor scenes are organized around stable semantic anchors, OASIS represents the environment using a sparse set of anchor nodes, each maintaining a unified multimodal representation that integrates semantic descriptions, visual features, and geometric states for reasoning, verification, and navigation. Its compact anchor-centric memory supports on-demand evidence extraction and hierarchical retrieval, reducing redundant storage while improving retrieval reliability. To handle scene dynamics, OASIS further reuses transient observations collected during navigation to relocalize moved objects. Experiments in both simulation and real-world environments, including standard retrieval benchmarks and navigation tasks under both static and dynamic settings, demonstrate that our system can perform complex open-vocabulary retrieval and navigation, while also enabling relocation of displaced objects after scene changes.

Abstract:
Reconstructing 4D objects from casually captured monocular videos remains challenging due to severe viewpoint ambiguity, occlusion, and incomplete temporal observations, which make dynamic geometry and appearance recovery highly under-constrained. Existing methods often rely on category-specific templates, weakly structured dynamic representations, or free-form deformation, and therefore struggle to maintain surface continuity, structural plausibility, and temporal coherence under complex dynamics. Moreover, directly applying generic diffusion priors to this task often leads to suboptimal reconstruction, as such priors lack awareness of the underlying motion structure. To this end, we present Skel4D, a skeleton-driven framework for 4D object reconstruction from casual monocular videos. Specifically, we first construct a structured Mesh-Gaussian representation by constraining Gaussians onto a deformable mesh, together with a cross-triangle transition design that adaptively reallocates Gaussians according to local surface variations. We then introduce a skeleton-driven deformation module initialized from a coarse skeleton prior estimated by a foundation model, along with an adaptive modulation term to mitigate skeleton noise and a residual non-rigid deformation module for fine-grained Gaussian refinement. Finally, we enhance diffusion-based supervision with skeletal conditioning and recursive temporal propagation, yielding more reliable and structurally consistent supervision over dynamic sequences. Extensive qualitative and quantitative experiments show that Skel4D achieves robust and high-quality 4D reconstruction on challenging dynamic scenes, while naturally supporting motion editing.

Abstract:
Generative world models have demonstrated strong capability in modeling complex environment dynamics. However, existing approaches mainly rely on data pattern matching or physics-informed generation with externally given physical priors, failing to learn the intrinsic and coupled physical laws. Those challenges are particularly severe in fluid environments, where governing physical laws are implicit and jointly determined by multiple related physical factors. To solve these challenges, we propose the Physics Attention World Model (PAWM), a physics-learned generative world model that autonomously learns and decorrelated informs latent physics from raw fluid observations. Specifically, PAWM comprises two core modules: i) the Physics Attention module, which learns disentangled latent physics and achieves physics-based consistent constraints, and ii) the Decorrelated Physics-informed Prediction module, which first conducts decorrelation reweighting to quantify the contribution of different physical factors in video generation with theoretical support, then performs state transition and prediction conditioned on these learned and decorrelated latent physics. Extensive experiments on three fluid benchmarks demonstrate that PAWM significantly outperforms existing generative world models with respect to accurate and physically consistent prediction across both simulated and real-world fluid scenarios, as well as maintaining strong generalization across diverse fluid properties.

Abstract:
Text-to-video (T2V) generation models have achieved strong visual realism, but improving physical plausibility can come at the cost of semantic consistency with the input text. This tension arises because physical preference is typically determined by comparing dynamics between two videos, without accounting for whether either video faithfully depicts the scene specified by the prompt, making physical-semantic conflict a systematic tendency under this supervision paradigm. We formulate this challenge as a constrained preference optimization problem and propose Physical and Semantic Direct Preference Optimization (PSDPO), which modulates each preference pair's contribution based on the agreement between its physical and semantic signals. A gradient-level analysis shows that PSDPO bounds the semantic drift from conflicting pairs to a controllable residual, and further motivates a staged optimization protocol that provably reduces cumulative drift. The resulting method operates entirely within the standard DPO framework, requiring no auxiliary models or additional loss terms. Experiments show that PSDPO improves physical plausibility by up to 2× over the baseline on VideoPhy-2, while maintaining strong semantic consistency on VBench, achieving a more reliable balance than existing preference-based methods.

Abstract:
In autonomous driving, 3D LiDAR semantic segmentation is expected to adapt to evolving environments and recognize previously unseen categories, making generalized category discovery under test-time adaptation (GCD-TTA) an important problem. However, existing works treat GCD and TTA separately, while the limited semantic information in 3D point clouds makes GCD-TTA under domain shift more prone to feature collapse when combining GCD with TTA. To address this issue, we propose a two-stage framework for GCD-TTA in 3D LiDAR semantic segmentation. Specifically, in the first stage, we adaptively aggregate local 3D features according to geometric density variations and obtain reliable pseudo labels under a global class-wise prototype consistency constraint. In the second stage, we cast category discovery as an optimal transport assignment process for globally consistent grouping of novel categories. Meanwhile, semantic features from 2D vision foundation models are fused with 3D point cloud features to enrich 3D semantic information and promote a more separable feature space. Extensive experiments on Synth4D to SemanticKITTI, SynLiDAR to SemanticKITTI, and Synth4D to nuScenes demonstrate the effectiveness of our framework. Under a unified evaluation protocol, our method achieves the best overall mIoU on 9 out of 10 benchmark splits and improves the overall mIoU by up to 12.21 points over the strongest re-implemented baseline.

Abstract:
Radar, with its all-weather sensing capability, serves as a core component for robust 3D scene motion perception in autonomous driving. Due to its lack of texture information and noisy measurements, radar requires complementary visual information for fine-grained motion perception. However, current radarcamera fusion methods still suffer from degraded visual features in challenging environments and excessively sparse radar representation that lacks sufficient information. To address these limitations, we propose RCSFlow, a novel framework that fuses raw 4D radar echoes with a camera for accurate motion perception across various driving conditions. We design a Radar Block Feature Encoder (RBFE) for efficient radar feature representation to mitigate redundancy in raw radar echoes, and a Feature Confidence Modulation Mechanism (FCMM) to selectively modulate visual and radar features during radarcamera fusion. To capture 3D motion cues across adjacent frames, we introduce a bidirectional BEV-to-3D motion correlation architecture that leverages a BEV-anchored irregular 3D cost volume to produce a stable and fine-grained 3D motion field, and a 3D-guided BEV aggregation scheme to maintain memory efficiency. Extensive experiments demonstrate that RCSFlow achieves state-of-the-art motion perception performance while remaining robust under adverse driving conditions.

Abstract:
Camera localization in bronchoscopy remains a challenging problem due to stringent accuracy requirements, real-time constraints, and limited training data. Compared to natural scenes, the confined anatomical structures demand millimeter-level precision, while intraoperative guidance necessitates low-latency inference. However, existing methods often fail to effectively exploit preoperative geometric priors, limiting their robustness and accuracy. To address these limitations, we propose a unified geometry-aware bronchoscope localization framework (GABL) that effectively fuses preoperative structural priors with paired intraoperative video to estimate 6-DoF camera poses. Specifically, to address visual ambiguity in complex airways, we propose a graph-guided coarse-to-fine localization scheme that effectively leverages structural priors for precise pose estimation. Furthermore, to mitigate pose jitter and bridge the visual-structural gap, we integrate a Transformer-based tracking model with a novel RGB-depth matching objective, jointly enforcing spatio-temporal and geometric consistency. Extensive experiments demonstrate that our method yields remarkable reductions of 8.37% and 31.76% in translation and rotation errors over the prior state-of-the-art, alongside 4 times inference speedup (33.6 FPS) for robust real-time bronchoscope localization. Project website: https://paulili08.github.io/GABL/.

Abstract:
Video Temporal Grounding (VTG) faces significant challenges when natural language queries must distinguish between multiple events involving visually similar entities, particularly when relying on fine-grained visual attributes that are difficult to describe accurately in words alone. To address this, we introduce Image-Disambiguated Video Temporal Grounding (ID-VTG), a task that leverages multimodal queries combining a reference image and a text description to precisely localize segments where a specific instance performs a described action. To facilitate research, we construct two benchmarks: IDVTG-Gym, focusing on fine-grained, compositionally ordered gymnastics actions with athletes in similar uniforms; and IDVTG-InternVid, an open-world dataset featuring diverse entities (e.g., humans, animals, fictional characters) and significant temporal distractors. Methodologically, we propose the Visually-Guided Disambiguation Aggregation (VGD-Agg) framework based on a dual-branch fast-slow architecture. The fast branch efficiently generates preliminary event proposals, while the slow branch performs fine-grained frame-level matching between video frames and the reference image. We enhance discriminability via two learnable tokens: a Compare Token, which represents hard negatives to probe for the presence of the target instance (as referred to by the query image), and a Depress Value, which represents text-irrelevant events. Proposals that the Compare Token identifies as lacking the target instance are pushed toward the Depress Value, thus easing disambiguation via the text query. Extensive experiments validate our approach, which achieves state-of-the-art results on the proposed benchmarks. Code is available at https://github.com/oceanflowlab/ID-VTG.

Abstract:
Video question answering (VideoQA) requires sophisticated interpretation of spatio-temporal dynamics. While grounded VideoQA has emerged to provide verifiable evidence, existing benchmarks are predominantly confined to Coarse-grained Spatio-Temporal (CST) evidence, such as event or object-level localization. Consequently, they fail to scrutinize the Fine-grained Spatio-Temporal (FST) evidence essential for clarifying subtle semantic states, allowing Multimodal Large Language Models (MLLMs) to exploit language priors instead of performing genuine visual grounding. To address this, we introduce NexusBench, a novel benchmark requiring joint reasoning across Multi-grained Spatio-Temporal evidence. Built upon a Hierarchical Video Scene Graph (HVSG), it explicitly models FST interactions to build logic-intensive queries. To diagnose current model limitations, we propose a metric termed NexusScore, assessing a model's capability to utilize evidence. Our analysis reveals that existing MLLMs exhibit inadequate FST grounding abilities, resulting in suboptimal performance, yet supplying them with the correct evidence drastically improves accuracy. Motivated by this, we present NexusQA, a training-free multi-agent baseline that accumulates evidence through collaborative reasoning. Extensive experiments demonstrate that NexusQA achieves state-of-the-art performance on both the proposed NexusBench and the public NExT-GQA dataset. Our code and benchmark will be released at https://github.com/Daria8976/NexusQA.

Abstract:
All-in-One Image Restoration (AiOIR) aims to recover high-quality images from diverse degradations within a unified framework. However, existing methods often fail to explicitly model degradation types and struggle to adapt their restoration behavior to complex or mixed degradations. To address these issues, we propose ClusIR, a Cluster-Guided Image Restoration framework that explicitly models degradation semantics through learnable clustering and propagates cluster-aware cues across spatial and frequency domains for adaptive restoration. Specifically, ClusIR comprises two key components: a Probabilistic Cluster-Guided Routing Mechanism (PCGRM) and a Degradation-Aware Frequency Modulation Module (DAFMM). The proposed PCGRM disentangles degradation recognition from expert activation, enabling discriminative degradation perception and stable expert routing. Meanwhile, DAFMM leverages the cluster-guided priors to perform adaptive frequency decomposition and targeted modulation, collaboratively refining structural and textural representations for higher restoration fidelity. The cluster-guided synergy seamlessly bridges semantic cues with frequency-domain modulation, empowering ClusIR to attain remarkable restoration results across a wide range of degradations. Extensive experiments on diverse benchmarks validate that ClusIR reaches competitive performance under several scenarios. The code is publicly available at: https://github.com/Husk021118/ClusIR.

Abstract:
Dataset distillation aims to distill the knowledge embedded in the original training set into a compact synthetic dataset, alleviating the reliance on large-scale datasets in modern model training. Dataset parameterization offers a more efficient representation of distilled datasets, reducing redundancy while preserving rich information for model training. This paper proposes GSDD, a novel and effective dataset parameterization technique by adopting Gaussian Splatting as the underlying representation of distilled images. By replacing pixels with Gaussian primitives as the basic units of distilled images, GSDD substantially reduces redundancy while preserving representational flexibility. To support this parameterization, we design efficient rendering operators and explore a series of distillation-oriented Gaussian representation strategies that improve expressiveness, stability, and rendering quality. Our experimental analysis reveals that Gaussian primitives capture meaningful training features, yielding a sparse yet expressive representation that enables GSDD to achieve a better trade-off between efficiency and representational capacity. Beyond achieving competitive results on standard benchmarks, we further demonstrate the effectiveness of GSDD on large-scale and modality-diverse distillation tasks. We also provide comprehensive benchmarks on its computational efficiency, memory footprint, and stability. Code is available at https://github.com/j-cyoung/GSDatasetDistillation.

Abstract:
Recent vision-language models (VLMs) typically rely on a single vision encoder trained with contrastive image-text objectives, such as CLIP-style pretraining. While contrastive encoders are effective for cross-modal alignment and retrieval, self-supervised visual encoders often capture richer dense semantics and exhibit stronger robustness on recognition and understanding tasks. In this work, we investigate how to scale the fusion of these complementary visual representations for vision-language modeling. We propose CoME-VL (Complementary Multi-Encoder Vision-Language), a modular fusion framework that integrates a contrastively trained vision encoder with a self-supervised DINO encoder. Our approach performs representation-level fusion by (i) entropy-guided multi-layer aggregation with orthogonality-constrained projections to reduce redundancy, and (ii) RoPE-enhanced cross-attention to align heterogeneous token grids and produce compact fused visual tokens. The fused tokens can be injected into a decoder-only LLM with minimal changes to standard VLM pipelines. Extensive experiments across diverse vision-language benchmarks demonstrate that CoME-VL consistently outperforms single-encoder baselines. In particular, we observe an average improvement of 4.9% on visual understanding tasks and 5.4% on grounding tasks. Our method achieves state-of-the-art performance on RefCOCO for detection while improving over the baseline by a large margin. Finally, we conduct ablation studies on layer merging, non-redundant feature mixing, and fusion capacity to evaluate how complementary contrastive and self-supervised signals affect VLM performance. Project page: \hrefhttps://mbzuai-oryx.github.io/CoME-VL/\textcolorteal\nolinkurlmbzuai-oryx.github.io/CoME-VL.

Abstract:
Despite the recent advances in layer decomposition, instruction-guided layer editing remains a pivotal yet under-explored challenge in image manipulation. Existing solutions typically rely on decoupled cascaded scheme: compositing layer onto solid background, applying off-the-shelf RGB editor, and then re-matting foregrounds. This way inherently neglects global contextual harmony across layers and struggles to preserve intricate boundary details during matting. To address this, we present LayerEdit, an end-to-end framework for context-aware image layer editing built upon a pre-trained layer decomposition foundation model. Unlike the decoupled cascaded scheme, LayerEdit learns joint layer editing and high-fidelity matting by seamlessly synergizing editing priors from massive RGB data with the intrinsic hierarchical perception embedded in the decomposition backbone. Furthermore, rather than processing the input layer in isolation, LayerEdit amplifies context awareness by integrating the global semantic guidance from the holistic composite into the input layer. Empowered by the introduced RGB-to-RGBA knowledge transfer strategy and the effective utilization of the native RGBA structural representations, our model achieves exceptional layer manipulation within a single network relying on limited curated layer editing data. Extensive experiments demonstrate that LayerEdit outperforms cascaded baselines. \hrefhttps://chenjw666.github.io/layeredit/Project page: https://chenjw666.github.io/layeredit/.

Abstract:
Multimodal Generative Recommendation (MGR) encodes items as discrete semantic identifiers (IDs) via vector quantization and predicts the next item through autoregressive token generation. Conventional multi-modal approaches construct item codes from text and image features through cross-modal RQ-VAE, consistently outperforming single-modality methods. However, existing methods build code spaces purely from content features, overlooking that frequently co-occurring items in user interaction sequences can differ greatly in content. Moreover, autoregressive decoding ranks candidates solely by token-level likelihood, unable to evaluate them in continuous semantic space. We propose Collaborative-Anchor guided multimodal Generative Recommender (CAGRec) to address both challenges. For the content-collaborative gap, we anchor the cross-modal RQ-VAE with collaborative embeddings derived from user interaction sequences, steering both text and image code spaces toward behavioral structure. We further propose Collaborative-Aware Regularization (CAR), a contrastive loss that clusters collaboratively related items in the latent space before code assignment, ensuring behavioral proximity survives quantization. The anchored RQ-VAE decoder gives rise to a collaborative-semantic space that jointly encodes behavioral and content signals. For the discrete-continuous gap, we design a semantic-aware reranking strategy that combines generation likelihood with collaborative-semantic similarity, pulling candidate evaluation back into the continuous space. Extensive experiments on three recommendation datasets show that CAGRec consistently outperforms state-of-the-art baselines. Our codes are available at https://github.com/YYYYXL1004/CAGRec.

Abstract:
Contrastive Language-Image Pre-training (CLIP) exhibits strong zero-shot transferability, but its direct use in semantic segmentation is hindered by two representation issues in vision transformers: concept entanglement and high-norm outlier tokens. Concept entanglement causes patch features to encode mixed semantics, leading to background confusion, while outlier tokens emerge in intermediate layers, accumulate excessively large norms, and progressively corrupt local representations through attention. Existing training-free methods mainly modify the final attention layer with class-agnostic operations, leaving both concept-specific feature disentanglement and early-stage outlier accumulation insufficiently addressed. We propose DisenCLIP, a training-free framework that jointly mitigates these two issues. Specifically, we introduce a dynamic concept-aware channel masking scheme to enhance concept-discriminative dimensions and suppress ambiguous shared channels, together with a soft contrastive masking strategy that produces continuous class-specific channel weights from inter-class feature divergence. We further propose a multi-layer norm-clipping mechanism that performs percentile-based compression on intermediate patch-token norms to prevent progressive feature corruption. Experiments show that DisenCLIP consistently surpasses existing training-free baselines and achieves state-of-the-art performance in both mIoU and boundary quality. The code is available at https://github.com/zixiwz/DisenCLIP.

Abstract:
Scratch Defect Segmentation (SDS) aims to identify scratch defects amidst a product's inherent surface texture with mask prediction. Recent methods have achieved promising results by introducing prompt mechanisms for this task. However, these methods often rely solely on spatial-domain prompts, limiting their performance in handling the high morphological variability of scratches and their high visual similarity to background textures. To address this limitation, we propose SFPSeg, a novel Spatial-Frequency dual-domain prompt learning framework for robust scratch segmentation. To this end, a novel Mixed-domain Prompt Learning Decoder (MPLD) is introduced in SFPSeg to decode complementary and diverse patterns of scratches from both domains. Specifically, a novel Morphology-aware Spatial Prompting (MSP) module is introduced in MPLD to adaptively learn hierarchical morphological details from the spatial domain. Meanwhile, another novel module, namely Wavelet Prompt Filter (WPF), is introduced in MPLD to enhance scratch-related frequency components while suppressing background responses from the frequency domain. Comprehensive experiments on four SDS benchmarks demonstrate that SFPSeg outperforms state-of-the-art methods, achieving an mIoU of 91.16% on NEUSD, 81.40% on Severstal, 91.01% on Car, and 78.33% on Chip. The code is available at https://github.com/ShanWuJ/SFPSeg.

Abstract:
The prompt learning paradigm for vision-language models is effective yet faces a granularity dilemma: global prompts lack fine-grained semantic awareness, while local prompts ignore contextual associations, limiting cross-task generalization. This dilemma exists in dense prediction tasks. Inspired by U-Net, which unifies multi-level representations across granularities, we propose UPrompt, a U-shaped multi-granularity prompt learning framework for vision-language models. Similar to how U-Net integrates fine and coarse features through symmetric encoder-decoder pathways with cross-level connections, UPrompt constructs parallel multi-granularity representations in both visual and textual modalities, where coarse-to-fine cascaded enhancement propagates global context to refine local details, while fine-to-coarse hierarchical supervision ensures semantic consistency across scales. Extensive experiments on 17 benchmarks validate our effectiveness. UPrompt outperforms MAMET and VPKE by 4.1 and 7.3 rSum on MSCOCO, surpasses CoCoA-Mix by 5.09% in base-to-novel generalization, while maintaining competitive performance with minimal overhead (coarse-grained) and matching PSRC with 1/3 cost (medium-grained). Our code is available at https://github.com/JustCoolPig/UPrompt.

Abstract:
Learning 3D human-object interaction aims to accurately estimate interaction poses and human-object contacts, which is essential for applications in embodied AI, AR/VR, and imitation learning. However, most existing HOI methods predominantly focus on single-object scenarios and rely on action-centric representations, which are coarse and non-compositional, limiting their ability to capture compatibility and coupled constraints in multi-object interactions and often resulting in locally plausible but globally inconsistent predictions. In contrast, object attributes grounded in part-level geometric and functional cues provide a more fine-grained, compositional semantic unit that maps objects to required body resources and the induced constraints on contact, alignment, and stability, enabling the composition and coordination of multi-object constraints for coherent whole-body interactions. Building on this insight, we propose ACRM-HOI (Attribute-Centric Reasoning for Human Multi-Object Interaction), an attribute-driven framework that infers per-object attributes through geometric, functional, and usage cues. These attributes are then composed to reason about concurrent multi-object interaction intents and finally grounded in human pose and contact regions to produce coherent whole-body interactions. We further collect AMHOI (Attribute-Based Human Multi-Object Interaction), an attribute-based human multi-object interaction dataset that serves as a testbed for training and evaluation. Experiments demonstrate that our approach produces more accurate interaction regions and more plausible, whole-body-consistent concurrent interactions than prior methods. The source code and dataset are available at \hrefhttps://github.com/soddday/ACRM-HOIhttps://github.com/soddday/ACRM-HOI.

Abstract:
Training-free Vision-Language Navigation (VLN) agents powered by foundation models can follow instructions and explore 3D environments. However, existing approaches rely on greedy frontier selection and passive spatial memory, leading to inefficient behaviors such as local oscillation and redundant revisiting. We argue that this stems from a lack of metacognitive capabilities: the agent cannot monitor its exploration progress, diagnose strategy failures, or adapt accordingly. To address this, we propose MetaNav, a metacognitive navigation agent integrating spatial memory, history-aware planning, and reflective correction. Spatial memory builds a persistent 3D semantic map. History-aware planning penalizes revisiting to improve efficiency. Reflective correction detects stagnation and uses an LLM to generate corrective rules that guide future frontier selection. Experiments on GOAT-Bench, HM3D-OVON, and A-EQA show that MetaNav achieves state-of-the-art performance while reducing VLM queries by 20.7%, demonstrating that metacognitive reasoning significantly improves robustness and efficiency. The code is available at https://github.com/perfectlxy7/MetaNav.

Abstract:
Recent advancements in 3D Gaussian Splatting (3DGS) have enabled photorealistic rendering of complex scenes, yet widespread adoption on mobile and Extended Reality (XR) devices is hindered by substantial computational and bandwidth requirements. While existing solutions often focus on model compression for client-side rendering, they still demand significant GPU power, limiting applicability on resource-constrained hardware. We propose TIGAS (Thin-client Interactive Gaussian Adaptive Streaming), an open modular research testbed for interactive 3DGS streaming that offloads rasterization to a GPU backend and streams view-dependent 2D projections to a lightweight web client over QUIC, minimizing Head-of-Line (HoL) blocking. Researchers can hot-swap Adaptive Bitrate (ABR) algorithms, movement traces, network conditions, and neural enhancement models under controlled conditions. We include a latency-oriented ABR baseline and an optional WebGPU super-resolution (SR) hook as demonstrators of this extensibility, not as primary algorithmic contributions. We evaluate TIGAS across multi-continental environments using 13 3DGS models and real Six Degrees of Freedom (6DoF) EyeNavGS movement traces. With a backend that renders frames in under 10 milliseconds, TIGAS achieves an average SSIM of 0.88 and keeps motion-to-photon latency within interactive thresholds for screen-based 6DoF navigation when clients are within the geographic reach of a single edge server. The source code is available at \urlhttps://github.com/Rekenar/GaussianAdaptiveStreamer.

Abstract:
Despite notable advances in image denoising, existing models often struggle to generalize beyond in-distribution noise patterns, particularly when confronted with out-of-distribution (OOD) conditions characterized by the spatially varying level map. This generalization gap remains a fundamental yet underexplored challenge. In this work, we investigate magnitude equivariance as a core inductive bias for improving OOD robustness. Specifically, incorporating magnitude-equivariant structures enables models to better adapt from training on the uniform noise map to inference on spatially non-uniform degradation intensities. Building on this insight, we propose a robust blind denoising framework equipped with two key components: a Heterogeneous Normalization Module (HNM) and an Interactive Gating Module (IGM). HNM stabilizes feature distributions and dynamically corrects features under varying noise intensities, while IGM facilitates effective information modulation via gated interactions between signal and feature paths. Extensive evaluations demonstrate that our model consistently outperforms state-of-the-art methods on various OOD noises. Furthermore, we extend the applicability of the magnitude-equivariant principle to real-world scenarios via the masked noise translator. The code is released at \urlhttps://github.com/DaweiZhang0304/MEVNet.

Abstract:
Predicting accurate spatiotemporal sequences in complex spatiotemporal systems is a fundamental challenge in machine intelligence. A critical aspect lies where real-world systems always appear critical transitions can be anticipated through early-warning signals. However, existing predictive models primarily learn visual representations from historical sequences, overlooking dynamic transient patterns governing behind system evolutions in several inherent challenges, including physical concealment, causal interpretability and hotspot effect. Thus, we propose a novel predictive framework called transient patterns discovered for spatiotemporal sequence forecasting (TransientCast), seamlessly integrating transient pattern discovery with spatiotemporal predictive learning. In order to quantitatively capture transient patterns, disentangled transient pattern modeling is proposed to extract transient patterns using transient pattern discovery and then disentangled as transient embeddings together with visual embeddings. Interpretable future transient patterns and sequences are jointly predicted by a multivariate causal transformer, ensuring causal dependency. To further enhance the predictive accuracy of identified critical regions, hotspot refinement is designed by focusing on regions with high transient modularity. Extensive experiments on several scenarios show significant improvement in predictive accuracy and plausible predicted future transitions. (https://github.com/miaomiaoToday/predictiveModel.git)

Abstract:
Current image-to-video models achieve visual realism and physical plausibility, but reasoning about mental states remains unexplored. Actions are driven by belief, desire, and perception, requiring inference beyond explicit instructions. We introduce MindWorldBench to evaluate mental-state-conditioned video generation. We formalize this as mental-state-to-behavior reasoning, where models generate actions from a world state and latent variables without explicit action prompts. MindWorldBench utilizes Zero-Action Prompting and a counterfactual design with 744 prompts to isolate the causal effects of mental states. An automated pipeline evaluates video quality, commonsense plausibility, and mental-state consistency. Evaluations of 11 models show that despite visual fidelity and physical reasoning, models fail to align behaviors with latent mental states. We identify a failure mode, termed Omniscient Bias, where models default to the objective world state rather than human's subjective belief. These results demonstrate a disconnect between visual generation and cognitive reasoning, suggesting a need for explicit mental-state modeling in video generation systems. Project website: https://richard2049-lee.github.io/MindWorldBench/

Abstract:
Manchu archives are historically valuable documents formed during the Qing Dynasty in China. However, due to limitations in Manchu word recognition technology, these archives have hardly been exploited and utilized. The primary reason lies in the fact that Manchu language has become critically endangered, as officially recognized by UNESCO in 2009. There are very few individuals who can read Manchu script, and annotating Manchu word images is extremely challenging. Consequently, no large-scale dataset for Manchu archive word recognition has been available to date. In light of this, after more than six years of effort by the authors, this paper presents MAWD15935, the first large-scale Manchu word dataset for word recognition in Manchu archives. The dataset comprises 1,092,744 samples derived from 15,935 Manchu words, including 975,301 printed samples and 117,443 handwritten samples which are sourced from scans of authentic Manchu archives. The dataset features three distinctive strengths: (1) authenticityit encompasses all Manchu vocabulary entries contained in the New Comprehensive Manchu-Chinese Dictionary , the most authoritative reference work with the largest collection of Manchu entries for archival research; (2) reliabilityall data have been annotated by professionals specialized in the Manchu language; and (3) richnessit features a wide variety of words and a large sample size. The contributions of this paper will significantly advance the development of Manchu archive digitization technology and accelerate the process of exploiting and utilizing Manchu archives. The dataset is publicly available at https://github.com/DevilEnfant/MAWD15935.

Abstract:
In this paper, we show for the first time that visual token pruning enhances the robustness of Multimodal Large Language Models (MLLMs), mitigating vulnerabilities such as jailbreak attacks and hallucinations. Given that vision and language modalities cannot be perfectly aligned, the misaligned visual tokens might act as out-of-distribution (OOD) inputs, leading to unpredictable outputs and introducing potential vulnerabilities. Building on this insight, we aim to enhance model robustness against jailbreaks and hallucinations by reducing OOD visual tokens at robust-pruning layers, while also reducing inference cost as a side benefit. Specifically, we measure the distance between each visual token and the language feature space. Then, visual tokens with large distances are identified as OOD tokens, which can be iteratively pruned. To demonstrate the effectiveness of our method, we evaluate it on seven diverse popular benchmarks. Notably, our method yields an average improvement of 13.29% in defending jailbreak attacks, consistently achieves competitive performance in mitigating hallucinations, and maintains strong results on general datasets like MME. Code is available at \urlhttps://github.com/Eurek001/OOD-VTP.

Abstract:
Multi-view clustering provides a natural formulation for exploiting heterogeneous data. In real tasks, however, noise and incompleteness often induce pseudo-label noise, which degenerates clustering performance. Previous studies frequently neglect this inherent noise, compromising robust semantic representation extraction. Although quantifying aleatoric and epistemic uncertainties can effectively identify and suppress unreliable pseudo-labels, current approaches rarely exploit this potential. In this paper, we propose the Evidential Multi-view Clustering with Dual Uncertainty-Aware Pseudo-Labels (EMDUP) approach. EMDUP introduces Evidential Dual-Uncertainty Quantification to encode views into Dirichlet distributions, jointly modeling prediction confidence and dual uncertainties. Then, an Uncertainty-Weighted Pseudo-Label Refinement module adaptively denoises supervision by down-weighting uncertain instances. Finally, a Dual-Correlation Consistency Strategy integrates mutual information maximization and contrastive learning to enforce consistency across feature representations and cluster assignments. Experimental results on six benchmark datasets validate that EMDUP improves clustering performance and robustness against noisy and partial multi-view data. The code is available at https://github.com/norrth-z/EMDUP.

Abstract:
Infants learn word-object associations from continuous first-person experience, where caregiver speech is sparse and often only weakly aligned with what is visible. Computationally, standard contrastive learning is poorly suited to this regime because it treats all frames within an utterance-window as positive matches to the utterance, introducing many false alignments. We present MILCVCL, a noun-centric vision-language framework for learning from child video under weak temporal alignment. MILCVCL replaces frame-level supervision with a bag-of-frames and a multiple instance contrastive objective, where at least one frame in the bag provides relevant visual evidence of the paired utterance. We further examine noun-centric supervision as a developmentally motivated prompt for object-centered grounding. Our work also extends SAYCam-S with roughly 80 hours of additional semi-automatically transcribed speech. Our results show that MILCVCL consistently improves over all baselines across in-domain word learning, four out-of-domain benchmarks, and a pooled rank-percentile evaluation. Our strongest variant, which combines noun supervision with untrimmed utterances, improves average out-of-domain accuracy by +3.5% and overall accuracy by +4.4% over prior infant-perspective baselines. It also achieves the best group score on in-domain compositional reasoning. Our findings follow infants developmental cognition, suggesting that relaxation of frame-level alignment yields more transferable object-centric representations from noisy infant-perspective videos. Additional analyses suggest that MILCVCL yields more semantically grounded and shift-robust object representations and that scaffolded noun-to-utterance training can further improve generalization. The code will be available at https://github.com/abrhamkg/milcvcl.

Abstract:
We present the first systematic study of Sparse Autoencoders (SAEs) on video representations. Standard SAEs decompose video into interpretable, monosemantic features but destroy temporal coherence: hard TopK selection produces unstable feature assignments across frames, reducing autocorrelation by 36%. We propose spatio-temporal contrastive objectives and Matryoshka hierarchical grouping that recover and even exceed raw temporal coherence. The contrastive loss weight controls a tunable trade-off between reconstruction and temporal coherence. A systematic ablation on two backbones and two datasets shows that different configurations excel at different goals: reconstruction fidelity, temporal coherence, action discrimination, or interpretability. Contrastive SAE features improve action classification by +3.9 pp over raw features and text-video retrieval by up to 2.8× R@1. A cross-backbone analysis reveals that standard monosemanticity metrics contain a backbone-alignment artifact: both DINOv2 and VideoMAE produce equally monosemantic features under an independent (CLIP) similarity space. Targeted feature ablation shows that contrastive training concentrates the probe's predictive signal into a small number of identifiable features. Supplementary material, code, configurations and evaluation scripts are available at \urlhttps://github.com/atahandokme/spatio-temporal-sparse-autoencoders-video.

Abstract:
Autoregressive (AR) models have achieved unified and strong performance across both visual understanding and image generation tasks. However, removing undesired concepts from AR models while preserving overall generative quality remains an open and critical challenge. Existing Concept Erasure (CE) methods, primarily designed for diffusion models, fail when applied to AR architectures due to fundamental differences in their sequential generative process. In this paper, we propose the Erasure Autoregressive Model (EAR), the first fine-tuning framework specifically designed for CE in AR models. EAR addresses the fact that concepts in AR models manifest as dynamic trajectories within the hidden states, not as static vectors. Specifically, it introduces Windowed Gradient Accumulation (WGA), which accumulates alignment loss over windows of the hidden state trajectory to target stable conceptual features rather than noisy, isolated patches, and Thresholded Loss Masking (TLM), which selectively applies gradients only to trajectory windows where the target concept is strongly present, preserving utility for non-target concepts. To facilitate rigorous evaluation, we also propose Erase Concept Generator and Visual Filter (ECGVF), a novel benchmark for the CE task. ECGVF uses multiple LLMs to generate high-quality contrastive prompt pairs and employs visual classifiers for rigorous filtering to ensure dataset quality and cross-model comparability. Extensive experiments on the AR model Janus-Pro using our ECGVF benchmark demonstrate that EAR achieves 91% average removal rate while maintaining generative fidelity on unrelated prompts. Additional cross-architecture experiments on Lumina-mGPT show that EAR also transfers beyond Janus-specific settings, while disentanglement analysis confirms that closely related but safe concepts remain largely intact. Our code is available at https://github.com/immc-lab/ear.

Abstract:
As generative AI tools become increasingly powerful and easy to use, people can easily edit portrait images with a prompt, necessitating the task of image editing attribution, which predicts the involved editing tools from the given image. Existing attribution methods hold the single-tool assumption and can only attribute a specific editing tool, but struggle to handle the more complex and increasingly common multi-tool editing scenarios, where artifacts left by different editing tools are composite and overlapped. To address this gap, we explore Multi-Tool Image Editing Attribution~(MIEA), which aims to identify multiple editing tools involved in a multi-tool edited facial image. To simulate the real-life editing operations on facial images, we then construct a new dataset, MultiEdit, which contains 500k+ edited facial images and covers six types of editing tools that support face swapping~(Deepfake) and various facial enhancements. Inspired by the findings from data analysis, we design MIEA, a multi-tool attribution method that can capture distinguishable, locality-aware editing tool traces from both spatial and frequency domains with the support of an error-based curriculum learning strategy. Experiments show \Method\ outperforms nine methods for facial images edited in at most five steps.\footnoteCode, Data \& Extended Materials: https://github.com/ICTMCG/MIEA

Abstract:
We introduce the first method, to the best of our knowledge, for adapting \emphimage-to-video models to layer-aware text (glyph) animation, a capability critical for practical dynamic visual design. Existing approaches predominantly handle the transparency-encoding (\alpha channel) as an extra latent dimension appended to the RGB space, necessitating the reconstruction of the underlying RGB-centric \emphvariational autoencoder (VAE). However, given the scarcity of high-quality transparent glyph data, retraining the VAE is computationally expensive and may erode the robust semantic priors learned from massive RGB corpora, potentially leading to latent pattern mixing. To mitigate these limitations, we propose TransText, a framework based on a novel \emphAlpha-as-RGB paradigm to jointly model appearance and transparency without modifying the pre-trained generative manifold. TransText embeds the \alpha channel as an RGB-compatible visual signal through latent spatial concatenation, explicitly ensuring strict cross-modal (RGB-and-Alpha) consistency while preventing feature entanglement. Our experiments demonstrate that TransText significantly outperforms baselines, generating coherent, high-fidelity transparent animations with diverse, fine-grained effects. More details are available at \urlhttps://sii-ferenas.github.io/TransText/.

Abstract:
Multimodal sequential recommendation (MSR) improves recommendation by incorporating heterogeneous information such as text, images, and user interactions. However, existing MSR methods often fail to capture user-level preference heterogeneity and dataset-level modality bias, limiting their adaptability across users and datasets. To address this issue, we propose Sequence-Guided Universal Multimodal Prioritization Calculation Framework (SG-UMP), a plug-and-play plugin for enhancing multimodal information processing in MSR. SG-UMP includes a Module Combiner for flexible multimodal processing and a Module Router for dynamic module ordering, enabling adaptation to both user preferences and dataset characteristics. Experiments on four real-world datasets show that SG-UMP consistently improves recommendation performance across different backbones and multimodal settings. The code is available at ~\urlhttps://github.com/esemsc-xz524/SG-UMP.

Abstract:
\beginabstract Modern Multimodal Retrieval-Augmented Generation (M-RAG) systems are fundamentally limited by the binary connectivity paradigm of traditional simple graphs, which fails to capture the intricate, high-order correlations among heterogeneous entitiessuch as the N-ary relationships between a visual chart, its scattered textual descriptions, and underlying numerical data. Furthermore, existing refinement strategies often rely on exhaustive, full-page reconstruction to align cross-modal information, leading to prohibitive computational redundancy and the introduction of contextual noise in long-form document processing. In this paper, we propose Hyper-M2RAG, a novel framework that redefines multimodal document retrieval through High-order Hypergraph Representation Learning. We first formalize the document structure as a Multimodal Hypergraph, utilizing hyperedges as unified semantic containers to encapsulate multi-way associations across text, images, and tables, thereby transcending point-to-point modeling. To mitigate semantic fragmentation caused by physical pagination, we introduce an Anchor-driven Incremental Refinement mechanism. Rather than performing a global sweep, our approach surgically identifies boundary-crossing anchorsnodes and reconstructs their local hyper-topology using one-hop neighborhood contexts. This targeted refinement effectively bridges cross-page knowledge gaps with minimal computational footprints. Extensive evaluations on multimodal benchmarking datasets demonstrate that Hyper-M2RAG significantly outperforms state-of-the-art methods in both retrieval precision and generation coherence. Our code is accessible at \urlhttps://github.com/ShenAoChen2001/MMHRAG. \endabstract

Abstract:
Dataset distillation compresses a large training set into a tiny synthetic set, greatly reducing training cost. However, models trained on distilled data remain highly vulnerable to adversarial attacks, especially in highly compressed regimes. A key challenge in robust dataset distillation is that adversarial robustness is not automatically inherited by synthetic data through standard distillation objectives. In particular, sharp robust teachers provide unstable supervision, while conventional feature-matching methods do not explicitly preserve adversarial distributions. To address this issue, we propose Adversarial Distribution Matching (ADM), a two-stage framework that explicitly transfers robust knowledge into the distilled dataset. First, ADM smooths the robust teacher via randomized weight perturbation during fine-tuning, producing a more stable feature space for distillation. Second, it aligns both clean and adversarial feature distributions between real and synthetic data using characteristic function matching in the spectral domain, which captures the complex shifts induced by adversarial perturbations. Extensive experiments on multiple datasets show that ADM consistently achieves substantially stronger adversarial robustness than prior methods while maintaining competitive clean accuracy. The code is publicly available at https://github.com/BaiqiWu/ADM_MM26.

Abstract:
Image stylization under multi-modal conditions faces challenges in semantic-style alignment due to limited cross-modal understanding. % While Multimodal Large Language Model (MLLM) offers a solution, it often introduces hallucinations. While Multimodal Large Language Model (MLLM) offers a solution, it often fails to ensure consistent semantic-style alignment. We propose MAIST, an MLLM-Augmented Image Stylization framework based on an Imagine-Infuse paradigm. The Imagine stage uses an MLLM to generate a text description from input conditions (text/image), which guides the creation of a stylized exemplar via a pre-trained diffusion model. In the Infuse stage, this text provides rich semantic and stylistic guidance, while the exemplar offers strong spatial constraints to enhance fidelity and controllability. This dual conditioning grounds structure while allowing flexible input control for content and style. Extensive experiments show MAIST achieves superior performance across diverse style transfer tasks. Project Page: \hrefhttps://github.com/Vincotto/MAIST \textcolorred!40\underline\textttgithub.com/Vincotto/MAIST

Abstract:
With the rapid advancement of Latent Diffusion Models (LDMs), watermarking has emerged as a crucial technique for intellectual property protection. However, existing methods rely on computationally intensive heuristic optimizationleading to high training overhead and local optima. To address these issues, we propose an \underlineAna\underlinelytical Watermark\underlineing Framework for Controllabl\underlinee Generatio\underlinen (ALIEN). Unlike optimization-based approaches, ALIEN develops the first analytical derivation of a time-dependent modulation coefficient, directly guiding the reverse SDE probability flow to embed watermark residuals without diffusion inversion. To accommodate diverse real-world deployment requirements, ALIEN enables a flexible transition between high-fidelity pixel alignment (ALIEN-Q) and high-resilience semantic embedding (ALIEN-R). ALIEN-Q functions as a strict content-preserving paradigm by injecting the watermark within a restricted window, ensuring zero semantic deviation from unwatermarked images while maintaining baseline robustness. ALIEN-R acts as a highly resilient semantic watermark by utilizing the full injection range. While still ensuring great visual quality, it achieves state-of-the-art resistance against adversarial perturbations. Extensive experiments demonstrate the superiority of both modes in their respective domains. ALIEN-Q outperforms existing methods by 33.1% across multiple fidelity metrics, while ALIEN-R delivers a 14.0% improvement in robustness against complex generative and stability threats across 15 distinct conditions. Code is available at https://github.com/REPO-EXP/ALIEN.git.

Abstract:
While multimodal large language models (MLLMs) achieve remarkable performance on generic image captioning, they frequently suffer from structural hallucinations in human-centric scenarios. % Accurately modeling human subjects is foundational for critical downstream applications, such as accurate avatar/video/image generation and fine-grained human action understanding. % However, these tasks require highly precise subject-centered spatial grounding, such as distinguishing egocentric left/right laterality and maintaining correct anatomical-object bindings. % Although catastrophic for structural integrity, these localized spatial inversions are often overshadowed by overall descriptive metrics in existing benchmarks. % To systematically expose and quantify this bottleneck, we introduce SPACE (Subject-centric Poses, Appearance, and Characteristics Evaluation), a benchmark designed to evaluate subject-centered spatial understanding. % On SPACE, we reveal that despite strong generic perception, current MLLMs consistently fail to ground descriptions in the subject's intrinsic frame of reference. To bridge this gap, we propose a specialized data construction and alignment pipeline. % We first extract structured spatial hints from fine-grained body-part localization to guide a two-stage caption rewriting process, yielding highly spatially-faithful training data. Furthermore, we design a rubric-based reward for Group Relative Policy Optimization (GRPO) that explicitly penalizes structurally critical spatial errors during alignment. Extensive experiments on SPACE demonstrate our framework significantly improves human-centric caption quality, particularly in subject-centered spatial reasoning, achieving performance competitive with strong closed-source models. %These results highlight the necessity of structured spatial alignment and set a new paradigm for generating fine-grained, anatomically grounded captions. % Our benchmark and code are available at \urlhttps://github.com/JHang2020/SPACE-Eval.

Abstract:
Dynamic representations based on 4D Gaussian Splatting (4DGS) have shown strong capabilities in reconstruction and rendering, yet current methods remain limited in supporting complex natural-language referring in dynamic scenes. Existing 4D Gaussian methods are primarily designed for rendering-oriented dynamic modeling, while current semantic extensions often rely on scene-specific semantic optimization or are evaluated on relatively simple query settings. In this paper, we study Referring 4D Gaussian Splatting (R4DGS), a task for grounding temporally varying, multi-target or reasoning-intensive, and zero-target or distractor expressions in dynamic 4D scenes. To support this task, we introduce R4D-Bench-QA, a benchmark with structured query annotations. We further present ReferGaussian, a unified framework built upon 4DGS that converts low-level dynamic Gaussians into persistent, language-addressable entities for training-free referring inference. ReferGaussian organizes mask-supported Gaussian entities in an EntityBank, uses Multimodal Large Language Models (MLLMs) to decompose complex queries into trackable object phrases and structured referring constraints, and performs constraint-aware grounding without scene-specific semantic optimization. Experiments on R4D-Bench-QA and the 4D LangSplat protocol demonstrate improved temporal target selection and competitive spatiotemporal grounding without scene-specific semantic training. Project resources are available at \urlhttps://trump0412.github.io/R4DGS/.

Abstract:
Identity-Preserving Personalized Generation (IPPG) has advanced film production and artistic creation, yet existing approaches overemphasize facial regionsresulting in outputs dominated by "facial close-ups". These methods suffer from weak visual narrativity and poor semantic consistency under complex text prompts, rooted in identity (ID) feature embeddings that undermine the semantic expressiveness of generative models. To address these issues, this paper presents an IPPG method that breaks the constraint of facial close-ups, achieving synergistic optimization of identity fidelity and scene semantic creation. Specifically, we design a Dual-Line Inference (DLI) pipeline with identity-semantic separation, resolving the ID-semantics representation conflict in traditional single-path architectures. Further, we propose an Identity Adaptive Fusion (IdAF) strategy that defers ID-semantic fusion to the noise prediction stage, integrating adaptive attention fusion and noise masking to avoid ID embedding interference on semantics without manual masking. Finally, an Identity Aggregation Prepending (IdAP) module aggregates ID information in place of random initializations, further enhancing identity preservation. Experimental results validate that our method achieves stable and effective performance in IPPG tasks beyond facial close-ups, enabling efficient generation without manual masking or fine-tuning. As a plug-and-play component, it can be rapidly deployed in existing IPPG frameworks, facilitating film-level character-scene creation and enriching personalized generation for related domains. Project page: https://dids-aigc.github.io/BeyondFacial/.

Abstract:
Music creation is fundamentally a process of revision. Yet symbolic music generation remains dominated by paradigms that produce complete sequences from scratch, with limited support for selective modification. Edit-based methods have proven effective for text transformation tasks, but remain largely unexplored for symbolic music. We trace this absence to the representational level: conventional event-based music encodings lack the structural properties required by explicit music editing. In contrast, the BEAT encoding, a beat-grid-anchored representation originally designed for autoregressive generation, possesses structural properties amenable to editing. We propose BeatEdit, the first framework for symbolic music generation based on explicit edit operations, recasting generation as producing new content by editing a draft rather than synthesizing from scratch. BeatEdit comprises three complementary mechanisms along an axis of increasing edit density: per-token sequence tagging for error correction, iterative refinement for accompaniment editing, and tag-then-fill for segment completion. All these mechanisms share a single encoding and pre-trained backbone, achieving higher precision and perceptual quality than autoregressive and diffusion methods across all three tasks, while remaining efficient, with single-pass inference completing in under 100 ms. Cross-encoding evaluation further reveals that encoding design substantially influences editing effectiveness, with notable encodingmethod interaction effects. Code is available at https://github.com/Haoyu-Gu/BeatEdit-code; an extended version containing Appendices AP is available at https://arxiv.org/abs/2607.11124.

Abstract:
Interpretation is essential to deciphering the language of art: audiences communicate with artists by recovering meaning from visual artifacts. However, current Generative Art (GenArt) evaluators primarily emphasize surface-level image quality or literal prompt adherence, failing to assess the deeper symbolic or contextual meaning intended by the creator. We formalize this observation using a Peircean computational semiotic theory that models Human-GenArt Interaction (HGI) as cascaded semiosis. This framework reveals that artistic meaning is conveyed through iconic, symbolic, and indexical modes, whereas appearance-centered evaluators tend to emphasize the iconic mode. Building on this account, we propose SemJudge, an interpretation-centric evaluation framework that uses a Hierarchical Semiosis Graph (HSG) to reconstruct meaning conveyance from the prompt to the generated artifact. Experiments on an interpretation-intensive benchmark, SemiosisArt, show that SemJudge aligns more closely with human judgments than existing baselines. User studies further demonstrate that SemJudge produces deeper, more insightful artistic interpretations. We hope these findings pave the way for GenArt to move beyond the generation of ``pretty'' images toward a medium capable of understanding and expressing complex human experience. Project page: \urlhttps://github.com/songrise/SemJudge

Abstract:
Dance-to-music (D2M) generation aims to synthesize music that is rhythmically and stylistically aligned with dance videos. A key challenge arises from the semantic mismatch between sparse dance cues, such as rhythm and style, and the dense information required for music composition, including structure, instrumentation, and expressive dynamics. Existing methods typically rely on these sparse cues and supervise only the final audio output, resulting in poorly learned music representations and generated music with limited musicality and structural coherence. To address these issues, we propose Curriculum-guided Multi-scale representation Alignment framework with scale-aware Optimal Transport (CMA-OT), a novel paradigm that leverages an external music expert to provide hierarchical supervision for the generators latent features, bridging the semantic gap and enhancing representation learning. To effectively incorporate hierarchical supervision, we introduce a curriculum-guided multi-scale learning strategy that progressively transfers musical knowledge from expert to the music generator, enabling stable and effective representation learning. Moreover, to accommodate the semantic and structural variations across different expert scales and achieve fine-grained alignment under temporal mismatch, we propose a scale-aware optimal transport alignment mechanism, which models soft correspondences between hierarchical expert representations and generator's latent features. Extensive experiments on two datasets demonstrate that CMA-OT achieves state-of-the-art (SOTA) performance in rhythmic synchronization, perceptual quality, and overall music generation. The project page is available at \hrefhttps://beria-moon.github.io/CMA-OT/CMA-OT.

Abstract:
Existing Vision-Language-Action (VLA) models often struggle to generalize to long-horizon tasks due to their heavy reliance on immediate observations. While recent studies incorporate retrieval mechanisms or extend context windows to handle procedural tasks, they often struggle to capture Non-Markovian dependencies, where optimal actions cannot be determined solely by the current observation, but rely on specific past states. To address this, we introduce Keyframe-Chaining VLA, a framework that extracts and links key historical frames to model long-horizon dependencies. Specifically, we propose an automatic keyframe selector that learns a discriminative embedding space, effectively identifying distinct state transitions. To capture task-critical information, we design a progress-aware query mechanism that dynamically retrieves historical frames based on their temporal relevance to the current execution phase. These selected keyframes are integrated into the VLA as interleaved visual tokens, explicitly grounding the policy in the long-horizon temporal context. Finally, we introduce a suite of four non-Markovian manipulation tasks built upon the ManiSkill simulator to measure task success rates. Experimental results demonstrate that our method achieves superior performance, effectively tackling robot manipulation tasks characterized by long-horizon temporal dependencies.

Abstract:
Multimodal Sentiment Analysis (MSA) aims to interpret complex human emotions by integrating natural language with non-verbal modalities. Non-verbal modalities share a structural isomorphism with natural language, as both can be viewed as feature sequences evolving over time. This isomorphism enables the transformation of non-verbal modalities into text-like tokens for unified semantic reasoning. Large Language Models (LLMs), designed to understand and generate sequential data, can thus be utilized to interpret complex affective sequences. However, existing LLM-based methods primarily capture low-level superficial features, failing to model affective semantics arising from structural variations and contextual interactions. To address this limitation, we propose SentiLLM, a unified framework that leverages Semantic-Aligned Structural Abstraction to distill continuous raw signals into compact, semantically meaningful tokens. Specifically, we introduce a Dual-Stream Salience-Context Calibration Mechanism, which disentangles non-verbal feature sequences into a focus stream and an ambient stream. The focus stream captures salient sentiment shifts (e.g., facial expressions) guided by textual priors, while the ambient stream characterizes stable background states. Through calibrating these dynamic sentiment shifts against background states, SentiLLM effectively projects non-verbal modalities into a unified semantic space, making them naturally understandable for LLMs. Serving as a plug-and-play module, SentiLLM significantly improves discriminative performance with only a small number of trainable parameters. Our method achieves superior performance on four datasetsMOSI, MOSEI, CH-SIMS, and CH-SIMS v2demonstrating the effectiveness of the structural abstraction paradigm in MSA. Our code is available at: \hrefhttps://github.com/especiallyW/SentiLLMhttps://github.com/especiallyW/SentiLLM.

Abstract:
Deep Neural Networks (DNNs) remain vulnerable to diverse adversarial perturbations, motivating multi-attack adversarial training (AT) for improved robustness. However, existing methods either incur prohibitive overhead by computing all attacks at each iteration, or rely on stochastic sampling over adversarial examples, which may cause excessive parameter drift. To address these issues, we propose Calibrated Adversarial Sampling (CAS), an efficient and stable framework that reformulates multi-attack AT as a multi-armed bandit optimization problem. By sampling a single attack per iteration that dynamically balances exploration and exploitation, CAS significantly reduces training cost while mitigating optimization conflicts across attacks and controlling excessive parameter drifts. Extensive experiments demonstrate that CAS achieves superior overall robustness at low computational cost, offering a scalable and principled approach to robust generalization against multi-attack settings. Our code is available at \urlhttps://github.com/1240148048/CAS.

Abstract:
Precipitation nowcasting remains challenging due to the highly localized, rapidly evolving, and heterogeneous nature of atmospheric dynamics. Although recent methods adopt attention-based architectures in both unimodal and multimodal settings, they mainly emphasize stronger representation learning and prediction capacity, while paying less attention to the stability of attention responses across samples. In this work, we show that cross-sample instability of attention-response energy is an important and previously underexplored source of forecasting unreliability. Empirically, inaccurate forecasts are associated with larger attention-response energy variance across heads and layers. Theoretically, we show that cross-sample variability can propagate through self-attention, and enlarge a lower bound on prediction error. Based on this insight, we propose HARECast, a Head-wise Attention Response Energy-regulated framework for precipitation nowcasting. HARECast explicitly models head-wise attention-response energy and stabilizes it through a group-wise regularization objective that reduces cross-sample fluctuations. The proposed formulation is generic and applicable to both unimodal and multimodal nowcasting. We instantiate HARECast in a standard forecasting pipeline with reconstruction branches and a diffusion-based predictor, and evaluate it on commonly used benchmarks--SEVIR and MeteoNet. Experiments demonstrate that HARECast achieves state-of-the-art performance. Our code is available at https://github.com/ph-w2000/HARECast.

Abstract:
Generating images directly from spoken language without intermediate text transcription remains a fundamental challenge in cross-modal perception. Existing approaches cascade Automatic Speech Recognition (ASR) with Text-to-Image (T2I) models, introducing compounding transcription errors and latency that might degrade quality of generated images. We eliminate this bottleneck entirely by proposing SPIMSyn and SPIMSyn-D, two SPeech to IMage Synthesis frameworks, for adapting Stable Diffusion~3.5 (SD~3.5), a popular T2I model, to direct Speech to Image (S2I) generation. We design a 3-branch cross-modal transformer decoder architecture to transform speech representations into text representations corresponding to three text encoders of SD~3.5 namely CLIP-L, CLIP-G, and T5-XXL. For obtaining speech representations, SPIMSyn uses a pre-trained Whisper encoder, while SPIMSyn-D employs training-free discrete mel-filterbank tokenization via dMel. Extensive experiments on direct S2I generation show that both SPIMSyn and SPIMSyn-D eliminate dependence on explicit text decoding from ASR, thereby reducing cascading errors, providing competitive image generation quality with lower end-to-end latency, when compared to conventional cascaded S2I baselines. \hrefhttps://adhithyaa15.github.io/Speech2Image/\textcolorblue\underlineProject page: https://adhithyaa15.github.io/Speech2Image/

Abstract:
Ordinal regression is widely used in scenarios where labels are discrete yet inherently ordered. In practice, however, ordinal labels are often obtained by discretizing underlying continuous semantics through subjective human judgment, resulting in ambiguous class boundaries and annotation noise. This poses a fundamental challenge to existing methods that rely on static supervision, as predefined labels may impose globally rigid and even biased ordering constraints throughout training. To address this limitation, we propose \modelname, a dynamic distribution distillation framework for ordinal regression. Beyond relying solely on predefined ordinal labels, \modelname introduces a dynamic evolution of label distributions via self-distillation. Specifically, we introduce a contrastive ordinal-aware label enhancement module that leverages visionlanguage alignment to recover ordinal label distributions capturing both inter-class ambiguity and instance-level uncertainty. These recovered distributions serve as dynamic supervision signals that are iteratively refined throughout training. Furthermore, we design a CDF-based cross-layer interaction distillation mechanism to propagate cumulative ordinal structure across network layers, improving hierarchical consistency of ordinal representations. Extensive experiments on four general ordinal regression tasks demonstrate that \modelname consistently outperforms existing methods, particularly under noisy and imbalanced settings. These results highlight the importance of moving beyond static supervision toward dynamic distribution refinement for robust ordinal representation learning. The code is publicly available at https://github.com/ZJUMAI/D3O.

Abstract:
Gaze Object Prediction (GOP) aims to localize and recognize the objects humans attend to, a task crucial for understanding human-centric interactions. However, existing methods are typically trained under a closed-vocabulary paradigm with a fixed label space and evaluated on scene-specific datasets, limiting their applicability to real-world scenarios where gaze targets often follow a long-tail distribution or belong to unseen categories. To address this gap, we introduce Diverse Scenes for Gaze object prediction (DiSG), a benchmark containing 86 in-the-wild categories that facilitates the evaluation of Open-Vocabulary GOP (OVGOP). Building on DiSG, we propose a framework that leverages text-driven object discovery to localize potential gaze candidates, with a gaze-guided selection module to pinpoint the intended target from the candidate objects. Furthermore, to better capture semantic knowledge across diverse in-the-wild categories, we introduce Gradient-Informed Selection Tuning (GIST) to selectively update parameters most relevant to a given class vocabulary. Extensive experiments demonstrate that our proposed model performs effectively in open-vocabulary settings and also outperforms existing methods in the conventional closed-vocabulary setting. The benchmark and code is available at \urlhttps://github.com/sensniu/ovgop.

Abstract:
Thermography plays a vital role in military and broader thermal analysis applications. Recent progress in 3D thermal reconstruction has extended temperature analysis from 2D to 3D space, yet most existing works assume static temperature distributions, neglecting the temporal dynamics of heat transfer in real-world environments. To address this limitation, we propose the first dynamic RGB-Thermal reconstruction framework for complex scenes. Our method jointly models RGB appearance, thermal observations, and scene geometry as they change over time. Specifically, we introduce a multimodal dynamic scene representation that anchors both the color and thermal modalities to a shared geometric substrate, ensuring their consistency under spatiotemporal deformations. We further design multimodal embeddings to enhance the motion expressiveness for each modality, and propose a multimodal routing mechanism that retains a unified set of shared multimodal Gaussians as the geometric backbone while adaptively spawning modality-specific Gaussians to strengthen the representational capacity in detail-rich regions of each individual modality. In addition, we contribute a novel benchmark dataset featuring high-frequency temperature variations to facilitate the evaluation of 4D reconstruction. Extensive experiments demonstrate that our method achieves high-fidelity spatiotemporal reconstruction of both appearance and temperature. Our code and dataset are available at: https://github.com/LinLif1869/DTG.

Abstract:
UAV Geo-Localization faces significant challenges due to the drastic appearance discrepancy between drone-captured images and satellite views. Existing methods typically assume a consistent scaling factor across views and rely on predefined partition alignment to extract viewpoint-invariant representations through part-level feature construction. However, this scaling assumption often fails in real-world scenarios, where variations in drone flight states lead to scale mismatches between cross-view images, resulting in severe performance degradation. To address this issue, we propose a scale-adaptive partition learning framework that leverages known drone flight height to predict scale factors and dynamically adjust feature extraction. Our key contribution is a height-aware adjustment strategy, which calculates the relative height ratio between drone and satellite views, dynamically adjusting partition sizes to explicitly align semantic information between partition pairs. This strategy is integrated into a Scale-adaptive Local Partition Network (SaLPN), building upon an existing square partition strategy to extract both fine-grained and global features. Additionally, we propose a saliency-guided refinement strategy to enhance part-level features, further improving retrieval accuracy. Extensive experiments validate that our height-aware, scale-adaptive approach achieves state-of-the-art geo-localization accuracy in various scale-inconsistent scenarios and exhibits strong robustness against scale variations. \urlhttps://github.com/C-water/SaLPN

Abstract:
Generalized Visual Grounding (GVG) task aims to localize targets in an image based on referring expressions, extends the classical visual grounding paradigm by integrating multi-target and non-target scenarios. Previous methods typically rely on global semantic matching or coarse-grained region interactions for localization, where the discriminative cues are primarily derived from sentence-level semantics or regional context. In complex multi-target scenarios, such approaches tend to confuse visually similar targets, making it difficult to establish stable instance-level decision boundaries. To address these limitations, this paper proposes a novel SemanticSpatial Discriminability Enhancement (SSDE) framework for generalized visual grounding, which aims to enhance the discriminative ability on fine-grained semantics and spatial localization, improving both cross-modal understanding and instance-level grounding. Specifically, to enhance the semantic discriminability of query representations at the fine-grained level, we propose a Semantic Discriminability Enhancement (SeDE) module, which leverages spatially guided cross-attention to disentangle fine-grained target-relevant visual attributes and integrates them with the textual subject semantics. Furthermore, to strengthen the spatial discriminability of the referred targets, we introduce a Spatial Discriminability Enhancement (SpDE) module, which models an instance center density map to characterize the spatial distribution of targets, and explicitly constructs instance separation structures in the spatial domain by employing them as an auxiliary supervision signal. Extensive experiments show that SSDE achieves superior performance on ten datasets across both classic and generalized visual grounding tasks. Code will be available at \urlhttps://github.com/Letitialky/GVG-SSDE.

Abstract:
Charts are ubiquitous in scientific and financial literature for presenting structured data. However, chart reasoning remains challenging for multimodal large language models (MLLMs) due to the lack of high-quality training data, as well as the need for fine-grained visual grounding and precise numerical computation. To address these challenges, we first propose \textscDuoChart, a scalable dual-source data pipeline that combines synthesized charts with real-world charts to construct diverse, high-quality chart training data. We then introduce \textscCharTool, which equips MLLMs with external tools, including image cropping for localized visual perception and code-based computation for accurate numerical reasoning. Through agentic reinforcement learning on \textscDuoChart, \textscCharTool learns tool-integrated reasoning grounded in chart content. Extensive experiments on six chart benchmarks show that our method consistently improves over strong MLLM baselines across model scales. Notably, \textscCharTool-7B outperforms the base model by +8.0% on CharXiv (Reasoning) and +9.78% on ChartQAPro, while achieving competitive performance with substantially larger or proprietary models. Moreover, \textscCharTool demonstrates positive generalization to out-of-domain visual math reasoning benchmarks. Our data and models are available at \urlhttps://github.com/OpenDFM/CharTool.

Abstract:
\beginabstract Text-to-image person retrieval (TIPR) aims to retrieve target person images based on textual descriptions and has gained attention in the cross-modal community. Despite recent progress, numerous methods implicitly assume that training image-text pairs are correctly aligned. In practice, real-world datasets inevitably contain weakly aligned or mismatched pairs, i.e., noisy correspondence (NC), which undermines the strict correspondence assumption and in turn degrades model training. To address this problem, we propose a novel Neighbor-guided Robust Learning (NRL) framework, which uses a layer-fused cross-modal backbone and is built on an intuitive yet effective idea of cross-modal neighbor relations: visually similar images should correspond to similar textual descriptions; conversely, textually similar descriptions should correspond to similar images. NRL evaluates the reliability of each potentially noisy pair by checking whether it is consistently supported by its clean neighbors in the shared embedding space, and further calibrates its reliability by comparing its optimization difficulty with that of its most consistent clean neighbor, thereby suppressing noisy pairs while retaining correct but under-represented ones. Extensive experiments on three public benchmarks demonstrate that NRL achieves superior performance and strong robustness, regardless of whether synthetic noisy correspondence is injected. Code is available at https://github.com/RunQing715/NRL. \endabstract

Abstract:
Audio-video (AV) generation has recently made strong progress in perceptual quality and multimodal coherence, yet generating content with plausible motion-sound relations remains challenging. Existing methods often produce object motions that are visually unstable and sounds that are only loosely aligned with salient motion or contact events, largely because they lack an explicit motion-aware structure shared by video and audio generation. We present Tora3, a trajectory-guided AV generation framework that improves physical coherence by using object trajectories as a shared kinematic prior. Rather than treating trajectories as a video-only control signal, Tora3 uses them to jointly guide visual motion and acoustic events. Specifically, we design a trajectory-aligned motion representation for video, a kinematic-audio alignment module driven by trajectory-derived second-order kinematic states, and a hybrid flow matching scheme that preserves trajectory fidelity in trajectory-conditioned regions while maintaining local coherence elsewhere. We further curate PAV, a large-scale AV dataset emphasizing motion-relevant patterns with automatically extracted motion annotations. Extensive experiments show that Tora3 improves motion realism, motion-sound synchronization, and overall AV generation quality over strong open-source baselines. Project page: https://ali-videoai.github.io/tora3_page.

Abstract:
Cross-modal retrieval aims to establish semantic alignments across different modalities. However, existing methods frequently assume the availability of perfectly paired training data. This assumption ignores the inevitable noisy correspondence (NC) within real-world datasets, which degrades the performance of retrieval. To address these limitations, we propose the Multi-phase Collaborative Representation (MCR) framework. To the best of our knowledge, MCR is the first framework that introduces cross-modal collaborative representation to tackle the NC problem, accurately identifying true correspondences. Specifically, MCR reconstructs a query feature using a linear combination of training samples within the batch from the other modality. By optimizing the objective of collaborative representation, we leverage the reconstruction coefficients to measure the semantic correlation between each pair. Furthermore, we propose a multi-phase iterative process that enhances the feature representations of clean samples and dynamically filters noisy ones. This process produces reliable matching confidence weights to effectively quantify the noise. Extensive experiments on three widely-used benchmark datasets, including Flickr30K, MS-COCO, and Conceptual Captions, validate the effectiveness of our proposed approach. The code is available at https://github.com/Ji-Xiaoying/MM26_MCR.

Abstract:
Our primary objective is to advance video object counting in crowded scenes, aiming to robustly count all instances of a target category based on given text or visual prompts. Existing methods rely on RGB information, limiting their discriminative ability in crowded and occluded conditions. To address this, we propose a Depth-Guided Detector (DG-Det) along with a general post-processing pipeline. By integrating depth cues with multi-scale RGB-D cross-attention and explicit occlusion prediction, our method enhances spatial understanding and achieves robust detection in crowded and occluded scenes. Furthermore, we introduce a unified de-duplication framework to eliminate cross-frame redundant counting. To facilitate future research, we also release a new RGB-D Video Object Counting dataset featuring depth information and multiple object categories persequence. Extensive experiments demonstrate that our method achieves a 62.01% reduction in MAE compared to existing baselines, and also produces consistent improvements in RMSE. We provide the source code at \urlhttps://github.com/streamer-AP/DG-Net and the dataset at \urlhttps://huggingface.co/datasets/aerospace123/RGBD-VideoCount.

Abstract:
Vision-Language Models (VLMs) are known to be vulnerable to adversarial attacks, where subtle perturbations to images or texts induce erroneous outputs. However, most text-based attacks are adapted from language-model-centric methods, in which the visual input is fixed during optimization, resulting in adversarial prompts that are tied to specific images and thus limiting their attack effectiveness. To this end, we first introduce a new research perspective: cross-image transferability for adversarial prompts. We then propose \ourmethod, an adversarial prompt that is optimized once and reused to steer VLM outputs toward attacker-specified responses across diverse images. \ourmethod employs a joint optimization that distills image-invariant adversarial features into the prompt by ``worst-case'' generation. Specifically, it alternates between constructing hard visual conditions for the current prompt and updating the prompt to remain effective under these conditions. Extensive experiments on prevalent VLMs verify that \ourmethod achieves an improvement of over 30% in attack success rates compared to state-of-the-art (SoTA) baselines, while reducing computation time by ～70%. Our code is avalable at~\urlhttps://github.com/Ye-ze-yu/GhostPrompt.

Abstract:
High-fidelity face normal estimation is a fundamental task for photorealistic 3D reconstruction, relighting, and avatar creation. Previous methods formulate this problem as a direct pixel-wise regression, which often leads to over-smoothed results, fails to capture fine-scale geometric details, and exhibits sensitivity to challenging in-the-wild lighting conditions. In this paper, we challenge this paradigm by reframing face normal estimation as a conditional generative modeling problem. We introduce \emphFNFlow, a novel framework that generates precise normal maps by leveraging the power and efficiency of latent flow matching. This formulation departs from conventional regression paradigms and enables more expressive modeling of fine-scale geometric details. By learning the entire conditional distribution instead of a single point estimate, \emphFNFlow implicitly learns a strong geometric prior of human faces, enabling it to robustly generate sharp, consistent, and detailed normal maps even from ambiguous inputs. Extensive quantitative and qualitative evaluations on benchmark datasets demonstrate that \emphFNFlow achieves state-of-the-art performance. Our code is available at \urlhttps://github.com/zhangquanbo/FNFlow.

Abstract:
Learning-based stereo matching has achieved remarkable progress in recent years. However, existing cost aggregation methods mainly exploit local information within the cost volume, which limits their aggregation range and degrades performance in ill-posed regions. To address this issue, we propose the Laplacian-guided Adaptive Cost Aggregation (LACA) network. Specifically, we propose a novel Disparity-Discontinuous Hybrid Scanning strategy to adapt Vision Mamba to stereo cost aggregation on 4D cost volumes, and build a Mamba-based Global Cost Aggregation (GCA) module upon it to enable global contextual modeling, thereby overcoming the limited aggregation range of existing methods. To compensate for the detail degradation caused by Mamba's low-frequency modeling bias, we further introduce Local Cost Aggregation (LCA) and a Laplacian-guided Adaptive Cost Fusion (LACF) module to balance robustness in ill-posed regions with accuracy in edge-detail regions. Extensive experiments demonstrate the effectiveness of LACA, including SOTA performance on KITTI 2012 Reflective and strong zero-shot generalization on multiple real-world benchmarks. The code is available at \urlhttps://github.com/TaiSen-WUhan/LACA.

Abstract:
Vocal-to-accompaniment (V2A) generation, which aims to transform a raw vocal recording into a fully arranged accompaniment, inherently requires jointly addressing an accompaniment trilemma: preserving acoustic authenticity, maintaining global coherence with the vocal track, and producing dynamic orchestration across a full song. Existing open-source approaches typically make compromises among these goals. Continuous-latent generation models can capture long musical spans but often struggle to preserve fine-grained acoustic detail. In contrast, discrete autoregressive models retain local fidelity but suffer from unidirectional generation and error accumulation in extended contexts. We present LaDA-Band, an endto-end framework that introduces Discrete Masked Diffusion to the V2A task. Our approach formulates V2A generation as Discrete Masked Diffusion, i.e., a global, non-autoregressive denoising formulation that combines the representational advantages of discrete audio codec tokens with full-sequence bidirectional context modeling. This design improves long-range structural consistency and temporal synchronization while preserving crisp acoustic details. Built on this formulation, LaDA-Band further introduces a dualtrack prefix-conditioning architecture, an auxiliary replaced-token detection objective for weakly anchored accompaniment regions, and a two-stage progressive curriculum to scale Discrete Masked Diffusion to full-song vocal-to-accompaniment generation. Extensive experiments on both academic and real-world benchmarks show that LaDA-Band consistently improves acoustic authenticity, global coherence, and dynamic orchestration over existing baselines, while maintaining strong performance even without auxiliary reference audio. Codes and audio samples are available at https://github.com/Duoluoluos/LaDA-Band .

Abstract:
Vision-Language-Action (VLA) models have recently emerged as a promising paradigm for end-to-end autonomous driving by enabling unified reasoning across perception, language, and planning. However, existing approaches lack mechanisms to exploit past failures or adapt to distribution shifts, causing the model to persistently underperform on similar scenarios where it has previously failed. In this paper, we propose DriveVLA-M0, a retrieval-augmented VLA with failure-aware latent memory. We construct a latent memory pool that stores failure cases along with their structure scene representations and expert trajectory labels, and design a dedicated Retrieve Model that decouples static road structure and dynamic agent interactions to enable structurally grounded retrieval. At inference time, retrieved cases are injected into the model via a lightweight decoupled LoRA-based test-time training (TTT) mechanism, allowing targeted and scenario-specific correction without modifying the backbone. Extensive experiments on NAVSIMv1 and NAVSIMv2 benchmark demonstrate that our approach consistently outperforms prior methods, achieving 94.1 PDMS on Navtest and 47.0 EPDMS on Navhard with only 26.44 ms TTT backward latency overhead. Furthermore, we show that DriveVLA-M0 scales effectively with additional memory, enabling training-free performance gains through memory expansion. The code is available at https://github.com/ZebinX/DriveVLA-M0.

Abstract:
Video object insertion places a user-specified object in an existing dynamic scene. Existing methods typically condition generation on text or a single reference image. Consequently, object appearance is underconstrained under viewpoint changes, often leading to identity drift, incorrect foreground-background layering, boundary artifacts, and temporal flickering. In this paper, we propose a video object insertion framework that incorporates multi-view object priors to address these limitations. The framework lifts a 2D reference image into a multi-view representation and uses view-consistent conditioning to provide stable identity guidance and view-adaptive appearance cues. A quality-aware weighting mechanism reduces the influence of noisy or imperfect reconstructed views. We further introduce an Integration-Aware Consistency Module that promotes plausible occlusion, clean boundaries, and temporal continuity. Experiments demonstrate that the proposed framework improves visual quality, controllability, identity consistency, and foreground-background integration for video object insertion compared to the baseline methods. Project page: \urlhttps://polarisxq.github.io/MOVI/.

Abstract:
To address the challenges of modeling hierarchical structures and complex motion in semi-supervised video object segmentation, we propose HyperVOS. By embedding features into a Poincar Ball, HyperVOS exploits hyperbolic properties, namely central flatness and peripheral exponential expansion, to achieve a more nuanced representation of global semantics and local details than is possible in Euclidean space.The framework consists of three core stages: Spatial Hierarchical Representation, Hyperbolic Temporal Flow and Hyperbolic Memory Matching.First, we propose Spatial Hierarchical Encoding, which anchors global structures at the manifold center via Hyperbolic Global Aligner (HGA) while pushing fine textures toward the edges using the Hyperbolic Local Adapter (HLA), achieving automatic hierarchical decoupling of features.Next, we introduce Hyperbolic Temporal Flow (HTF) to model motion as a trajectory of geometric states on the manifold. This approach enables dynamic parameterization of the manifold based on kinetic cues.To suppress identity drift in long sequences, we introduce Hyperbolic Memory Matching. We incorporate Hyperbolic First Frame Anchor (HFA) and utilize Hyperbolic Geodesic Memory Matcher (HGM) based on geodesic distance to perform high precision multiscale feature retrieval.Experimental results on standard VOS benchmarks demonstrate that HyperVOS exhibits superior robustness when handling fast motion and occlusion scenarios, significantly outperforming existing Euclidean space VOS methods.We have released source code on https://github.com/RaRYoMe/HyperVOS.

Abstract:
Multimodal attributed graphs (MAGs) combine multimodal node attributes with structured relations. However, existing methods usually perform shared message passing on a single graph and implicitly assume that the same neighbors are equally useful for all modalities. In practice, neighbors that benefit one modality may interfere with another, blurring modality-specific signals under shared propagation. To address this issue, we propose RoleMAG, a multimodal graph framework that learns how different neighbors should participate in propagation. Concretely, RoleMAG distinguishes whether a neighbor should provide shared, complementary, or heterophilous signals, and routes them through separate propagation channels. This enables cross-modal completion from complementary neighbors while keeping heterophilous ones out of shared smoothing. Extensive experiments on node classification and link prediction across three graph-centric MAG benchmarks show that RoleMAG achieves the best results on RedditS and Bili\_Dance, while remaining competitive on Toys. Ablation, robustness, and efficiency analyses further support the effectiveness of the proposed role-aware propagation design. The source code and supplementary material are available in our \hrefhttps://github.com/laonuo2004/RoleMAGpublic repository.

Abstract:
Multi-modal Large Language Models (MLLM) enable training-free Video Anomaly Detection (VAD). However, most existing methods operate only on isolated frames or short clips. This leads to fragmented predictions due to subtle anomalies and short-term occlusion. We propose Prior-enhanced Event Refinement for Video Anomaly Detection (PEER-VAD). It is a training-free, plug-and-play refinement framework that corrects arbitrary baseline predictions via prior disambiguation and bidirectional window analysis. PEER-VAD transforms fragmented frame-level predictions into semantically consistent event-level predictions. Extensive experiments demonstrate AUC gains, including 4.5% from refining LAVAD on UCF-Crime and 3.1% from refining AnomalyRuler on ShT. Code is available at https://github.com/HaochengY/PEER-VAD.

Abstract:
Efficient solutions for monocular depth estimation (MDE) constitute fundamental components in robotics, autonomous driving, and 3D reconstruction, particularly for resource-constrained platforms requiring real-time processing. Recent research has explored leveraging knowledge distillation to generate more lightweight methods. Despite progress in this learning paradigm, distilling knowledge to compact networks faces multiple challenges: 1) The reliability of teacher supervision often fails to remain consistent across all pixel locations. 2) Wide depth value ranges and scale inconsistency complicate the implementation of normalization. 3) Heterogeneous teacher-student architectures may compromise knowledge transfer effectiveness. To overcome these limitations, we propose the unified knowledge distillation framework RD-KD. This framework integrates three core components: Teacher Reliability-Boosted Mask (TRM) filters low-reliability regions, Soft Segment Normalization (SSN) aligns depth while suppressing boundary artifacts, and Distribution-Level Semantic Alignment Module (DSAM) reduces feature discrepancies. By synergistically leveraging these components, we achieve significant performance gains during distillation. Comprehensive experiments demonstrate that our framework consistently outperforms existing lightweight depth estimators across multiple benchmarks, delivering dual improvements in accuracy and generalization while maintaining compact parameters and computational efficiency. Source code is available at \urlhttps://github.com/XTU-PR-LAB/RD-KD.

Abstract:
Recent vision-language models for chest X-ray understanding are largely built on image-report alignment and therefore rely heavily on MIMIC-CXR as the dominant pretraining source. While effective at scale, this paradigm underexplores an important alternative source of supervision: a range of existing multi-label classification datasets, which provide cleaner and more explicit disease signals than free-text reports, and can offer broader pathology coverage when combined across sources. However, learning from such heterogeneous datasets is nontrivial, as differences in label ontologies, annotation protocols, acquisition pipelines, and report styles can cause models to entangle clinical semantics with dataset identity, leading to poor transfer despite increased scale. In this work, we revisit radiology VLM construction from the perspective of harmonized multi-source learning. We propose HarMoE, a dataset-aware mixture-of-experts framework that learns shared cross-dataset medical semantics while confining source-specific variation to lightweight residual experts in deeper decoder layers. To further exploit clean supervision from labeled datasets, we train in a unified disease vocabulary with masked multi-dataset supervision, enabling the model to leverage complementary annotations without introducing false negatives. Experiments on large-scale chest X-ray benchmarks show that HarMoE consistently improves zero-shot classification, out-of-distribution transfer, and grounding over strong baselines. Our results suggest that building robust radiology VLMs requires moving beyond single-source image-report alignment toward structured knowledge construction from heterogeneous datasets with cleaner supervision and broader coverage. Code and the 873k harmonized dataset will be released at https://github.com/Roypic/harmoe.

Abstract:
Vision-Language-Action (VLA) models, as large foundation models for embodied control, have shown strong performance in manipulation tasks. However, their performance comes at high inference cost. To improve efficiency, recent methods adopt action chunking, which predicts a sequence of future actions for open-loop execution. Although effective for reducing computation, open-loop execution is sensitive to environmental changes and prone to error accumulation due to the lack of close-loop feedback. To address this limitation, we propose Speculative Verification for VLA Control (SV-VLA), a framework that combines efficient open-loop long-horizon planning with lightweight closed-loop online verification. Specifically, SV-VLA uses a heavy VLA as a low-frequency macro-planner to generate an action chunk together with a planning context, while a lightweight verifier continuously monitors execution based on the latest observations. Conditioned on both the current observation and the planning context, the verifier compares the planned action against a closed-loop reference action and triggers replanning only when necessary. Experiments demonstrate that SV-VLA combines the efficiency of chunked prediction with the robustness of closed-loop control, enabling efficient and reliable VLA-based control in dynamic environments.

Abstract:
Although text-to-image diffusion models generally exhibit strong prompt-following ability, we identify a persistent and previously underexplored failure pattern in which a small subset of prompts differing only in the object consistently fails to realize the same target concept under identical generation settings. We term this phenomenon object-dependent concept brittleness. Such cases suggest systematic internal blind spots rather than random sampling noise. In this paper, we present an interpretability-oriented framework to audit and minimally correct these failures. Our key idea is to analyze denoising trajectories in a step-wise sparse autoencoder (SAE) space, where abstract style and attribute concepts become more separable than in the raw denoising representation. This sparse space enables us to compare successful and failed generations, identify concept dimensions whose evidence is missing, weakened, or temporally delayed, and construct class-level concept prototypes from reliable class-consistent samples. Based on this audit process, we introduce a lightweight inference-time correction strategy that interpolates denoising features toward the corresponding prototype in SAE space. Rather than serving as a task-specific retraining method, this intervention acts as a validation of the diagnosed concept deficiency. We evaluate the proposed framework on style and attribute failure cases across multiple diffusion backbones, with significant improvements in concept consistency, text fidelity, and repair success. Further analyses show that deeper denoising representations provide clearer concept structure, while early-stage intervention offers the strongest correction leverage. Code is available at GitHub.

Abstract:
Dunhuang mural restoration aims to reconstruct faithful semantic content within degraded regions, which is crucial for the continuation of cultural heritage. However, cross-cave style diversity introduces severe inter-domain confounding, making semantic reconstruction susceptible to cave style interference, which yields spurious textures. To mitigate this issue, we revisit mural restoration from a causal perspective, formulating cave styles as confounding factors that interfere with intrinsic semantic recovery. We propose a Causal Deconfounding for Cross-Cave Style-Aware Dunhuang Mural Restoration (ReStyleDH), which explicitly blocks the confounding influence of cave styles through causal intervention. To render cave style variables explicit and intervenable, we first construct a cave-stratified mural dataset. Building upon this, our approach employs a Causal Graph Inference module to reconstruct causal semantic content via graph-based context aggregation. Subsequently, a Style Intervention module performs a formal causal intervention by retrieving and injecting target cave styles, which effectively blocks the spurious correlation between semantics and styles. Extensive experiments demonstrate that ReStyleDH achieves superior restoration quality in cross-cave scenarios, validating the effectiveness of causal deconfounding for style-aware mural restoration. Our code is available at \urlhttps://github.com/J-B-Zh/ReStyleDH.

Abstract:
Reinforcement learning (RL) has been successfully applied to autoregressive (AR) and diffusion models. However, extending RL to hybrid AR--diffusion frameworks remains challenging due to interleaved inference and noisy log-probability estimation. In this work, we study masked autoregressive models (MAR) and show that the diffusion head plays a critical role in training dynamics, often introducing noisy gradients that lead to instability and early performance saturation. To address this issue, we propose a stabilized RL framework for MAR. We introduce multi-trajectory expectation (MTE), which estimates the optimization direction by averaging over multiple diffusion trajectories, thereby reducing diffusion-induced gradient noise. To avoid over-smoothing, we further estimate token-wise uncertainty from multiple trajectories and apply multi-trajectory optimization only to the top-k% uncertain tokens. In addition, we introduce a consistency-aware token selection strategy that filters out AR tokens that are less aligned with the final generated content. Extensive experiments across multiple benchmarks demonstrate that our method consistently improves visual quality, training stability, and spatial structure understanding over baseline GRPO and pre-RL models. Our code is available at \hrefhttps://github.com/AMAP-ML/mar-grpo\textcolorbluehttps://github.com/AMAP-ML/mar-grpo.

Abstract:
Most existing cross-modal hashing methods rely on pairwise or triplet supervision, which provides only fragmented local guidance and neglects the global semantic structure of heterogeneous samples. Such limited supervision fails to capture how all samples collectively shape the geometry of the cross-modal embedding space, leading to locally aligned but globally inconsistent hash representations with reduced discriminability. Although some previous studies have attempted to establish interrelationships between more samples, they are still limited to local relational metrics rather than global geometric distributions. To address this issue, we propose Deep Discriminative Global Relevance Hashing (DDGRH), a novel framework that introduces Global Relevance Propagation (GRP) into deep cross-modal hashing. Specifically, DDGRH constructs a semantic graph within each batch and employs an iterative message propagation mechanism to propagate relevance across heterogeneous embeddings, enabling the network to perceive the global distribution of multi-modal samples. By jointly optimizing local relational metric loss and global semantic preservation loss, the model learns a discriminative embedding co-space across modalities. Extensive experiments on three benchmark datasets demonstrate that DDGRH significantly outperforms state-of-the-art hashing methods in retrieval accuracy and discriminability. The source code is available in \urlhttps://github.com/QinLab-WFU/DDGRH.

Abstract:
The rapid evolution of autonomous agents has enabled increasingly complex long-horizon tasks in multi-modal environments. However, these systems remain vulnerable: subtle reasoning errors or visual misinterpretations can propagate through an execution trajectory and ultimately cause task failure. Failure Attribution (FA) faces two key challenges. (1) The heterogeneous nature of visual and textual modalities hinders accurate semantic alignment, making it difficult to ground high-level intentions in fine-grained visual evidence. (2) Long execution trajectories obscure long-range dependencies, where early reasoning errors often surface only after many subsequent actions. To address these challenges, we present VISTA, a training-free, plug-and-play \ulineVisual-textual \ulineInspector for \ulineStep-wise \ulineTrajectory Failure \ulineAttribution. VISTA decomposes diagnosis into Perception, Analysis, Deduction, and Evaluation, bridging cross-modal gaps through explicit visual-textual alignment while mitigating error propagation via milestone summarization and forward state extrapolation. Experiments across diverse benchmarks show that VISTA consistently outperforms existing baselines, providing an interpretable \hrefhttps://github.com/JLZhong23/VISTA_MM_Failure_Attribution\textcolorpurpleframework for failure diagnosis, agent harness and real-time optimization.

Abstract:
Interactive prompts in online video must be engaging yet lightweight enough not to disrupt viewing. Existing video question generation (VQG) systems are predominantly designed to produce questions answerable directly from the video content itself---a paradigm fundamentally misaligned with the goal of stimulating viewer curiosity and deeper exploration. We introduce Video-triggered Curiosity Question Generation (VCQG), a new task that breaks this convention: rather than asking what the video can answer, we ask what it should inspire. Our questions are temporally grounded in the current scene, yet deliberately extend beyond it---probing knowledge, context, and implications that the video alone cannot resolve. This demands that a capable system exhibit strong domain generalizability, safety-awareness against harmful prompts, real-world format compatibility, and compelling engagement quality. To address these challenges, we propose PIQUE, an agentic reinforcement learning framework that tackles inaccurate step-level credit assignment and unstable optimization under heterogeneous constraints. PIQUE combines Trajectory-aware Selective Reinforcement (TSR), which suppresses updates on unreasonable tool-use steps, with Curriculum Reward Shaping (CRS), which progressively introduces reward signals from basic validity to higher-level quality. On the test set, PIQUE-8B achieves 39.6% High-Quality and 63.9% Qualified rates, outperforming strong closed-source and open-source baselines, demonstrating deeper video understanding and effective user engagement. Code is released at \hrefhttps://github.com/pikapi-pipipi/PIQUECode URL.

Abstract:
The rapid advancement of generative technologies has made synthetic images increasingly difficult to distinguish from real ones, creating an urgent need for robust detectors that can counter misinformation. Existing methods, however, mainly rely on universal artifact cues shared across generators. We observe that the intersection of these cues shrinks as generator diversity increases, which weakens generalization. Conversely, relying only on generator-specific artifacts can overfit the detector to particular forgery patterns. To address this challenge, we propose LEGO (LoRA-Enabled Generator-Oriented Framework). LEGO uses an MLP to modulate multiple Low-Rank Adaptation (LoRA) blocks, each pretrained to capture artifacts from a specific generator, and then fuses their features through attention. Unlike methods that seek a single universal solution, LEGO learns and composes specialized artifact representations in two stages. First, each LoRA module is trained on a single-generator dataset to learn generator-specific representations. The LoRA modules are then frozen, and the MLP router and attention layers are trained on mixed-generator data to regulate each module's contribution dynamically. This modular design allows LEGO to incorporate new LoRA modules as generators emerge. With fewer than 30,000 training images, less than 10% of the training data used by major baselines, and only five epochs per stage, LEGO nevertheless achieves substantially better average performance than prior state-of-the-art methods. The code is available at https://github.com/seanstep/LEGO_LoRA_Enabled_Generator_Oriented_Framework.

Abstract:
Rendering clear and accurate text is a critical capability of text-to-image (T2I) models. Despite rapid advances in T2I generation, the systematic and human-aligned evaluation of text rendering remains underexplored. Existing benchmarks often suffer from limited scenario coverage, a lack of bilingual support, and an over-reliance on unvalidated automated metrics. To address this gap, we introduce AGTI-Bench, a dedicated benchmark for evaluating text-aware AI-generated text images (AGTIs). AGTI-Bench is characterized by three core features: (1) Bilingual and Diverse Scenarios, covering typical real-world text-bearing image types with native Chinese--English support; (2) Multi-dimensional Evaluation, focusing on text accuracy, text rendering quality, and overall image quality; and (3) Human Preference Alignment, supported by large-scale subjective annotations. Our analysis reveals that while OCR-based methods excel at verifying text content, multimodal large language model (MLLM) evaluators achieve stronger alignment with human preferences for perceptual quality assessment. Furthermore, our benchmarking of representative T2I models uncovers significant performance disparities across the three evaluation dimensions, underscoring the necessity of multi-dimensional assessment for text-aware image generation. The dataset and code are publicly available at \urlhttps://github.com/mzcha0/AGTI-Bench-dataset.

Abstract:
Visual tokenizers and VAEs are core components of modern visual generation and multimodal models, yet the information loss from compression fundamentally limits generation fidelity. We show that conventional evaluation metrics such as PSNR, SSIM, and LPIPS fail to reliably assess reconstruction quality for text and faces, two content types that are particularly vulnerable to degradation under compression and highly sensitive to human perception. To address this, we propose TokBench, a benchmark comprising over 12K images and 400 video clips rich in text and face content, curated from diverse public datasets spanning scene text, documents, and naturalistic multi-face scenarios. We introduce task-aware metrics leveraging established OCR and face recognition models, providing a lightweight yet reliable evaluation pipeline. Using TokBench, we conduct comprehensive evaluations of existing image and video tokenizers across various resolutions and object scales, revealing that current tokenizers still exhibit significant deficiencies in preserving fine-grained text and facial features. A user study further confirms that our proposed metrics align well with human judgments, serving as an effective complement to conventional metrics. Our benchmark and evaluation code are publicly available at https://wjf5203.github.io/TokBench.

Abstract:
Video traffic constitutes a significant share of global web traffic. To reduce its volume, video codecs have been developed and continuously improved. While the industry has achieved substantial progress in traditional video coding, neural video codecs (NVCs) have recently emerged as a new approach that applies deep learning to video compression. This creates new challenges for compression quality assessment, which is essential for the further development and improvement of such codecs. In particular, it is important to evaluate the novel temporal compression paradigms introduced by NVCs. In this work, we present a large-scale subjective dataset of videos compressed with both neural and traditional video codecs. The subjective scores were collected through crowd-sourced pairwise comparisons. The proposed dataset provides a valuable resource for the development and benchmarking of video quality metrics tailored to neural video codecs. The dataset is available at the following link: https://videoprocessing.github.io/nvc-dataset-benchmark.

Abstract:
Zero-shot classification needs efficient label retrieval and fine-grained visual reasoning, yet discriminative and generative vision-language models fail in complementary ways. When CLIP's top-1 prediction is wrong, the correct label often remains in its top-K shortlist, making disambiguation rather than recall the key challenge. Standalone generative models, however, are hindered by large label spaces and unconstrained outputs. This complementarity motivates separating broad candidate retrieval from fine-grained, image-grounded verification. We propose G2D, a training-free framework that uses a generative VLM to verify CLIP-retrieved candidates against the image. Candidate names and CLIP probabilities provide a structured prior for resolving visually similar classes. Fixed confidence routing, entropy-adaptive candidate sizing, and trie-constrained decoding focus generative reasoning on uncertain samples and ensure one valid output for each input at test time. Across eight benchmarks, G2D achieves 68.85% average accuracy, versus 59.35% for CLIP and 63.11% for the standalone VLM. Across seven generator configurations, candidate-set verification improves average accuracy by 1.08--27.42 percentage points. G2D also transfers to DCLIP, WaffleCLIP, and CuPL, supporting a practical interface between discriminative proposal and generative visual reasoning. Code:https://github.com/Harzva/G2D.

Abstract:
One-dimensional (1D) image tokenizers have emerged as key enablers for efficient visual generation by compressing images into compact discrete sequences. However, existing approaches adopt a coupled architecture that concatenates learnable latent tokens with image patches at the input. This tight entanglement creates three critical bottlenecks: (1) structural incompatibility with powerful pretrained foundation models, (2) quadratic complexity that hinders scaling to larger token budgets, and (3) a representation bottleneck caused by single-scale feature extraction. To address these limitations, we propose QTok, a novel framework that introduces a paradigm shift to decoupled tokenization. We reformulate 1D tokenization as an ``Encode-then-Query'' process: a standard encoder first extracts a Hierarchical Multi-Level Memory to capture both global semantics and fine-grained details. Subsequently, a lightweight Query Attention Module distills this memory into compact latent tokens via cross-attention. This decoupling converts the scaling complexity from quadratic to linear and enables seamless integration with foundation models. Experiments on ImageNet-1K demonstrate that QTok establishes a new state-of-the-art for 1D tokenizers in generation fidelity, semantic consistency, and efficiency. Notably, our empirical results indicate that leveraging pretrained visual encoders achieves superior performance with only 50% of the standard training schedule. Code is available at \urlhttps://github.com/Wenzhao-Xiang/QTok.

Abstract:
Video slicing is the foundation of video content creation. However, existing automated video trimming methods suffer from coarse granularity due to fixed-duration splitting, misplaced dialogue breakpoints in vision-only approaches, and excessive redundancy from coarse scene-level segmentation. This paper presents SeamSlicer, which is compatible with existing segment-based captioning methods. Our core insight is that the semantic integrity of atomic units depends on comprehensive multimodal analysis, integrating visual continuity, narration, multi-speaker dialogue, background music, sound effects, on-screen text, and ambient audio. SeamSlicer consists of three stages: multimodal boundary extraction stage for candidate boundary detection, visual boundary refinement stage for maintaining semantic completeness, and a multi-strategy proper length control stage for length control. We also construct AtomicBench, a pioneering multimodal dataset specifically designed for video trimming segmentation tasks, addressing the limitations of existing datasets by providing atomic-level annotations that ensure both visual smoothness and semantic coherence. Extensive experiments show that our method performs well on boundary detection, and the trimmed videos outperform existing methods across multiple dimensions, validating that this approach can provide high-quality material for automated video trimming. The code is available at https://github.com/suiyingliuxin/SeamSlicer.

Abstract:
As large language models (LLMs) evolve into tool-augmented agents, tool selection becomes a key interface between the model and external resources. However, existing methods struggle to model interactions among tools and generally overlook the causal structure inherent in the tool selection process. In this paper, we propose Autoregressive Tool Selection (ARTS) to explicitly model causal dependencies among tools. Specifically, ARTS formulates toolset construction as a sequential decision-making problem, where the next tool is selected conditioned on the query and the previously selected tools. To further capture rich interactions during this process, ARTS employs a graph neural network (GNN) to jointly model tooltool and toolquery relationships. In practice, ARTS constructs a graph comprising a query node and candidate tool nodes and explicitly marks the nodes of the selected tools. The resulting marked graph is fed into a GNN to autoregressively predict the next tool until a termination condition is met. Experiments on two public benchmark datasets demonstrate that ARTS consistently outperforms relevant baselines in terms of accuracy and exhibits clear advantages in inference latency. The code is available at \urlhttps://github.com/qizhihuanying/ARTS_MM.

Abstract:
Recent geometric foundation models (e.g., Metric3D, Depth Anything and UniDepth) have greatly advanced monocular depth estimation (MDE), yet their cross-scene and metric-scale capabilities remain difficult to transfer to tiny models. We bridge this gap with DepthART (Depth Anything Rethought for Tiny Models), a compact MDE model for on-device deployment. We target two capacity-driven failures: (i) overfitting to dataset-specific distribution bias and (ii) geometry forgetting during metric fine-tuning. Accordingly, DepthART combines two simple strategies: a bias-resistant data sampling scheme to reduce distribution bias under the same training budget, and camera-conditioned fine-tuning to recover metric scale from intrinsics with a frozen encoder. Across datasets, DepthART consistently surpasses previous tiny baselines in zero-shot generalization and metric depth (e.g., zero-shot \delta_1\!\!=\!\!0.964 for DepthART-S on NYUD v2), and sometimes approaches much larger models. In model-only inference, DepthART-S reaches 347/245 FPS (strict FP32) on an RTX A6000 at 224^2/448^2, 102 FPS (FP32 with TF32 enabled) on an Orin NX 8GB, and over 15 FPS (FP32) on a Jetson Nano 4GB. Code: \hrefhttps://github.com/xuefeng-cvr/DepthARTGitHub; Project: \hrefhttps://xuefeng-cvr.github.io/DepthARTDepthART.

Abstract:
Diffusion models have shown remarkable success in generating 3D human motion from text descriptions, demonstrating superior realism and diversity. However, directly modeling the full raw motion sequence is computationally intractable. Existing diffusion-based methods either compromise essential computations on the raw sequence or perform low-dimensional modeling in a latent space, inevitably introducing semantic deviations. To overcome this challenge, we propose SegMoDiff, a text-to-motion diffusion framework that enables adequate and fine-grained semantic modeling directly on raw sequences by partitioning them into tractable segments. SegMoDiff pivots on a segmented motion diffusion transformer (SegMoDiT), where overlapping contextual attention is introduced to preserve cross-segment coherence. On top of this, segment-wise text enhancement based on large language models is leveraged to achieve precise text-motion alignment. To further accelerate inference speed, we incorporate distribution matching distillation to reduce sampling steps. Extensive experiments demonstrate that our method significantly outperforms state-of-the-art methods in semantic fidelity while delivering superior inference speed. The code is available at https://github.com/wjxxyz/SegMoDiff.

Abstract:
2D Gaussian Splatting is an attractive direction for image representation due to its explicit formulation, fast rasterization, and favorable decoding efficiency. The representation quality of this paradigm depends on the proper allocation of Gaussian capacity to the demanding regions. However, existing methods fail to allocate Gaussian capacity efficiently during optimization: under-reconstructed content is often refined in a fragmented pixel-wise manner, while neighboring optimized Gaussians with similar attributes are redundantly retained. This inefficiency motivates the need for a density control framework that jointly addresses insufficient allocation in under-reconstructed regions and redundant allocation in over-reconstructed regions. Our key insight is that this framework should exploit two complementary forms of locality: the local continuity of reconstruction errors in image space for improved Gaussian allocation, and the local similarity of neighboring Gaussians in Gaussian space for redundant elimination. Based on this insight, we propose Locality-Aware Density Control (LocoADC), a plug-and-play framework that improves Gaussian capacity utilization through Region-wise Gaussian Densification (RGD) and Similarity-Driven Gaussian Merging (SDGM) strategies, together with a local color consistency constraint for more reliable merging. Extensive experiments on diverse datasets show that LocoADC consistently improves multiple baselines by enabling more effective local Gaussian allocation, including a 2.93 dB PSNR gain over GI on the CLIC dataset under the same 30k Gaussian budget. Code is available at: https://github.com/ChenJiaCong-1005/LocoADC.

Abstract:
EEG-based visual decoding provides a non-invasive pathway to interpret visual semantics. However, existing methods ignore the perceptual asymmetry between foreground and background in complex scenes, leading to background interference and semantic misalignment. Compounding this challenge, EEG signals exhibit rapid temporal dynamics and non-stationary spatial patterns, posing a challenge for capturing the time-varying brain connectivity activated by focal visual attention. To address these limitations, we propose FSDBN, a unified framework for robust EEG-visual decoding. Specifically, we introduce a Semantic-Consistent Saliency Alignment (SCSA) module to disentangle semantic foregrounds from background noise under the saliency-semantic joint constraints. Furthermore, we devise a Semantic-Prior Dynamic Gating Foreground Fusion to adaptively regulate the contributions of foreground and background features for better representation learning. Correspondingly, we model EEG signals as adaptive spatiotemporal brain networks, where functional connectivity dynamically reorganizes to capture neural responses to salient foregrounds, facilitating robust alignment between dynamic neural representations and the visual structure. Our method achieves a top-1 accuracy of 69.0% and a top-5 accuracy of 92.2% on the zero-shot brain-to-image retrieval task, surpassing previous state-of-the-art methods. Code is available at \urlhttps://github.com/xbrainnet/FSDBN.

Abstract:
Recent studies have extended chain-of-thought (CoT) reasoning to Vision-Language Models (VLMs) to improve their performance in complex multimodal tasks. However, the underlying mechanisms of how VLMs learn inference patterns from CoT reasoning to process multimodal input information remain in a black-box state. To this end, we define and quantify the inference patterns learned by VLMs from CoT prompts, which help us analyze the representation characteristics of VLMs and reveal the mechanisms underlying CoT reasoning in VLMs. Specifically, we disentangle the output score of a VLM into a set of interactions between input variables in the CoT prompt and take these interactions as detailed inference patterns used by the VLM. Based on the interaction patterns, we analyze several classical VLMs and draw the following new insights on CoT reasoning in VLMs. (1) The interaction patterns used by VLMs exhibit high complexity, which means that VLMs can comprehend and internalize complex logical chains constructed by CoT prompts. (2) The interaction patterns used by VLMs manifest a high degree of adjacency, which implies that VLMs can truly learn and understand the adjacent logic underlying CoT prompts. (3) The text-based interaction patterns used by VLMs change accordingly when the visual information in input samples changes, which indicates that VLMs can model cross-modal inference patterns from CoT prompts. The code is available at \urlhttps://github.com/Error404-stack1024/CoT-Interaction.

Abstract:
Rendering complex reflections of real-world scenes using 3D Gaussian splatting has been a quite promising solution for photorealistic novel view synthesis, but still faces bottlenecks especially in rendering speed and memory storage. This paper proposes a new Hybrid Splatting (HybridSplat) mechanism for much faster complex reflection scene rendering with significantly reduced Gaussian primitives, while preserving a similar level of high-fidelity rendering quality compared with state-of-the-art Gaussian tracing. Our first novel idea is a new reflection-baked Gaussian tracing method, which bakes the view-dependent reflection within each Gaussian primitive. Thus we can hybridly integrate the reflective Gaussian primitives with base Gaussian primitives together using the efficient tile-based Gaussian splatting framework for high-fidelity complex scene rendering. Moreover, we further introduce end-to-end reflection-sensitive Gaussian pruning to reduce the Gaussian primitives for lower memory storage while preserving the reflection rendering quality. Through extensive evaluation, our HybridSplat accelerates about 7× rendering speed across complex reflective scenes from Ref-NeRF and NeRF-Casting with 4× fewer Gaussian primitives than similar ray-tracing based Gaussian splatting baselines. Project Page: https://aetheryne.github.io/HybridSplat-Page/

Abstract:
\beginabstract Text-to-image diffusion models enable personalization of specific visual concepts from a small number of reference images. However, generating a single image that contains multiple personalized subjects, each bound to user-specified attributes such as clothing, accessories, and held objects, remains largely unaddressed. Without explicit spatial constraints, concurrently activated concept checkpoints produce overlapping cross-attention responses, causing per-subject identity degradation and attribute misalignment. Moreover, no established benchmark jointly evaluates these two failure modes in the personalized multi-subject setting. We present MultiCompose, a composition framework that decouples per-concept personalization from multi-subject inference. A semantic preservation regularization maintains attribute binding capacity during fine-tuning, while a two-phase inference procedure automatically establishes subject layout and composes per-concept predictions through spatially exclusive masks. We further introduce MSP-Bench, a benchmark that jointly evaluates identity fidelity (ID), attribute binding accuracy (BIND), and attribute misalignment (MIS) through a dual-pathway protocol. Experiments show that MultiCompose outperforms existing methods on both conventional metrics and MSP-Bench, confirming the benchmark's ability to reveal failure modes that conventional metrics overlook. Code is available at \urlhttps://github.com/I2-Multimedia-Lab/MultiCompose. \endabstract

Abstract:
Collaborative perception enhances the perceptual capabilities of autonomous vehicles by sharing information among multiple agents. However, existing studies typically define "agents" as those equipped with full sensor suites, limiting real-world scalability. In reality, any traffic participant capable of sharing data through a cooperative network should be considered an agent, especially the vehicles that provide only coarse localization data. This broader definition introduces a new open problem: How to effectively integrate such heterogeneous yet widely available localization data into collaborative perception to improve perception performance? To address this problem, we propose CoLD, a novel extensible collaborative perception framework. CoLD first introduces a Position-Aware Mapping Module (PAMM) to bridge the heterogeneity between discrete localization data and dense perceptual features, mapping the former into sparse yet critical heatmaps that guide the model's attention toward key areas. To mitigate localization errors, CoLD introduces a Localization-Uncertainty Denoising Diffusion (LUDD) module. LUDD utilizes Collaborative Occupancy Flow (CO-Flow) as a spatiotemporal condition to guide the diffusion process, enabling robust denoising of perceptual features through implicit alignment. Evaluations on OPV2V, V2XSet, and DAIR-V2X demonstrate that CoLD outperforms state-of-the-art methods in perception, with only a negligible communication overhead. Moreover, the core components are plug-and-play modules, enabling easy integration into existing methods and consistently delivering significant performance gains, demonstrating its versatility and applicability. The code is available at https://github.com/Attenre/CoLD.

Abstract:
Inversion-free image editing using flow-based generative models offers an efficient alternative to prevailing inversion-based pipelines. However, existing approaches rely on fixed Gaussian noise to initialize the source trajectory. This content-agnostic stochastic noise severely disrupts the spatial topology of the original image, causing severe trajectory drift and structural degradation. To address this source--noise mismatch, we introduce SNR-Edit, a training-free framework that actively prevents structural drift right from the initialization phase. Mechanistically, SNR-Edit performs structure-aware noise rectification by extracting semantic masks and injecting their RoPE-encoded spatial-frequency signatures directly into the initial noise. Instead of applying post-hoc trajectory corrections, this approach provides a stable geometric anchor for the generative flow, explicitly encouraging the model to respect the original layout constraints during source--target transport. This lightweight modification ensures high-fidelity structural preservation without requiring model tuning or inversion. Extensive evaluations across SD3 and FLUX on PIE-Bench and SNR-Bench demonstrate that SNR-Edit achieves superior performance on pixel-level metrics and VLM-based scoring, while adding merely ~1s overhead per image. Code is released at https://github.com/Tankowa/SNR-Edit.

Abstract:
Spiking neural networks (SNNs) are promising for edge sensing due to their event-driven computation and temporal filtering capability. However, standard leaky integrate-and-fire (LIF) neurons communicate only through binary spikes, which severely limit representational capacity. Existing multi-level spiking neurons improve information transmission, but often rely on uniform quantization that mismatches membrane-potential distributions or introduces costly synaptic multiplications. In this paper, we propose ShiftLIF, a multi-level spiking neuron that maps membrane potentials to a logarithmically spaced power-of-two spike set. This design provides finer representation in the small-amplitude regime, where membrane potentials are densely concentrated, while enabling multiplier-free synaptic computation through bit-shift and accumulation operations. As a result, ShiftLIF improves spike-level expressiveness without sacrificing the hardware-friendly nature of standard SNN computation. We evaluate ShiftLIF on 10 datasets spanning wireless, acoustic, motion, and visual sensing tasks. Results show that ShiftLIF consistently matches or exceeds the accuracy of existing multi-level spiking neurons while maintaining synaptic energy consumption close to standard binary LIF. These results indicate that ShiftLIF provides a favorable accuracy-efficiency trade-off for cross-modal edge sensing. Our code is publicly available at https://github.com/Kaiwen-Tang/ShiftLIF.

Abstract:
LiDAR point cloud (LPC) compression serves as a main component of 3D representation, where octree-based entropy models emerge as the prevailing technique, benefiting from their superior organization structure and ability to model dependencies. However, existing methods usually rigidly map rich siblings to fixed sequence indices, leading to context-entropy misalignment and hindering feature extraction. To address this problem, we propose EADR, an Entropy-Aware Dynamic Routing model for LPC compression, which combines a difficulty-aware gating network to dynamically route nodes and an interleaved context coupling backbone to comprehensively fuse routing priors. Specifically, a gating mechanism routes nodes through coding-cost supervision to overcome the inflexibility of topological splitting, and a grouping operation constrains routing to ensure local spatial coherence. Besides, we propose an interleaved context coupling backbone, which drives a coupled globallocal refinement loop to enhance contextual representations and provide effective routing priors. The experimental results show that our EADR outperforms state-of-the-art works, achieving 41.07% and 54.00% BD-Rate gains on SemanticKITTI and Ford datasets, respectively. Moreover, it allows extension to downstream tasks (e.g., vehicle detection and semantic segmentation), demonstrating its practical utility. The code is available at https://github.com/LPCcompression/EADR-framework.

Abstract:
Video anomaly detection (VAD) is a critical yet challenging task due to the complex and diverse nature of real-world scenarios. Traditional deep learning approaches are fundamentally limited by poor generalization across diverse scenarios. While multimodal agents offer a promising tool-learning paradigm for VAD, current systems relying on supervised fine-tuning struggle with complex orchestration, and standard reinforcement learning often causes premature termination due to coarse-grained outcome rewards. To address these challenges, we propose VTO, a process-supervised reinforcement learning framework. Moving beyond static tool usage, VTO enables the agent to dynamically explore and interact with the environment. Specifically, we introduce a foundation model-driven cognitive evaluator to provide context-aware semantic feedback, which is seamlessly integrated into a Process-Supervised Cognitive Alignment that delivers fine-grained, step-wise supervision. By explicitly penalizing logical truncation and rewarding complete causal chains, the agent optimizes its multi-step reasoning policy for interrelated tool orchestration. To support our proposed framework, we meticulously crafted VAD-Tool, a hierarchical visual tool set comprising 12 specialized vision tools spanning from entity tracking to high-stakes hazard detection, and established the corresponding benchmark for rigorous multi-step reasoning evaluation. Extensive experiments on VAD-Tool demonstrate that VTO significantly outperforms baselines, achieving up to a 10.2% absolute accuracy improvement in tool scheduling. Code and data are available at \urlhttps://github.com/MICLAB-BUPT/VTO.

Abstract:
Processing continuous video streams requires multimodal large language models (MLLMs) to maintain temporally coherent memory. Existing architectures predominantly rely on heuristic feature similarity, such as cosine similarity, for memory updates, making them blind to high-level semantics and susceptible to catastrophic fragmentation from camera ego-motion. To overcome this, we construct SemBound-1K, a multi-dimensional soft-labeled dataset that factorizes video boundaries into visual, action, and subject transitions while explicitly penalizing camera noise. Building upon this foundation, we propose SemStream, an intent-driven streaming architecture realized through semantic boundary-aware memory management. SemStream replaces heuristic matching with a learned Semantic Continuity Controller (SCC) for dynamic event partitioning, and employs a Budgeted Episode Merger (BEM) for long-term memory consolidation under strict token constraints. Extensive experiments demonstrate that SemStream achieves state-of-the-art performance, with accuracies of 78.6% on StreamingBench, 74.1% on MLVU, and 62.3% on ODV-Bench. By natively modeling semantic transitions, SemStream effectively prevents memory fragmentation, enabling robust long-horizon reasoning in complex, unconstrained real-world environments. The project page is available at \urlhttps://jasontan2002.github.io/SemStream.github.io/.

Abstract:
Reinforcement Learning with Verifiable Rewards (RLVR) has improved the reasoning capabilities of Multimodal Large Language Models (MLLMs), yet existing frameworks rely on coarse, sequence-level reward signals that lack the fine-grained supervision over the visually-grounded steps within a multimodal reasoning chain. We investigate this gap through the lens of two token-level metrics: visual dependency (i.e. how much a token's prediction relies on the input image features) and predictive entropy. Our empirical analysis reveals two key findings: (1) correct reasoning chains exhibit a markedly sharper entropy reduction as visual grounding intensifies, compared to incorrect ones; (2) pivotal tokens, those whose misprediction triggers reasoning collapse, are statistical outliers in the joint distribution of visual dependency and predictive entropy derived from correct chains. Motivated by these findings, we propose token-level perception-grounded advantage estimation (TPAE), which estimates token-level advantages by measuring each token's statistical consistency with the vision-entropy patterns of correct rollouts. TPAE leverages this granular score to modulate the sequence-level advantage, producing a fine-grained supervision signal that can be integrated into various RLVR frameworks. Extensive experiments on seven benchmarks show that TPAE consistently outperforms leading strong baselines, yielding more stable and efficient optimization for multimodal reasoning. The code is publicly available at https://github.com/Zhihan72/TPAE.

Abstract:
Personalized text-to-image generation enables users to adapt pretrained generative models to specific concepts through lightweight adapters such as LoRAs. With the rapid growth of community-trained LoRAs, combining multiple independently trained LoRAs at inference time has become an appealing solution for multi-concept generation without additional finetuning. However, naive multi-LoRA combination often encounters parameter conflicts, concept entanglement, spatial confusion, and degraded generation quality. In this paper, we propose LoRAMix, a global-to-local multi-LoRA conciliation framework for test-time text-to-image generation. Our key idea is to coordinate style and content LoRAs according to their distinct scopes of influence on image generation. Specifically, we treat style LoRAs as \bf global modifiers and merge them with content LoRAs to construct style-enhanced content-specific branches, together with magnitude-based sparse pruning and compensation to suppress weak or conflicting parameters. We then treat content LoRAs as \bf local concept controllers and perform region-aware multi-LoRA composition during denoising, where each LoRA mainly acts on its assigned spatial region based on concept-specific masks derived from cross-attention and self-attention. We evaluate LoRAMix on 840 combinations constructed from community-trained LoRAs under three task settings. LoRAMix achieves the best results on all nine metrics, with the largest gains in the style-content setting, improving CLIP-I by 7.53 and DINO by 15.39 over the best baseline. These results validate its effectiveness in concept preservation and style-content coordination. Codes are available at \urlhttps://github.com/ychdata/LoRAMix.

Abstract:
Reconstructing 3D scenes from unordered images remains bottlenecked by expensive Structure-from-Motion (SfM) preprocessing and frozen pose interfaces. We present SalientGS, a unified SfM-to-3D Gaussian Splatting (3DGS) pipeline. Its central contribution is importance-guided Markov Chain Monte Carlo (MCMC) Gaussian allocation, which aggregates multi-view residuals into per-Gaussian underfit and redundancy signals. These signals define a smooth importance-weighted sampling distribution that biases both birth and relocation toward underfit regions. This reallocates capacity from well-fit areas without altering the underlying stochastic gradient Langevin dynamics (SGLD). In a released-code verification over 13 scenes and three benchmarks, SalientGS achieves the best cross-benchmark macro-average PSNR, SSIM, and LPIPS (27.65~dB / 0.876 / 0.147) among the compared methods, while also providing the fastest end-to-end runtime (10.62 minutes) with 1.5M Gaussians. Code, per-scene measurements, and evaluation scripts are available at \urlhttps://github.com/Six-Bit-TX/SalientGS.

Abstract:
Supervised low-level vision models rely on pixel-wise losses against paired references, yet paired training sets exhibit per-pair photometric inconsistency, say, different image pairs demand different global brightness, color, or white-balance mappings. This inconsistency enters through task-intrinsic photometric transfer (e.g., low-light enhancement) or unintended acquisition shifts (e.g., de-raining), and in either case causes an optimization pathology. Standard reconstruction losses allocate disproportionate gradient budget to conflicting per-pair photometric targets, crowding out content restoration. In this paper, we investigate this issue and prove that, under least-squares decomposition, the photometric and structural components of the prediction-target residual are orthogonal, and that the spatially dense photometric component dominates the gradient energy. Motivated by this analysis, we propose Photometric Alignment Loss (PAL). This flexible supervision objective discounts nuisance photometric discrepancy via closed-form affine color alignment while preserving restoration-relevant supervision, requiring only covariance statistics and tiny matrix inversion with negligible overhead. Across 6 tasks, 16 datasets, and 16 architectures, PAL improves metrics and generalization. The implementation is simple at https://github.com/lime-j/paloss.

Abstract:
Nighttime image dehazing remains a formidable challenge due to the complex interplay between non-uniform illumination and dense haze. Existing supervised and semi-supervised methods are inherently limited by their inability to leverage real-world clear nighttime distributions, while conventional unsupervised approaches frequently struggle with unstable domain alignment. To address these issues, we propose MA-UND, a Multi-Perspective Alignment framework for Unsupervised Nighttime Image Dehazing that optimizes synergistic distribution alignment, feedback-driven structural alignment, and perturbative distortion alignment. Specifically, we first establish Synergistic Distribution Alignment via joint self-reconstruction and decoupled branches to anchor scene statistics in low-light nighttime scenes. Building upon this, Feedback-driven Structural Alignment is implemented using a loopback uncertainty mechanism that adaptively gates alignment signals to ensure geometric restoration of dehazed nighttime images. Furthermore, Perturbative Distortion Alignment enforces latent space invariance against non-structural fluctuations, including noise and illumination jitter, through stochastic perturbations. Finally, these alignment objectives are dynamically balanced via a Bayesian-driven weighting strategy. Extensive experiments demonstrate that MA-UND suppresses haze more effectively than state-of-the-art approaches. The code are available at \urlhttps://github.com/lovekang3344/MA-UND.

Abstract:
Reconstructing 3D Hand-Object Interaction (HOI) from monocular RGB images is pivotal for understanding human activities. However, existing methods often assume idealized, single-object scenarios, leading to severe target ambiguity in real-world cluttered environments. To address the inability to distinguish active targets from background distractors, we propose CHORES, an end-to-end reconstruction architecture with geometry-conditioned target disambiguation. Unlike direct regression models, the proposed approach explicitly models the geometric relationship between the hand and potential targets. Specifically, the integration of global metric depth priors provides absolute scale constraints, mitigating monocular depth uncertainty. Central to the framework is the Physical-Aware Interaction Refinement (PAIR) module, which identifies the interaction target by analyzing the geometric affordance between hand poses and candidate objects. Furthermore, a geometry-guided anchor initialization strategy dynamically adapts point queries to ensure precise structure generation. Extensive experiments on the DexYCB, HO3D, and ObMan benchmarks demonstrate that CHORES achieves SOTA performance. Notably, the framework improves the F-score by 4.8% and reduces the Chamfer Distance by 9.1% relative to the strongest compared baseline on the challenging DexYCB dataset, demonstrating its effectiveness under the evaluated multi-object setting. Code and pretrained models will be released soon at \urlhttps://github.com/AiPhDres/CHORES.

Abstract:
Recently, speech deepfake detection (SDD) has achieved significant progress. However, its robustness evaluation remains largely confined to controlled additive noise scenarios, lacking systematic investigation of the complex distortions introduced by acoustic front-end (AFE) processing pipelines in real-world deployments. In this work, we simulate a unified AFE pipeline comprising acoustic echo cancellation, noise suppression, automatic gain control, and voice activity detection (VAD), and conduct a comprehensive evaluation of current state-of-the-art models. The results show that the nonlinear and time-frequency coupled distortions introduced by AFE significantly degrade detection performance. To address this issue, we propose a Time-Frequency Consistency Learning (TFCL) framework, which aims to learn invariant spoofing representations that remain stable before and after AFE processing. We observe that AFE not only introduces temporal misalignment (e.g., segment-level shifts caused by VAD), but also weakens or distorts critical frequency-domain cues. To this end, TFCL employs an attention-driven soft alignment mechanism to capture cross-temporal dependencies, along with frequency-domain structural consistency constraints to enforce feature invariance. As a result, the model is able to maintain stable representations under both temporal perturbations and spectral distortions. Extensive experimental results demonstrate that the proposed method effectively mitigates the performance degradation caused by AFE processing, significantly improving the robustness of SDD in real-world scenarios. The code is available at \urlhttps://github.com/JunXue-tech/TFCL.

Abstract:
Lyric-to-melody generation is an important yet highly challenging task in automatic songwriting. We study the problem of editable lyric-to-melody generation, motivated by the iterative nature of real-world music creation, where lyrics and melodies are frequently revised while preserving surrounding musical context. However, existing methods primarily focus on generation from scratch and lack support for flexible editing, leaving this problem under-explored. In this paper, we propose HE-L2M, a novel hierarchical editable lyric-to-melody generation framework that provides a unified formulation for lyric-to-melody generation and editing by modeling melody editing as a lyric-conditioned infilling problem, enabling flexible modifications across multiple structural levels (word, phrase, sentence, section, and song) while maintaining global musical coherence and accurate lyricmelody alignment. Specifically, we introduce a hierarchical lyric--melody alignment modeling approach to capture accurate lyric--melody alignment, and develop a lyric-conditioned melody infilling Transformer that leverages bidirectional melodic contexts via a reordered decoding strategy. We further incorporate structural conditioning with reference phrases and phrase-level musical attributes to enhance controllability and structural consistency, and adopt a curriculum learning strategy to effectively train the model across diverse hierarchical editing tasks. Extensive experiments demonstrate that HE-L2M achieves superior performance in both melody generation and editing across objective and subjective evaluations. The project is available at \urlhttps://lichaiustc.github.io/HE-L2M/.

Abstract:
Zero-shot voice conversion (VC) aims to convert a source utterance into the voice of an unseen target speaker while preserving its linguistic content. Although recent systems have improved conversion quality, building zero-shot VC systems for interactive scenarios remains challenging because high-fidelity speaker transfer and low-latency streaming inference are difficult to achieve simultaneously. In this work, we present X-VC, a zero-shot streaming VC system that performs one-step conversion in the latent space of a pretrained neural codec. X-VC uses a dual-conditioning acoustic converter that jointly models source codec latents and frame-level acoustic conditions derived from target reference speech, while injecting utterance-level target speaker information through adaptive normalization. To reduce the mismatch between training and inference, we train the model with generated paired data and a role-assignment strategy that combines standard, reconstruction, and reversed modes. For streaming inference, we further adopt a chunkwise inference scheme with overlap smoothing that is aligned with the segment-based training paradigm of the codec. Experiments on Seed-TTS-Eval show that X-VC achieves the best streaming WER in both English and Chinese, strong speaker similarity in same-language and cross-lingual settings, and substantially lower offline real-time factor than the compared baselines. These results suggest that codec-space one-step conversion is a practical approach for building high-quality low-latency zero-shot VC systems. Our code, checkpoints and demo pages are released at \urlhttps://github.com/Jerrister/X-VC.

Abstract:
We propose RefineSVG, a single-step closed-loop visual feedback framework that enables multimodal large language models (MLLMs) to perform high-fidelity image-to-SVG generation through self-correction. Existing MLLM-based approaches rely on single-pass open-loop inference, where the model receives visual input only once and must generate thousands of SVG code tokens without intermediate verification. This paradigm inevitably leads to geometric drift, error accumulation, and visual hallucination on complex images. RefineSVG overcomes this limitation by invoking an external rendering engine after an initial SVG generation pass to compare the rendered output against the target image. The comparison yields a multi-dimensional visual residual map (Diff-Map) that is fed back to the model as a ReAct-style correction signal, driving a targeted correction step. To support this render-observe-correct interaction, we further introduce an SVG-oriented semantic vocabulary that compresses token sequences by over 52%. A progressive training pipeline spanning supervised fine-tuning, rejection-sampling cold-start data construction, and end-to-end agentic reinforcement learning aligns the model with closed-loop visual correction. Extensive experiments show that RefineSVG consistently outperforms existing baselines in reconstruction fidelity, structural accuracy, and code efficiency.

Abstract:
Understanding 3D scenes is fundamental to embodied intelligence, requiring joint reasoning over heterogeneous information from multiple modalities, including visual and geometric cues. However, the relevance of these modalities often varies across queries. Existing Multimodal Large Language Models (MLLMs) typically rely on fixed modality combinations, overlooking query-dependent modality needs. Such a rigid design can introduce semantic noise from irrelevant modalities while underutilizing more informative ones, leading to wasted computation and diluted reasoning. To address these challenges, this paper proposes SmartMage, a unified MLLM that dynamically orchestrates heterogeneous modalities for semantic-aware 3D scene understanding. Specifically, SmartMage incorporates: (1) a Semantic-guided Modality Adaptive RouTing (SMART) module that selects task-relevant modalities using semantic priors, textmodality alignment, and modality quality; and (2) a Modality-Aware Gating Expert (MAGE) module that leverages modality priors to guide expert activation, fostering adaptive specialization in multimodal reasoning. Empirically, SmartMage achieves state-of-the-art performance across five 3D scene understanding benchmarks, and attains competitive results on RGB-only video understanding benchmarks. In our diagnostic benchmark ScanFacet, tasks are divided into fine-grained semantic categories, enabling analysis of modality combinations preferred by each semantic type. The observed modalitysemantic patterns provide further evidence of SmartMages effectiveness. Project page: \textcolorblue\urlhttps://yuecheong.github.io/SmartMage/.

Abstract:
Autonomous driving planning requires selecting trajectories that remain safe, rule-compliant, and comfortable under complex multi-agent interactions, yet robust closed-loop decision making in planning-critical scenes remains challenging. Existing diffusion- and RL-based planners often rely on generic rewards and use relational structure mainly for representation, limiting how planning-critical rollouts shape policy improvement. We present GRID-Planner, a graph-reward integrated diffusion planning framework that aligns generative planning with structured relational signals. Specifically, GRID-Planner combines graph-aware importance weighting, a time-indexed relational reward surrogate, and an advantage-weighted diffusion objective with on-policy self-distillation, forming a rollout-return-advantage-update loop without explicit likelihoods. Experiments on nuPlan and interPlan show that, compared with the diffusion-only baseline, GRID-Planner reduces collision rate, lane deviation, and jerk by 35.8%, 60.8%, and 27.4%, respectively, and improves the planning score by 3.4%. Project website: https://github.com/yangzhaojason/GRID-Planner.

Abstract:
Out-of-distribution (OOD) detection aims to identify samples from unseen classes and prevent unreliable predictions under open-world conditions. While recent studies explore visionlanguage models (VLMs) for OOD detection, most approaches rely on unimodal similarity or confidence scores, limiting their ability to exploit complementary multimodal information and handle uncertainty. In this paper, we propose an uncertainty-guided imagetext evidence fusion framework (UGE-Fusion) for VLM-based OOD detection. Specifically, visual and textual representations are transformed into evidence representations and integrated using DempsterShafer Theory (DST), enabling principled modeling of both modality-specific evidence and the associated uncertainty. By explicitly capturing inter-modal consistency and knowledge ambiguity, the proposed framework improves OOD detection through uncertainty-aware evidence fusion. Extensive experiments on multiple OOD benchmarks demonstrate that UGE-Fusion achieves state-of-the-art performance and generalizes well across different VLM architectures. In particular, on the large-scale ImageNet-1k OOD benchmark with CLIP, our method achieves 86.4% AUC and 49.4% FPR, outperforming existing methods. Moreover, the proposed uncertainty-aware fusion mechanism improves robustness under both challenging close-OOD scenarios and domain-shifted in-distribution (ID) data. Code is available at \urlhttps://github.com/BillaBNB/UGE-Fusion

Abstract:
Recent multimodal models have substantially improved scene understanding from heterogeneous inputs such as images, text, and audio. However, rich multimodal description alone is often insufficient for explicit and machine-interpretable scene representation. To address this gap, we introduce Multimodal Object-Attribute-Relation (OAR) Graph Construction, a new task that formulates image, text, and audio inputs describing the same scene as unified structured graph prediction. We propose a modality-aware framework that combines task-specific multimodal fusion, cascaded O-A-R prediction, and learned candidate-pair pruning for efficient graph inference. To facilitate systematic study of this task, we construct VG-Tri, a controlled tri-modal benchmark that extends scene graph annotations with aligned textual and acoustic descriptions, enabling rigorous analysis of how heterogeneous modalities contribute to structured graph prediction. Experiments on VG-Tri show that the proposed method consistently outperforms unimodal baselines, naive multimodal fusion strategies, and extraction-based alternatives, with leakage-controlled analyses confirming that the gains stem from genuine cross-modal semantic integration. The code is available at https://github.com/lcx529955/Multimodal-OAR.

Abstract:
Dual-Camera Compressive Hyperspectral Imaging (DCCHI) introduces an auxiliary panchromatic (PAN) observation to mitigate spatial information loss in compressed hyperspectral measurements. Despite recent advances in deep unfolding for DCCHI, two challenges remain: physically coupled PAN-guided reconstruction and effective modeling of spatial degradations caused by coded modulation and spectral dispersion. To address these issues, we propose PPU-Net, a Physically-coupled PAN-guided Unfolding Network. Derived from a joint optimization model, PPU-Net unfolds the reconstruction into three interpretable steps: PAN-consistency updating, hyperspectral-consistency updating, and learned refinement. It explicitly enforces PAN consistency through the spectral response operator and its adjoint, establishing principled physical coupling between the two observations. Furthermore, we propose a Dispersion-Guided Self-Attention (DGSA) tailored to capture both spatial structures and dispersion-induced artifacts. DGSA adopts a dual-branch design with shared query, key, and value features: a structural branch models local spatial structures with enlarged receptive fields, while a dispersion branch leverages stripe-shaped windows aligned with the dispersion direction to suppress ghosting artifacts. Experiments on both simulated and real datasets demonstrate that our method consistently outperforms state-of-the-art approaches in reconstruction quality and computational efficiency. Code is available at \urlhttps://github.com/M1zuk1/PPU-Net.git

Abstract:
Infrared-visible image fusion facilitates robust multimodal perception by integrating complementary textural nuances from visible sensors with thermal signatures from infrared systems. Due to the task's inherently ill-posed nature, existing methods heavily rely on structural priors but typically enforce rotation equivariance uniformly across all features. Such a holistic approach overlooks a critical distinction where low-frequency shared structures strictly adhere to equivariant constraints while high-frequency modality-specific details require greater flexibility to preserve unique information. To bridge this gap, we propose RoES, a Rotational Equivariant Selective-frequency fusion network. Instead of employing static decomposition, we introduce a trainable rotation-enhanced updater/predictor module to dynamically decouple low- and high-frequency components. The resulting representations are then processed through a dual-branch fusion module tailored for spectral consistency. Specifically, a rotation-equivariant Mamba is employed to capture long-range structural dependencies in the low-frequency domain, while a polar spectral attention-based Dual-Fourier block refines high-frequency details under explicit low-frequency guidance. Extensive experiments demonstrate that RoES consistently achieves state-of-the-art performance in both fusion quality and downstream object detection, establishing a robust solution for multimodal fusion by reconciling frequency-selective features with equivariant constraints. The source code is available at https://github.com/BryceLosky/RoES-Fusion.

Abstract:
In recent years, image-to-video (I2V) generation models have made remarkable progress in subject consistency and temporal coherence, enabling high quality video synthesis. However, these advances also introduce new safety risks. Existing studies mainly focus on jailbreak attacks involving single frame violations, while largely overlooking the temporal dimension unique to video generation models. In this paper, we investigate three attack scenarios and uncover a temporal vulnerability in I2V systems: unsafe semantics may emerge not from a single frame, but from semantic composition over time. We further identify two key challenges in such attacks: temporal abstraction and semantic camouflage. To address these issues, we propose TempJail, a novel temporal jailbreak framework for I2V systems. For temporal abstraction, we decompose a target malicious caption into an initial frame visual condition and a temporal text instruction. For semantic camouflage, on the image side we model semantic injection as controlled latent perturbation in diffusion sampling and introduce gradient guidance from pretrained encoders. On the text side, we rewrite the caption into an innocuous ``subject-action-scene'' template that bypasses safety filters while preserving temporal guidance. In the black-box inference phase, these two modalities jointly enable malicious semantics to be gradually triggered over time. Experiments on closed-source commercial models, including Kling, Seedance, Veo and PixVerse, show that TempJail improves attack success rate over prior state-of-the-art methods by 23.3% under GPT-5.2 evaluation and 22.0% under human evaluation. Our codes are available at \hrefhttps://github.com/luqi-glory/TempJailGitHub.

Abstract:
The growing demand for rapid and scalable 3D asset creation has driven interest in feed-forward 3D reconstruction methods, with 3D Gaussian Splatting (3DGS) emerging as an effective scene representation. While recent approaches have demonstrated pose-free reconstruction from unposed image collections, integrating stylization or appearance control into such pipelines remains underexplored. Existing attempts largely rely on image-based conditioning, which limits both controllability and flexibility. In this work, we introduce AnyStyle, a feed-forward 3D reconstruction and stylization framework that enables pose-free, zero-shot stylization through multimodal conditioning. Our method supports both textual and visual style inputs, allowing users to control the scene appearance using natural language descriptions or reference images. To the best of our knowledge, AnyStyle is the first feed-forward 3D stylization method that supports style control with a text prompt. We propose a modular stylization architecture that augments a pretrained state-of-the-art feed-forward 3D reconstruction backbone with only minimal modifications. Experiments demonstrate that AnyStyle improves the controllability and flexibility of style transfer over prior feed-forward stylization methods while preserving high-quality geometric reconstruction. A user study further confirms that AnyStyle achieves superior stylization quality compared to an existing state-of-the-art approach.

Abstract:
Long-tailed recognition suffers not only from data imbalance, but also from severe structural squeeze in the learned representation space, where tail samples are often embedded in densely populated head-class neighborhoods with weak local support. To diagnose this geometric fragility, we propose Local Manifold Sensitivity (LMS), a label-free, online metric that measures semantic discrepancy relative to local structural support, thereby identifying regions in which the current geometry is insufficient to sustain healthy decision boundaries. Building on this diagnosis, we further propose Manifold Expansion Regularization (MER), a plug-in regularizer that strengthens semantic correction in weakly supported regions by weighting a semantic divergence term with detached inverse structural support, thereby progressively expanding collapsed local manifolds. Experiments on CIFAR-LT, ImageNet-LT, and iNaturalist demonstrate that LMS consistently reveals the structural fragility of tail classes, while MER yields great improvements across diverse long-tailed recognition settings. Code is available at \urlhttps://github.com/PolarisLight/LMS-MER.

Abstract:
Text-driven image stylization has made substantial progress in digital content creation. Nevertheless, achieving fine-grained multi-attribute control still relies on large-scale stylized datasets, which are often isolated in privacy-sensitive silos. To overcome this limitation, we introduce federated text-to-image stylization, a new task where distributed clients collaboratively learn a text-driven stylization model without sharing private image data. However, achieving precise and stable multi-attribute stylization in this setting remains highly challenging. The first challenge is Semantic-Visual Conflict between clients, where the same style prompt may correspond to substantially different visual patterns across clients. The second challenge is Multi-Attribute Entanglement and Aggregation Interference, where entangled style attributes can easily cause destructive interference during federated aggregation and joint stylization. To address these challenges, we propose FedTIS, a federated text-to-image stylization framework built upon topology-guided multi-attribute composition. Specifically, FedTIS introduces a client-side Multi-Attribute Stylization Adapter to transform attribute-specific text prompts into style control residuals for fine-grained stylization. On the server side, it builds a Tri-level Style Tree to hierarchically organize client style knowledge and mitigate cross-client semantic-visual conflicts, while a Multi-Attribute Compatibility Graph explicitly models cross-attribute compatibility for stable joint stylization. Extensive experiments show that FedTIS consistently outperforms representative baselines in style consistency, controllability, and personalization. Code is available at https://github.com/FanQi-AI/FedTIS.

Abstract:
Breast ultrasound diagnosis relies on clinically meaningful semantic concepts, yet most deep learning methods adopt end-to-end image-to-label paradigms that lack interpretability and robustness. While concept-based approaches offer a promising alternative, they often assume complete annotations or require multimodal inputs at inference, which significantly limits their real-world applicability. To tackle these issues, we propose Training-time Report-guided and Clinically Ordered Concept Editing (TRACE), a training-time report-guided framework that leverages structured radiology reports as privileged concept supervision while enabling image-only diagnosis at test time. TRACE refines image-derived concepts through a teacher-guided editing mechanism within a malignancy-aware ordered concept space. To address incomplete annotations, we introduce Strategic Concept Missing Training (SCMT) and train an image-only self-editor via edit distillation for autonomous concept refinement. Besides, we introduce BUSC, a concept-enriched benchmark linking images, labels, and structured attributes. Experiments across multiple datasets demonstrate that TRACE achieves superior performance and improved cross-domain robustness compared to existing methods. Code is available in our GitHub repository: \urlhttps://github.com/wentao-2/TRACE

Abstract:
Face-swap detectors often generalize poorly because they rely on renderer-specific artifacts that vary across synthesis pipelines and acquisition conditions. We instead target identity-replacement forgeries, where donor identity information is combined with the recipient's facial context. We introduce DIMF, a source-aware training and source-free inference framework that uses paired source referencesidentity donors for identity-replacement samplesas privileged training supervision. A shared ConvNeXt-V2 backbone is fine-tuned through a training-only Cross-Identity Mamba branch and Identity Gap Supervision, while a frozen diffusion pathway provides auxiliary regularization. At inference, all source-dependent and diffusion components are removed, leaving an 88 M-parameter single-image detector. Trained only on FaceForensics++ and evaluated without target-domain fine-tuning on five unseen benchmarks, DIMF achieves an average AUC of 84.25, exceeding the strongest protocol-aligned diffusion-guided baseline by 2.37 points, including a 7.23-point gain on DiffSwap. Donor-assignment and matched-backbone controls support structured sourcetarget correspondence as a contributor to the observed gains. Code: https://github.com/aiai-9/DIMF.

Abstract:
Human action-reaction synthesis, a fundamental challenge in modeling online causal human interactions, plays a critical role in applications ranging from virtual reality to social robotics. While diffusion-based models have demonstrated promising performance, they exhibit two key limitations for interaction synthesis: reliance on complex noise-to-reaction generators with intricate conditional mechanisms, thus limiting to unidirectional generation, and frequent body penetrations in generated motions. To address these issues, we propose Action-Reaction Flow Matching (\methodname), a novel paradigm that establishes direct action-to-reaction mappings, eliminating the need for complex conditional mechanisms and supporting bi-directional generation. Directly applying traditional guidance algorithms tends to undermine the quality of generated reaction motion. We analyze the sampling of flow matching in depth and reveal an issue (Initial Point Deviation) which causes the sampling trajectory to ever farther from the initial action motion. Thus, we propose a reprojection guidance method, R\scalebox0.8E-G\scalebox0.8UID, to correct this deviation to enable better interaction. To further enhance the reaction diversity, we incorporate randomness into the sampling process. Extensive experiments on NTU120, Chi3D and InterHuman datasets demonstrate that \methodname~not only outperforms existing methods in terms of FID and motion diversity but also significantly reduces body collisions, as measured by our introduced Intersection Volume and Intersection Frequency metrics. Project Page: \urlhttps://arflow2025.github.io/

Abstract:
Social deduction games (SDGs) such as Werewolf have become challenging testbeds for AI agents. These games require complex social skills such as reasoning, deception, and collaboration. While recent advances in large language models (LLMs) have driven significant progress in SDG agents, current approaches are predominantly text-based, overlooking the multimodal nature that is fundamental to human social interaction. To bridge this gap, we introduce CaM-Wolf, the first SDG agent that integrates multimodal perception and generation. CaM-Wolf processes video inputs from other players, employs a causal-aware Reasoner trained via reinforcement learning to establish logical chains between observable behaviors and hidden roles, and presents itself through an animated avatar. Our experiments and user study show that CaM-Wolf achieves superior agent gameplay performance and enhances the quality of human-AI interaction. This work represents a significant advancement towards creating more human-like AI agents capable of participating in nuanced social dynamics. Our code is available at https://3dagentworld.github.io/avatar_wolf.

Abstract:
Video misinformation on social media threatens information credibility, yet its diversity makes automated fact-checking challenging. Multimodal Large Language Models (MLLMs) and Multi-Agent Systems (MAS) enable collaborative reasoning with explainable decisions. However, existing methods often rely on trial-and-error reasoning without reusing prior experience, leading to redundant and suboptimal reasoning. Moreover, processing entire videos introduces irrelevant information that can obscure key evidence. To address these challenges, we propose a framework for VIdeo fact-checking via experience-Guided multi-agent cOllaborative Reasoning (VIGOR). Specifically, we introduce a Video Fact-Checking Experience Memory (VFEM) with a non-parametric reinforcement learning mechanism inspired by value iteration, allowing agents to retrieve past experiences to reduce reasoning costs and improve performance. We represent videos as a Multimodal Evidence Memory (MEM) of temporal segments and use cross-modal retrieval to adaptively isolate key evidence and reduce interference. Extensive experiments show that VIGOR improves veracity classification by 7.16% and explanation quality by 5.72% over existing methods. Our code is available at https://github.com/immc-lab/VIGOR.

Abstract:
Self-generated skills for web agents are often unstable and can even hurt performance relative to direct acting. The key bottleneck is not only skill generation quality, but the fact that web skills remain implicit and therefore cannot be checked or locally repaired. To address this, we present ContractSkill, a framework that converts a draft skill into an executable artifact with explicit procedural structure, enabling deterministic verification, fault localization, and minimal local repair. This turns skill refinement from full rewriting into localized editing of a single skill artifact. Experiments on VisualWebArena show that ContractSkill is effective in realistic web environments, while MiniWoB provides a controlled test of the mechanism behind the gain. Under matched transfer layers, repaired artifacts remain reusable after removing the source model from the loop, demonstrating cross-model artifact reuse within web-agent benchmarks. These results show that the central challenge is not merely generating skills, but making them explicit, executable, and repairable.

Abstract:
Temporal grounding in egocentric videos, known as Natural Language Query (NLQ), aims to localize query-relevant moments in long first-person videos. To facilitate multimodal alignment, existing methods typically learn video representations via vision-language pretraining or contrastive learning. However, these approaches primarily focus on capturing and modeling full-body poses and environmental context, while overlooking a crucial aspect: egocentric activities are fundamentally human-driven (\ie, human intent and action). In light of this, we revisit NLQ from the human-centric representation learning perspective and propose HuRe-NLQ, a framework that jointly models human attention and human actions. We learn gaze and hand-object-interaction related cues to highlight intention-relevant visual patterns, enabling the model to describe what the actor truly attends to. Concurrently, to localize fine-grained moments under the frequently changing scene, we build a soft positive anchor annotation mechanism for better query alignment with sub-second action. Extensive experiments on Ego4D-NLQ and Ego-Exo4D benchmarks demonstrate the superiority of HuRe-NLQ over state-of-the-art methods. The code is available at \urlhttps://github.com/Sunburst7/HuRe-NLQ.

Abstract:
Multi-track music generation requires token representations that are not only reconstructable, but also easy to model and structurally aligned across tracks. In the dual-track vocalaccompaniment setting, this requirement creates a fundamental tension: reconstruction-oriented codecs preserve acoustic detail but are difficult for sequence modeling, while more semantic tokenizers are often more predictable yet may lose fidelity or cross-track structure. We present DuoTok, a source-aware dual-track music tokenizer designed as a multi-track generation interface through staged disentanglement. DuoTok first learns a semantic audio representation via self-supervised pretraining, then shapes source-aware structure in continuous space with feature replacement noise and multi-task supervision, including spectral reconstruction, music source separation regularization, and an ASR head for lyric alignment. It then freezes the encoder and learns hard-routed dual codebooks for vocals and accompaniment, while a diffusion decoder restores fine acoustic detail from the discrete tokens. Across public benchmarks, DuoTok achieves a better predictabilityfidelity trade-off than prior tokenizers at ultra-low bitrate. Under held-constant dual-track language modeling, it further improves both unconditional vocalaccompaniment modeling and vocal-conditioned accompaniment prediction, indicating that its token space provides a stronger interface for dual-track sequence modeling. Controlled diagnostics show larger predictability costs under cross-track corruption and larger gains from longer temporal context, supporting the interpretation that these gains arise from stronger structural usage rather than merely easier local prediction. At the same time, DuoTok maintains competitive reconstruction quality and preserves control-relevant musical attributes in the discrete space. Together, these results suggest that, for multi-track music generation, tokenizer design should be treated as a core modeling problem rather than only a compression problem.

Abstract:
Multi-reference image generation aims to synthesize images by integrating attributes from multiple reference images under textual instructions. As the number of references increases, the task necessitates complex semantic comprehension, such as correctly associating attributes with the intended subjects and planing out coherent spatial arrangement between subjects and their environments. Existing approaches, which rely solely on natural language instruction, often fail to capture these complex intentions precisely, leading to semantic misalignment and inconsistent generation. We identify two key factors behind these limitations: natural language instructions are often verbose and ambiguous, and high-quality multi-reference data is scarce. To address these issues, we propose StructGen, which employs a structured, dictionary-like format to encode multiple reference images, thereby enabling explicit and unambiguous specification of generation intentions. To support this design, we construct a structured dataset based on high-quality real images and develop a corresponding training framework, along with a dedicated benchmark for challenging multi-reference scenarios. Extensive experiments on both public benchmarks and our proposed benchmark demonstrate that StructGen consistently outperforms existing methods on both semantic alignment and detailed reference-generation consistency, especially under complex instructions with multiple references.

Abstract:
Test-Time Adaptation (TTA), particularly the training-free paradigm, has garnered widespread attention due to its significant advantages in addressing test-time distribution shifts in Vision-Language Models. However, existing training-free approaches are often constrained by the inherent limitations caused by an over-reliance on a single textual or visual modality, resulting in suboptimal performance and robustness. To address this issue, we propose a novel training-free dynamic multimodal synergy strategy. First, to extract textual modality features with enhanced robustness, we construct a cache-based, hierarchical retrieval-augmented TTA paradigm. Building upon this, we introduce a negative-entropy-driven dynamic synergy mechanism that adaptively fuses the optimized textual modality, the visual modality, and the original CLIP modality. This integration ultimately yields a predictive distribution characterized by both high robustness and high confidence. Our approach significantly enhances class discriminability while maintaining exceptional robustness during testing. Extensive experiments across multiple benchmarks demonstrate that the proposed method consistently outperforms state-of-the-art approaches in both performance and efficiency. Code is available at \urlhttps://github.com/Alian2014/MM26-MMS-TTA

Abstract:
Cross-embodiment navigation is a key challenge in embodied intelligence. Due to differences in embodiment, the same visual observation may imply different actions for different agents, making prediction ambiguous when relying solely on vision. Existing studies mainly rely on reinforcement learning, which requires large-scale interaction and careful reward design, making it difficult to support scalable pretraining and real-world adaptation. In contrast, imitation-learning-based approaches remain limited. To address these challenges, we propose an imitation-learning-based embodiment-aware navigation framework with a modular multi-stage design. In pretraining, we construct a cross-embodiment navigation dataset from Internet videos and introduce embodiment geometry as conditional tokens to reduce action ambiguity under the same observation. In fine-tuning, we design a multimodal information injection mechanism based on a decoupled architecture. Specifically, we design a trajectory augmentation strategy to generate high-risk samples, which are used to train spatial perception and risk-aware correction separately, thereby explicitly incorporating embodiment geometry for safe navigation. Experimental results show that the proposed method effectively improves navigation performance across different embodiment settings, demonstrating the effectiveness of incorporating embodiment geometry into embodied navigation. Project page: \urlhttps://ea-nav.github.io/.

Abstract:
Single-image human material estimation aims to recover physically based rendering (PBR) attributes of a human from a single image, enabling realistic relighting for immersive multimedia applications. Although recent methods have achieved promising results on synthetic benchmarks, their predictions on real images may suffer from residual illumination artifacts, which lead to unrealistic relighting results. We argue that this limitation mainly arises from the domain gap between synthetic and real data in both material appearance and illumination, which aggravates the ambiguity between illumination and reflectance during inference. To address this issue, we propose S2R, a novel two-stage synthetic-to-real framework for single-image human material estimation. In the first stage, we train the model on high-quality synthetic data with supervision under multiple constraints and introduce an illumination-aware feature injection module to improve the disentanglement of material properties from shading. In the second stage, we selectively fine-tune the model on real images to transfer the knowledge learned from synthetic supervision to the real domain. This adaptation process is guided by self-supervised reconstruction and teacher-student regularization, which together reduce the synthetic-to-real domain gap while preserving stable material estimation. Extensive experiments on both synthetic and real-image benchmarks demonstrate that our method produces cleaner material estimates with less residual illumination and achieves more realistic relighting results than existing approaches. Project materials are available at https://jiangyu1181.github.io/S2R/.

Abstract:
Underwater vision serves as a critical role in exploration and understanding of marine environments. However, the water body induces light absorption and scattering, resulting in degraded image quality and impaired downstream vision tasks. Existing methods primarily focus on global feature modeling or pixel-wise reconstruction and overlook structural information related to spatial distribution and object boundaries, leading to limited performance of downstream vision tasks. To address this issue, we propose a Structure-Aware Transformer Network (SATNet) to bridge the gap between low-level pixel representations and high-level semantic representations by enhancing structural information in underwater images. Specifically, a Structure-guided Cascaded Attention (SCA) mechanism is proposed to collaboratively optimize pixel-wise spatial features and structure-wise semantic representations, enabling long-range dependency modeling while enhancing attention to informative regions. In addition, to better restore details during the image reconstruction procedure, a Structure-guided Dynamic Upsampling (SDU) strategy is proposed to impose structural constraints, thereby reducing blurring and artifacts. Extensive experiments demonstrate the proposed model achieves the best performance on multiple public datasets in quantitative metrics. Compared with the state-of-the-art baseline, it yields a +0.65 dB improvement in PSNR while reducing computational complexity and parameter count by 65.5% and 91.1%, respectively. Furthermore, the proposed model achieves superior visual quality with enhanced detail restoration and more consistent preservation of structural information. The code is available at https://github.com/xsml1/SATNet.

Abstract:
Embodied intelligent ultrasound scanning enables the automation and standardization of the ultrasound examination process by integrating perception, decision-making, and execution capabilities. However, existing methods suffer from loosely coupled modeling between force and ultrasound modalities and lack awareness of scanning stages, which limits their ability to capture dynamic probetissue interactions. To address these issues, we propose ForceU-VLA, a force-aware VisionLanguageAction model for autonomous embodied ultrasound scanning, which leverages force signals and ultrasound image feedback throughout the scanning process to enable accurate and high-quality ultrasound acquisition. Firstly, we propose a Force-Ultrasound Synergistic Fusion Module (FUSFM) that synergistically fuses ultrasound visual and force-feedback information to provide stable, reliable guidance for probe motion. Secondly, a Stage-Adaptive Modulation Mechanism (SAMM) is proposed to accommodate the task requirements across different scanning stages by adaptively modulating multimodal features to enhance their representation quality. Additionally, we introduce ForceU-VLA-Data, a real-world, force-aware embodied ultrasound dataset that integrates visual, force, and action signals, including data from two organs across five representative clinical scanning views, and comprising 450 expert-collected trajectories with approximately 100,000 synchronized multimodal frames. Extensive experimental results demonstrate that ForceU-VLA significantly improves contact stability and probe pressure regulation in embodied ultrasound scanning, thereby effectively enhancing task execution quality and overall system reliability. The source code is available at https://github.com/VMVLab/ForceU-VLA.

Abstract:
Transferable adversarial attacks pose significant security threats to neural networks due to their strong black-box effectiveness against real-world models. While neuron attribution-based attacks have shown promise in enhancing transferability, existing methods are often constrained by integration paths that accumulate noise or lack structural diversity. Moreover, they typically rely on single-layer neuron attribution, failing to exploit multi-layer information. To address these limitations, we propose COunterfactual-Guided Diversified Path Integration Neuron Attribution Attack, termed COPIA. Specifically, we incorporate counterfactual-guided paths into the neuron attribution framework, to effectively mitigate noise accumulation in integration trajectories and enable model-adaptive construction. Building on this, we diversify path update mechanisms and counterfactual samples to enhance path diversity and improve adversarial transferability. Furthermore, to capture the complementary semantics across different layers, we design a Cross-layer Attribution (CA) mechanism that progressively aggregates attribution information from deeper layers into the target attack layer via attribution flow. Extensive experiments on the ImageNet dataset across normally trained models, adversarially trained CNNs, ViTs, and various defense models demonstrate that COPIA consistently outperforms state-of-the-art methods in black-box attack settings. In addition, the proposed CA mechanism is highly generalizable and effectively improves the black-box transferability of existing single-layer neuron attribution attacks. Code and supplementary materials are available at https://github.com/AdvML-Group/COPIA.

Abstract:
While 3D Gaussian Splatting (3DGS) has advanced open vocabulary scene understanding, existing methods remain confined to explicit queries. They struggle to interpret implicit intents, complex spatial constraints, and commonsense reasoning required for practical embodied interactions. To address this gap, we introduce the task of reasoning 3D Gaussian segmentation and construct two benchmarks, Causal-LERF and Causal-ScanNet. These benchmarks systematically evaluate commonsense, spatial, affordance, and counterfactual reasoning. Evaluations reveal that current state of the art methods perform poorly on these reasoning challenges. Therefore, we propose CausalSplat, a framework that integrates vision-language models with 3D scene graphs to disentangle explicit structural perception from implicit logical inference. Extensive experiments demonstrate that CausalSplat achieves state of the art performance on our reasoning benchmarks while showing strong generalizability on standard referring and open vocabulary 3D segmentation tasks. Project Page: https://jiayuding031020.github.io/CausalSplat

Abstract:
3D Gaussian Splatting (3DGS) has shown strong performance in novel view synthesis (NVS). However, sparse-view aerial NVS is further complicated by regions that fall entirely outside all training frustums. These regions receive neither geometric nor photometric supervision, leaving them fundamentally under-constrained. Existing methods either address only observed regions or rely on ground-level priors that transfer poorly to aerial NVS. To tackle this challenge, we propose Structure-Spreading Gaussian Splatting (S^2-GS), which extends geometry and appearance from observed to unobserved aerial regions. Consistent Initialization with Geometric Anchoring enforces cross-view metric consistency among depth estimates via dual-frequency residual refinement, producing a dense and reliable point cloud covering observed regions. Structure Spreading via Observed Geometry extends this initialization into unobserved regions. It extrapolates depth from observed regions and back-projects the resulting values into 3D space. Reliability-Aware Progressive Optimization refines the unobserved Gaussian appearance. The first stage applies pseudo-view supervision across all unobserved regions to establish a coherent reconstruction foundation. The second stage freezes geometry and switches to reliability-filtered supervision to suppress artifacts. Experiments on LEVIR-NVS and 3D-AS show that S^2-GS surpasses all comparative methods. The code is publicly available at \hrefhttps://github.com/TangXu-Group/S2-GShttps://github.com/TangXu-Group/S2-GS.

Abstract:
Human-Object Interaction (HOI) detection aims to localize humans and objects in images and classify their interactions. Zero-shot HOI focuses on recognizing interactions that are not observed during training, requiring models to generalize beyond seen verb-object compositions. Recent approaches leverage Vision-Language Models (VLMs), benefiting from rich semantic representations. However, they often rely on global or detector-centric features that compress interaction cues and hinder fine-grained spatial reasoning. To overcome this limitation, we propose \method, a zero-shot HOI framework that explicitly models interactions from dense patch-level features. Our approach is motivated by the observation that human-object interactions are defined by localized spatial relationships, which are not preserved by global and detector-centric representations. To this end, we introduce an \adaptivepatchmodule module that decomposes humans and objects into semantically coherent parts via unsupervised clustering, and re-weights them based on their interaction relevance. These representations are then integrated through a \regiontransformermodule that integrates part-aware and global features and produces the final HOI embedding. Extensive experiments demonstrate that \method consistently outperforms existing zero-shot HOI methods, achieving particularly strong gains on unseen interactions. Code is available \hrefhttps://github.com/francescotonini/fine-hoihere.

Abstract:
Human personality analysis is crucial for personalized dialogue systems, mental health assessment, and humancomputer interaction, but existing resources are mostly text-only and fail to capture rich cognitive cues (e.g., speaker-awareness and emotions) in real conversations. We present a new benchmark for human personality analysis in multimodal dialogues (HPAMD), built on 218 real-world IELTS-style interview videos (RIIV) and segmented into 2,322 turn-level units with aligned emotion-related dialogue text, acoustic, visual, and continuous Big Five annotations. We further propose DAMP, a dialogue-aware multimodal framework that models turn-level question-response interactions and integrates emotion reasoning for personality prediction. Experiments show dialogue modeling outperforms whole-video modeling, emotion-enhanced representations bring additional gains, and DAMP achieves the best overall performance under the full setting. The dataset is publicly available at: https://github.com/lulvhui/RIIV.

Abstract:
Recent advances in AI have revolutionized speech processing, yet effective speech understanding requires discerning not just what is said, but how it is said. Speech Sentiment Analysis plays a critical role in decoding these paralinguistic cues for diverse real-world applications such as recruitment and customer service. However, existing Speech Sentiment Analysis research faces two primary limitations. First, dominant approaches rely on text-centric pipelines that cascade Automatic Speech Recognition with text analysis. This process inevitably discards essential acoustic features like prosody and tone, failing to capture attitudinal meanings in acoustically ambiguous utterances. Second, current benchmarks suffer from a mismatch in label granularity, prioritizing basic emotions (e.g., happy, sad) over the nuanced interpersonal stances (e.g., confident, impatient) necessary for social sensitivity. To address these limitations, we propose a novel dataset, SpeechSense, for fine-grained speech sentiment analysis. Specifically, we define a specialized 8-class taxonomy of interpersonal stances detectable primarily through prosodic cues beyond lexical content alone. We then construct a curated dataset based on this taxonomy, built from high-fidelity speech synthesis and rigorous human validation. Comprehensive experiments across multi-modal LLMs, text-only LLMs, and speech encoders demonstrate that models with acoustic access consistently outperform text-only baselines. These results empirically validate the primacy of acoustic cues in detecting subtle speaker attitudes, highlighting the necessity of SpeechSense.\footnoteDataset and supplementary materials: \urlhttps://github.com/Sher13cked/SpeechSense

Abstract:
The rapid growth of video data has created an urgent need for efficient video understanding systems that can handle long and continuous visual inputs, posing significant challenges for Multimodal Large Language Models (MLLMs) due to severe memory bottlenecks. Moreover, existing compression methods often overlook the explosive accumulation of approximation errors. To address this issue, we propose Video-MemComp, a training-free and sensitivity-aware context management framework for memory-constrained streaming video understanding. Our method theoretically decomposes compression error into orthogonal local approximation error and layer-wise structural sensitivity via a first-order Taylor expansion. To minimize local approximation error, we perform token aggregation in a pre-RoPE safe zone with an attention-conserving logit bias, mitigating geometric distortions caused by RoPE rotations. To manage layer-wise structural sensitivity, we dynamically allocate the KV cache budget based on offline gradient priors and online context entropy, prioritizing high-sensitivity layers. Video-MemComp outperforms or achieves performance comparable to full-cache and state-of-the-art compression methods on benchmarks including MLVU, VideoMME, and EgoSchema. Notably, it improves the overall accuracy of Qwen2.5-VL-3B from 62.3% to 64.1% while reducing the KV cache to an equivalent of 6,000 tokens per layer, saving over 90% of KV memory. Code: \urlhttps://github.com/hhnk666/VideoMemComp.

Abstract:
Existing token compression methods for omnimodal large language models typically rely on one modality to determine what to retain in the other. We show that this assumption often breaks down: for the same query, audio and video relevance often peaks at different moments. This cross-modal salience mismatch makes unidirectional guidance prone to discarding answer-critical cues under aggressive compression. We propose OmniScope, a training-free token compression framework that uses the query as a shared semantic anchor while estimating relevance separately for audio and video. OmniScope allocates modality-specific token budgets, prunes visual tokens with an anchor-delta strategy that preserves both global context and temporal changes, and merges audio tokens within each second to reduce redundancy while maintaining temporal continuity. Across four audio-video benchmarks and two Qwen2.5-Omni model scales, OmniScope achieves the best average accuracy across all compression settings. At 25% overall token retention, it delivers up to 3.53x prefill speedup and more than 15% GPU memory reduction, with only a 0.35-point drop in average accuracy. These results suggest a simple design principle for OmniLLM inference: share the query across modalities, but not the salience estimates.

Abstract:
Existing music-driven dance generation approaches have achieved strong realism and effective audio-motion alignment. However, they generally lack semantic controllability, making it difficult to guide specific movements through natural language descriptions. This limitation primarily stems from the absence of large-scale datasets that jointly align music, text, and motion for supervised learning of text-conditioned control. To address this challenge, we propose TeMuDance, a framework that enables text-based control for music-conditioned dance generation without requiring any manually annotated music--text--motion triplet dataset. TeMuDance introduces a motion-centred bridging paradigm that leverages motion as a shared semantic anchor to align disjoint musicdance and textmotion datasets within a unified embedding space, enabling cross-modal retrieval of missing modalities for end-to-end training. A lightweight text control branch is then trained on top of a frozen music-to-dance diffusion backbone, preserving rhythmic fidelity while enabling fine-grained semantic guidance. To further suppress noise inherent in the retrieved supervision, we design a dual-stream fine-tuning strategy with confidence-based filtering. We also propose a novel task-aligned metric that quantifies whether textual prompts induce the intended kinematic attributes under music conditioning. Extensive experiments demonstrate that TeMuDance achieves competitive dance quality while substantially improving text-conditioned control over existing methods. Code and qualitative results are available at: https://github.com/surrey-nlp/TeMuDance.

Abstract:
Graph-based models have shown strong potential for modeling patient health trajectories from multimodal electronic health records (EHRs). However, existing methods still struggle to capture the longitudinal dynamics of patient health states and often overlook subtle yet clinically critical risk signals, limiting the accurate characterization of health evolution. To address these limitations, we propose HyperCare, a longitudinal heterogeneous hypergraph neural network for modeling patient health trajectories from multimodal EHRs. HyperCare represents each patient trajectory as a longitudinal heterogeneous hypergraph, where heterogeneous clinical events serve as nodes, and time- and type-aware hyperedges jointly capture intra-visit interactions and long-term cross-visit dependencies, yielding a unified representation of evolving patient health states. We further enhance the hypergraph representation with multimodal semantics derived from large language models (LLMs) and leverage population-informed progression patterns to provide contextual information and directional priors. Based on this representation, HyperCare employs a guided hypergraph message passing mechanism to preserve clinically meaningful directional dependencies and capture longitudinal health dynamics. Experiments on two representative healthcare prediction tasks over MIMIC-III and MIMIC-IV demonstrate that HyperCare consistently outperforms state-of-the-art methods and exhibits strong few-shot generalization. Notably, for diagnosis prediction, it improves the F1 score over the best baseline by 10.5% on MIMIC-III and 9.1% on MIMIC-IV. The code is available at \urlhttps://github.com/low86/HyperCare.

Abstract:
Audio-driven video generation (A2V) has achieved promising progress in synthesizing temporally coherent and audio-visually aligned videos, yet its inference remains expensive due to the iterative denoising process of diffusion models. Existing caching methods mainly exploit temporal redundancy in visual features, while overlooking the cross-modal alignment of A2V, where audio drives visual generation with highly non-uniform temporal importance. In this paper, we identify two levels of misalignments in existing A2V caching methods: temporal-semantic and computation-storage. To address them, we propose EchoCache, an energy-guided cross-modal caching framework for efficient A2V generation. EchoCache leverages audio time-frequency energy as a saliency anchor to guide latent-level cache updates, and further introduces a dynamic timestep-latent caching mechanism with quantized cache management for joint efficiency and memory optimization. Extensive experiments on mainstream A2V models show that EchoCache consistently improves the latency-quality trade-off while preserving generation quality and audio-visual consistency. In particular, on Wan2.2-S2V over the EMTD benchmark, EchoCache achieves 2.46× speedup with the best overall performance.

Abstract:
Recent work has demonstrated that sparse memory fine-tuning offers a principled solution to catastrophic forgetting in continual learning by decoupling task-specific knowledge through selective slot-level updates. However, a critical pre-training bottleneck limits practical adoption. The memory layer replaces an internal feed-forward network of the Transformer, and the resulting architecture fundamentally departs from any standard pre-trained model, necessitating expensive re-pre-training from scratch. To overcome this bottleneck, we propose Sparse Retrieval Memory, an external sparse memory architecture that can be directly deployed on frozen pre-trained Vision Transformers without architectural surgery or re-pre-training. Concretely, we attach learnable key-value memory modules as parallel residual branches alongside the multi-head self-attention at specific intermediate layers. Through Top-K sparse routing, each input token dynamically retrieves a small subset of the most relevant memory slots, significantly mitigating gradient interference via sparse parameter activation. Furthermore, the memory pool grows via coverage-based dynamic expansion, automatically adapting its capacity to the complexity of the incoming data stream. Extensive experiments on ImageNet-R, CIFAR-100, and DomainNet demonstrate state-of-the-art performance, achieving 80.50% on 10-task ImageNet-R and 84.68% on 10-task DomainNet, while training fewer than 3% of backbone parameters and storing zero historical exemplars. The source code is available at https://github.com/xyz-npu/SRM-CIL.

Abstract:
Recent generative video codecs (GVCs) have achieved impressive reconstruction fidelity at ultra-low bitrates (< 0.02 bits per pixel) by compressing the tokens from generative tokenizers. However, existing GVCs generally require considerable computation time and model complexity, which hinder their deployment on compute-limited devices and in real-time applications. To bridge this gap, we systematically identify the computational bottlenecks and propose GVC-RT, which redesigns the generative latent coding framework to realize real-time video coding without sacrificing compression performance. Specifically, built on a pretrained lookup-free quantization (LFQ) tokenizer, GVC-RT adopts an asymmetric architecture that directly learns to match the LFQ latent distribution, while generative-space alignment is enforced via a regularization loss term only during training. In this manner, we bypass heavy tokenization and entirely remove the complex feature-alignment process at inference time. Moreover, we further introduce a lightweight de-tokenizer architecture to resolve the final latency bottleneck during decoding. Experimental results demonstrate that GVC-RT outperforms the previous SOTA model, GLC-Video, with average BD-rate savings of 12.4% and 48.8% in terms of DISTS and LPIPS, while achieving encoding/decoding speeds of 123.1/55.1 fps for 1080p video. The code is at \urlhttps://github.com/semcomm/GVC-RT.

Abstract:
Automatic white balance (AWB) correction in multi-illuminant scenes remains a challenging yet practically important problem. Recent methods typically rely on complex architectural designs, such as attention-based illumination decomposition or learned preset fusion, to model spatially varying illumination, often at the expense of simplicity, interpretability, and computational efficiency. In this paper, we propose a Chromaticity-Weighted mixing Network (CWNet) as a simple yet effective baseline for multi-illuminant AWB. In our design, we reformulate AWB as the joint estimation of global chromaticities and their corresponding spatially varying weight maps within a unified U-Net architecture. Specifically, global chromaticities are inferred from bottleneck features to capture scene-level illumination characteristics, while spatial weight maps are generated by the decoder to model local illumination variations. These two components are subsequently fused to construct the final illumination map. In this design, CWNet requires neither iterative optimization nor computationally intensive attention. To mitigate the scarcity of real multi-illuminant training data, we further introduce a physics-based rendering pipeline that generates large-scale, photorealistic sRGBground-truth image pairs under controllable mixed illumination. Extensive experiments demonstrate that CWNet consistently improves color fidelity over existing methods, while achieving competitive or state-of-the-art (SOTA) performance across multiple benchmarks with lower computational complexity. Our code will be available at: https://github.com/wzhsysu/CWNet.

Abstract:
Vision-language models (VLMs) such as CLIP exhibit remarkable zero-shot capabilities, yet their performance frequently degrades sharply under unexpected test-time distribution shifts. While Test-Time Adaptation (TTA) offers a promising solution, continuously adapting VLMs over an unlabeled test stream presents fundamental challenges. Conventional top-1-centric updates often reinforce errors by corrupting the local semantic geometry among related classes, while iterative adaptation exacerbates progressive bias accumulation, ultimately driving the model toward mode collapse. To overcome these coupled vulnerabilities, we propose Local Margin Restoration (LMR), a lightweight, one-step TTA framework. At the sample level, our Protected Margin Restoration (PMR) objective recovers local semantic geometry by shielding plausible near-top candidates from external hard negatives. Concurrently, to combat stream-level degradation, we introduce a dual-stage stabilization mechanism, featuring an Adaptive Margin (AM) controller and Bias Correction (BC), to dynamically disrupt progressive bias accumulation and prevent mode collapse. Extensive experiments on CIFAR-C, ImageNet-C, and ImageNet variants demonstrate that LMR consistently outperforms state-of-the-art TTA baselines, proving exceptionally robust and efficient even in challenging low-batch test-time regimes. Our code is available at \urlhttps://github.com/DennisHuangYan/LMR.

Abstract:
3D Gaussian Splatting (3DGS) has emerged as a highly efficient representation for novel view synthesis. However, extracting accurate surfaces from 3DGS remains challenging due to unreliable geometric supervision. Existing methods heavily rely on depth-based reprojection for visibility estimation and multi-view consistency. This creates a fundamental circular dependency: precise visibility estimation requires accurate depth, yet depth supervision itself is conditioned on visibility. To break this cycle, we rethink multi-view geometric supervision through the lens of visibility, modeling it directly at the Gaussian level. Specifically, we propose a Gaussian Visibility-aware Multi-View (GVMV) geometric consistency formulation. By aggregating the cross-view visibility of shared Gaussians, GVMV enables robust supervision specifically over co-visible regions. Furthermore, to effectively integrate monocular depth priors, we introduce a progressive Quadtree-calibrated Depth Constraint (QDC). Guided by our visibility formulation, QDC performs block-wise affine calibration to mitigate scale ambiguity while strictly preserving local geometric structures. Extensive experiments on the DTU and Tanks and Temples benchmarks demonstrate that our approach consistently outperforms prior Gaussian-based methods in surface reconstruction accuracy. Our code is fully open-sourced and available at \urlhttps://github.com/GVGScode/GVGS.

Abstract:
Vision foundation models are increasingly reused as frozen backbones for downstream visual recognition, making parameter-efficient adaptation a central problem. Prompt-based adaptation, including Visual Prompt Tuning (VPT), provides a lightweight way to specialize these models, but its layer-wise behavior remains poorly understood: performance is sensitive to prompt depth, placement, and task distribution, and gains on standard in-domain benchmarks do not always translate into robust generalization. We argue that this limitation is not solely an optimization issue, but a layer-wise information allocation issue: existing prompt-based methods lack principled control over what prompt-conditioned representations should preserve, suppress, and propagate across depth. Inspired by the Information Bottleneck principle, we introduce Prompted Information Bottlenecks (PIB), a framework that regularizes layer-wise compression-sufficiency trade-offs and promotes a more coherent cross-layer information path. The key idea is that effective adaptation should be minimal yet sufficient, retaining task-relevant local evidence in earlier layers while progressively discarding nuisance factors and redundant details in deeper layers. Extensive experiments show that PIB achieves strong performance across 34 datasets, reaching 92.1% on FGVC, 93.01% on HTA, and 77.33% on VTAB-1k, while tuning only 0.35% parameters on average across the main settings. Beyond benchmark accuracy, PIB helps explain the non-monotonic behavior of prompt capacity scaling, reduces shortcut reliance, and improves robustness under distribution shift and fine-grained recognition settings. These results position PIB as both a practical method and an information-allocation perspective for adapting frozen vision foundation models.

Abstract:
Recent advances in Vision-Language Models (VLMs) have significantly improved image geo-localization, yet existing models remain susceptible to landmark bias, causing them to overlook geographical cues or form spurious correlations, ultimately resulting in inaccurate localization. To systematically investigate this issue, we first design two quantitative metrics, Bias Intensity (BI) and Bias Harmfulness (BH), to characterize the impact of landmarks exerted on model reasoning, and establish a comprehensive benchmark, LandmarkBias-3K. To mitigate landmark bias, we further propose an evidence-driven reasoning framework, HoloGeo, to improve the reliability of geo-localization. HoloGeo is supported by a high-quality dataset, BF-30k, annotated with structured multi-evidence \underlinebias-\underlinefree reasoning chains. By incorporating multi-dimensional rewards, HoloGeo explicitly encourages balanced attention over diverse visual cues and achieves evidence-driven joint reasoning. Extensive experiments demonstrate that HoloGeo not only maintains excellent performance on IM2GPS3K and YFCC4k but also significantly outperforms existing open-source VLMs on LandmarkBias-3K, validating its effectiveness for robust geospatial reasoning. Project page: \urlhttps://hologeo.github.io/.

Abstract:
Multi-modal Sequential Recommendation (SR) incorporates rich side information (e.g., textual and visual features) to enhance dynamic user preference modeling. However, existing frameworks inevitably suffer from a Dual-Noise Dilemma: (1) Feature-level redundancy stemming from the semantic gap between generic pre-trained representations and fine-grained recommendation intent; and (2) Sequence-level stochasticity induced by spurious interactions such as accidental clicks. To break this bottleneck, we propose DDMSR, a novel Dual-level Denoising Multi-modal Sequential Recommendation framework that systematically purifies signals from both feature-topological and sequence-frequency perspectives. Specifically, we first design a graph-based feature denoising module that leverages Laplacian smoothing on item semantic graphs as a structural low-pass filter, effectively suppressing high-frequency semantic noise while preserving salient features. For sequence purification, we introduce a frequency-domain sequence denoising module, utilizing the Fast Fourier Transform and a learnable frequency filter to adaptively modulate the interaction spectrum and attenuate anomalous signals. Furthermore, a multi-modal contrastive alignment objective is incorporated to bridge the heterogeneity gap and enforce cross-modal semantic consistency. Extensive experiments on four public benchmark datasets demonstrate that DDMSR consistently outperforms state-of-the-art baselines, providing a highly robust and efficient solution for multi-modal sequential recommendation. The source code is available at: ~\hrefhttps://github.com/jluo00/DDMSR\textcolorbluehttps://github.com/jluo00/DDMSR.

Authors: Shuo Wang (Institute of Computing Technology, Chinese Academy of Sciences), Jilin Mei (Institute of Computing Technology, Chinese Academy of Sciences), Fuyang Liu (Institute of Computing Technology, Chinese Academy of Sciences), Wenfei Guan (University of Chinese Academy of Sciences), Fanjie Kong (Xi'an Jiaotong University), Zhihua Zhao (Beijing Institute of Technology), Shuai Wang (Institute of Computing Technology, Chinese Academy of Sciences), Chen Min (Institute of Computing Technology, Chinese Academy of Sciences), Yu Hu (Institute of Computing Technology, Chinese Academy of Sciences)

Abstract:
Feedforward Gaussian Splatting has recently emerged as an efficient paradigm for 4D reconstruction in autonomous driving. However, in unstructured off-road scenes, its performance degrades due to high-frequency geometry, ego-motion jitter, and increased non-rigid dynamics. These factors introduce conflicting Gaussian observations across timestamps, leading to either over-smoothed renderings or structural artifacts. To address this issue, we propose Ground4D, a spatially-grounded 4D feedforward framework for pose-free off-road reconstruction. The key idea is to resolve temporal conflicts through spatially localized conditioning. Specifically, we introduce voxel-grounded temporal Gaussian aggregation, which partitions the canonical Gaussian space into spatial voxels and performs query-conditioned temporal attention within each voxel. Intra-voxel softmax normalization ensures that temporal selectivity and spatial occupancy become mutually reinforcing rather than conflicting. We furthermore introduce surface normal cues as auxiliary geometric guidance to regularize the geometry of Gaussian primitives. Extensive experiments on ORAD-3D and RELLIS-3D demonstrate that Ground4D consistently outperforms existing feedforward methods in reconstruction quality and generalizes zero-shot to unseen off-road domains.

Abstract:
Video fusion aims to integrate complementary cues from heterogeneous sources, producing visually informative videos with coherent spatio-temporal structures. Jointly learning cross-modal complementarity and temporal dependencies typically relies on large-scale, high-quality video datasets, whose storage and repeated use during optimization incur substantial costs, making data-efficient learning highly desirable. To address this challenge, we rethink video fusion from the perspective of dataset condensation and propose STAR-VF, the first framework that condenses spatio-temporal representations for data-efficient video fusion. STAR-VF condenses the original training set into a compact set of learnable spatio-temporal representations, which are decoded into synthetic video clips and optimized using feature-level supervision provided by a frozen proxy model. The resulting condensed data preserves fusion-relevant spatial and temporal cues without requiring ground-truth fused videos as supervision. To adapt fusion learning to condensed representations, we equip the final fusion model with flow-free inter-frame temporal modeling and query-driven intra-frame aggregation, enabling direct temporal interaction and adaptive selection of fusion-relevant information in representation space. Extensive experiments on infrared-visible, multi-exposure, multi-focus, and medical video fusion benchmarks demonstrate that STAR-VF, trained with only approximately 1/16 of the original temporal training set, achieves fusion quality comparable to methods trained on the full data. Further analyses validate the contributions of learned condensation and flow-free temporal modeling, and demonstrate the robustness of the condensed supervision across different proxy encoders and downstream fusion models. Project page: https://zju-xyc.github.io/STAR-VF-Project-Page/.

Abstract:
Point cloud quality plays a critical role in 3D acquisition, reconstruction, rendering, and perception, yet existing point cloud quality assessment (PCQA) research remains largely centered on scalar score prediction. In practical inspection scenarios, quality assessment often involves identifying defects, characterizing dominant issue types, assessing downstream usability, and providing evidence-supported descriptions, which are not explicitly evaluated by current benchmarks. We introduce PointQ-Bench, a benchmark designed to extend PCQA from scalar scoring toward comprehensive quality understanding. PointQ-Bench consists of 3,083 point clouds spanning authentic scans, simulated distortions, and AI-generated content, covering eight major issue types. Each sample is annotated with mean opinion scores (MOS), quality levels, issue tags, expert-grounded descriptions, and 12,332 questionanswer pairs. The benchmark supports three perception-oriented tasksanomaly sensing, defect diagnosis, and usability grading, as well as a cognition-oriented task of open-ended quality reporting. To evaluate free-form quality descriptions, we further propose SSFRQ-5D, a five-dimensional evaluation protocol validated through humanAI agreement analysis. Extensive experiments on 14 visionlanguage models and traditional PCQA baselines reveal a consistent perceptiondiagnosis gap: while current models exhibit emerging abilities in coarse defect perception, they struggle with grounded diagnosis and quality calibration. Strong 2D MLLMs generally outperform existing 3D VLMs, and the benefit of additional views or point-level inputs is non-uniform, varying across tasks, data sources, and modelsparticularly under boundary-ambiguous conditions. Overall, PointQ-Bench provides a diagnostic testbed for advancing reliable and interpretable point cloud quality understanding. Project repository: https://github.com/Mujeco/PointQ-Bench.

Abstract:
3D Gaussian Splatting (3DGS) achieves high-quality real-time novel-view synthesis, yet its GPU efficiency is often limited by heterogeneous workloads and sequential alpha compositing. In practice, Gaussian primitives exhibit highly uneven spatial density and footprint sizes, causing substantial workload variance across threads/warps and introducing heavy-tailed latency in both per-Gaussian preprocessing and per-tile rasterization. Moreover, conventional tile-based rasterization relies on strictly ordered alpha blending, which exposes limited intra-pixel parallelism and amplifies warp divergence under irregular visibility and contribution counts. We present a scale-aware acceleration framework that reshapes 3DGS workloads to improve parallel efficiency without modifying the underlying representation. For small-footprint Gaussians, we aggregate multiple primitives and evenly distribute their computation across warp lanes to reduce underutilization. For large-footprint Gaussians, we decompose each primitive into independent sub-tasks and dispatch them across blocks to avoid single-block stragglers. To address the sequential dependency of alpha blending, we further introduce bucket-based parallel local compositing: per-tile primitives are partitioned into fixed-size buckets that are rendered independently and merged through an associative compositing rule, preserving correctness and differentiability. Across diverse standard benchmarks, our method consistently reduces training and rendering latency while maintaining comparable visual performance. Source code is available at \hrefhttps://github.com/GoogolplexGoodenough/FasterGShttps://github.com/\allowbreakGoogolplexGoodenough/\allowbreakFasterGS.

Abstract:
Diffusion models primarily rely on general-purpose ODE-based numerical methods for training-free accelerated sampling, but performance degrades with coarse time discretization. In this paper, we propose optimal linear leaps (LEAP), a specialized method grounded in theoretical and empirical investigations of sampling trajectory geometry. Built upon low-dimensionality and cross-trajectory consistency, LEAP formulates a supervised least-squares optimization problem within the local trajectory subspace, achieving a precise approximation of future states without model evaluations. Extensive experiments demonstrate that LEAP outperforms previous methods across diverse modalities, including text-to-image (Stable Diffusion 3.5, FLUX.1-dev) and image-to-3D (TRELLIS.2) generation. Furthermore, LEAP can be seamlessly integrated into existing solver based methods, offering a plug-and-play enhancement for modern diffusion models. Code is available at \urlhttps://github.com/zju-pi/LEAP.

Abstract:
Joint audio-video generation models, such as dual-branch diffusion transformers, have demonstrated that generating audio and video within a unified model yields stronger cross-modal coherence than cascaded pipelines. However, existing joint audio-video diffusion models typically couple the two modalities throughout the entire denoising process via pervasive attention mechanisms, treating both high-level semantics and low-level signal details in a fully entangled manner. We argue that such uniform coupling is suboptimal for talking head synthesis. While audio and facial motion are strongly correlated at the semantic and temporal levels, their low-level realizationsacoustic signals and visual texturesfollow distinct rendering processes. Enforcing joint modeling across all levels can therefore lead to unnecessary entanglement and reduced modeling efficiency. Instead, we posit that joint modeling should primarily occur at high-level semantic and temporal information, where cross-modal alignment is essential, while low-level refinement should be handled by modality-specific decoders. Building on this observation, we propose Talker-T2AV, an autoregressive diffusion framework that realizes this principle: high-level cross-modal modeling in a shared autoregressive backbone, with low-level refinement delegated to modality-specific decoding. A shared autoregressive language model serves as a high-level temporal planner, jointly reasoning over audio and video in a unified patch-level token space with autoregressive left-to-right generation. Two lightweight diffusion transformer heads then decode the language model's hidden states into frame-level audio and video latents. Experiments on talking portrait benchmarks show that Talker-T2AV outperforms dual-branch diffusion transformer baselines in lip-sync accuracy, video quality, and audio quality, and further achieves stronger cross-modal consistency than cascaded pipelines, highlighting the complementary benefits of joint modeling. The samples are available at https://talker-t2av.github.io/.

Abstract:
Robust low-light imaging remains challenging for the community. Recent studies have explored fusing Near-Infrared (NIR) with noisy RGB to achieve improved enhancement, yet most methods depend on carefully curated training data pairs, with limited robustness under different scenarios. This paper offers a new perspective for RGB-NIR low-light imaging by incorporating 3D-aware neural modeling. Without using clean RGB supervision, a powerful model can be optimized to implicitly fuse extremely noisy RGB observations with NIR cues in 3D space, effectively recovering clean RGB images. The proposed model obviates the requirement for clean RGB data collection, generalizes across different noise levels. Extensive evaluations on synthetic and real data demonstrate its superiority. Codes available: \urlhttps://github.com/MyNiuuu/3DarkFusion

Abstract:
Face restoration has achieved significant advancements through the years of development. However, maintaining high fidelity and authenticity while avoiding artifacts remains challenging, especially in extreme degradation scenarios. This highlights the need for models that are more ``honest'' in their reconstruction from low-quality inputs, accurately reflecting original characteristics. In this work, we propose HonestFace, a novel approach for honest face restoration with identity consistency and realistic textures. To achieve this, HonestFace incorporates several key components. First, we propose an identity embedder to effectively capture and preserve crucial identity features from both the low-quality input and multiple reference faces. Second, a masked face alignment method is presented to enhance fine-grained details and textural authenticity. Furthermore, we present a new real-world reference-based face restoration dataset, MultiRefCeleb-Test, with multiple high-quality references and low-quality inputs. Leveraging these contributions within a one-step diffusion model framework, HonestFace delivers excellent restoration results in terms of facial fidelity and realism. Experiments demonstrate that our approach surpasses state-of-the-art methods, achieving superior performance in both visual quality and quantitative assessments. The code and pre-trained models are available at \urlhttps://github.com/jkwang28/HonestFace.

Abstract:
Facing scaling laws, video data from the internet becomes increasingly important. However, collecting extensive videos that meet specific needs is extremely labor-intensive. In this work, we study the way to expedite this collection process and propose VC-Agent, the first interactive agent that is able to understand users' queries and feedback, and accordingly retrieve/scale up relevant video clips with minimal user input. At the user workflow, our agent defines various user-friendly ways for the user to specify requirements based on textual descriptions and confirmations. For agent functions, we leverage existing multi-modal LLMs to connect users' requirements with video content. More importantly, we propose two novel filtering policies that can be updated when user interaction is continually performed. Finally, we provide a new benchmark for personalized video dataset collection, where extensive experiments demonstrate the effectiveness and efficiency of our agent. Moreover, we carefully conduct a user study, where our agent successfully curated 10 large-scale video datasets for diverse customized applications, demonstrating the systems robustness and real-world applicability for large-scale production.

Abstract:
During sessions with scientific instructional videos, students are often overly captivated by vivid experimental phenomena, while neglecting to explore the underlying principles, consider potential safety hazards and clarify procedural steps. Recently, video-based question generation has emerged as an effective educational approach for fostering cognitive engagement, with large multimodal models (LMMs) forming its backbone and facilitating deeper conceptual understanding. However, LMMs exhibit cross-modal instructional misalignment in laboratory video understanding, often leading to unreliable or incomplete instructional guidance. In this paper, we make the following contributions: (i) We construct LabQG, a diverse dataset of scientific instructional videos comprising 13K question-answer pairs from 6K video clips across 21 scientific subjects, featuring five distinct question types grounded in laboratory instruction. (ii) We propose LabSeed, a semantic-adaptive tool-integrated framework that leverages knowledge retrieval, procedural modeling and keyframe selection for generating pedagogically meaningful laboratory questions. (iii) Extensive experiments demonstrate that LabSeed consistently outperforms 14 foremost LMMs baselines in generating effective and faithful laboratory questions, validating the superiority of our dataset and model. Our deeper analysis further uncovers an implicit toolscenario relationship, revealing underlying mechanisms of specific tool effectiveness. We believe that LabQG and LabSeed provide a solid and promising foundation for advancing AI-assisted learning in scientific instructional videos. The project page is available at https://github.com/Gary-code/LabSeed.

Abstract:
LLM-based multimodal emotion recognition relies on static parametric memory and often hallucinates when interpreting nuanced affective states. In this paper, given that single-round retrieval-augmented generation is highly susceptible to modal ambiguity and therefore struggles to capture complex affective dependencies across modalities, we introduce AffectAgent, an affect-oriented multi-agent retrieval-augmented generation framework that leverages collaborative decision-making among agents for fine-grained affective understanding. Specifically, AffectAgent comprises three jointly optimized specialized agents, namely a query planner, an evidence filter, and an emotion generator, which collaboratively perform analytical reasoning to retrieve cross-modal samples, assess evidence, and generate predictions. These agents are optimized end-to-end using Multi-Agent Proximal Policy Optimization (MAPPO) with a shared affective reward to ensure consistent emotion understanding. Furthermore, we introduce Modality-Balancing Mixture of Experts (MB-MoE) and Retrieval-Augmented Adaptive Fusion (RAAF), where MB-MoE dynamically regulates the contributions of different modalities to mitigate representation mismatch caused by cross-modal heterogeneity, while RAAF enhances semantic completion under missing-modality conditions by incorporating retrieved audiovisual embeddings. Extensive experiments on MER-UniBench demonstrate that AffectAgent achieves superior performance and robust generalization across complex emotional scenarios. Our code will be released at: \urlhttps://github.com/Wz1h1NG/AffectAgent. Extensive experiments on MER-UniBench demonstrate that AffectAgent achieves superior performance across complex scenarios. Our code will be released at: \urlhttps://github.com/Wz1h1NG/AffectAgent.

Abstract:
With the integration of multimodal large language models (MLLMs) into robotic systems and AI applications, embedding emotional intelligence (EI) capabilities is essential for enabling these models to perceive, interpret, and respond to human emotions effectively in real-world scenarios. Existing static, text-based, or text-image benchmarks overlook the multimodal complexities of real interactions and fail to capture the dynamic, context-dependent nature of emotional expressions, rendering them inadequate for evaluating MLLMs EI capabilities. To address these limitations, we introduce EmoBench-M, a systematic benchmark grounded in established psychological theories, designed to evaluate MLLMs across 13 evaluation scenarios spanning three hierarchical dimensions: foundational emotion recognition (FER), conversational emotion understanding (CEU), and socially complex emotion analysis (SCEA). Evaluation was conducted on 27 state-of-the-art MLLMs, using both objective task-specific metrics and LLM-based evaluation, revealing a substantial performance gap relative to human-level competence. Even the best-performing models, Gemini-3.0-Pro and GPT-5.2, achieve the highest scores on EmoBench-M, 70.5 and 66.5 points. Specialized models such as AffectGPT exhibit uneven performance across EmoBench-M, demonstrating strengths in certain scenarios but generally lacking comprehensive emotional intelligence. By providing a comprehensive, multimodal evaluation framework, EmoBench-M captures both the strengths and weaknesses of current MLLMs across diverse emotional contexts. All benchmark resources, including datasets and code, are publicly available at \hrefhttps://emo-gml.github.io/\textcolormyblueEmobench-M, facilitating further research and advancement in MLLM emotional intelligence.

Abstract:
While Vision-Language Models (VLMs) excel at multimodal tasks, they are increasingly vulnerable to sophisticated multimodal jailbreak attacks. Conventional defenses typically introduce significant training overhead, risk degrading model capabilities, or cause severe inference latency via external modules. Additionally, current internal state detection methods rely on coarse-grained feature analysis, often failing to capture the micro-level conflicts generated by jailbreak attacks. To overcome these challenges, we analyze multimodal jailbreaks through mechanistic interpretability. We demonstrate that successful attacks typically do not eradicate intrinsic safety knowledge but instead cause a widespread attenuation of refusal signals. Even when global safety alignment is bypassed, a small, robust subset of Attention and FFN components continues to exhibit directional resistance against malicious queries. Building upon this empirical finding, we introduce Dual-grained Internal Refusal Estimation (DIRE), a training-free detection framework. Unlike prior internal monitors that are either macro-only or attention-only, DIRE jointly probes resilient Attention routing and FFN semantic memory within effective semantic layers to detect multimodal jailbreaks in a training-free manner. Comprehensive evaluations on mainstream VLMs and diverse benchmarks demonstrate that DIRE achieves competitive and strong performance, with an average AUROC of 0.9800. For better reproducibility, the anonymized code is available at: https://github.com/xiyuzzz-njust/DIRE.

Abstract:
Referring Camouflaged Object Detection (Ref-COD) requires segmenting hidden targets guided by reference cues. While supervised methods are annotation-heavy and training-free approaches via sparse point-prompting are sensitive to localization errors, we propose GFR-SAM, a robust three-stage training-free framework. GFR-SAM shifts the paradigm from fragile point-matching to a "Generate-Filter-Refine" pipeline. First, we introduce In-Context Exemplar-guided Segmentation, empowering SAM3 with cross-image inference to generate candidate masks via holistic visual exemplars, bypassing its native intra-image constraints. Second, a Region-Global Contrastive Filtering module ranks candidates through DINOv3-based prototypical alignment, effectively suppressing background distractors. Finally, a Geometric-Semantic Refinement module synergizes bounding box and text prompts to recover fine-grained boundaries and enhance instance recall. Evaluated on the R2C7K benchmark, GFR-SAM outperforms existing training-free methods by 8.7% in weighted F-measure (F_\beta^w) and competes with supervised state-of-the-art counterparts. Ultimately, this work underscores the potential of unlocking SAM3's latent capability for cross-image In-Context prompting, establishing a robust, training-free paradigm that effectively bridges the gap between general-purpose foundation models and specialized, label-intensive perception tasks without the need for task-specific fine-tuning. Code is available at \urlhttps://github.com/ynulonger/GFR-SAM.

Abstract:
We introduce Latent-WAM, an efficient end-to-end autonomous driving framework that achieves strong trajectory planning through spatially-aware and dynamics-informed latent world representations. Existing world-model-based planners suffer from inadequately compressed representations, limited spatial understanding, and underutilized temporal dynamics, resulting in sub-optimal planning under constrained data and compute budgets. Latent-WAM addresses these limitations with two core modules: a Spatial-Aware Compressive World Encoder (SCWE) that distills geometric knowledge from a foundation model and compresses multi-view images into compact scene tokens via learnable queries, and a Dynamic Latent World Model (DLWM) that employs a causal Transformer to autoregressively predict future world status conditioned on historical visual and ego status representations. Extensive experiments on NAVSIM v2 and HUGSIM demonstrate new state-of-the-art results: 89.3 EPDMS on NAVSIM v2 and 28.9 HD-Score on HUGSIM, surpassing the best prior perception-free method by 3.2 EPDMS with significantly less training data and a compact 104M-parameter model. The code and model weights have been open-sourced at \urlhttps://github.com/yukki555/Latent-WAM.

Abstract:
\beginabstract Industrial anomaly generation is a crucial method for alleviating the data scarcity problem in anomaly detection tasks. Most existing anomaly synthesis methods rely on single-step generation mechanisms, lacking complex reasoning and iterative optimization capabilities, making it difficult to generate anomaly samples with high semantic realism. We propose AnomalyAgent, an anomaly synthesis agent with self-reflection, knowledge retrieval, and iterative refinement capabilities, aiming to generate realistic and diverse anomalies. Specifically, AnomalyAgent is equipped with five tools: Prompt Generation (PG), Image Generation (IG), Quality Evaluation (QE), Knowledge Retrieval (KR), and Mask Generation (MG), enabling closed-loop optimization. To improve decision-making and self-reflection, we construct structured trajectories from real anomaly images and design a two-stage training framework: supervised fine-tuning followed by reinforcement learning. This process is driven by a three-part reward mechanism: (1) task rewards to supervise the quality and location rationality of generated anomalies; (2) reflection rewards to train the model's ability to improve anomaly synthesis prompt; (3) behavioral rewards to ensure adherence to the trajectory. On the MVTec-AD dataset, AnomalyAgent achieves IS/IC-L of 2.10/0.33 for anomaly generation, 57.0% classification accuracy using ResNet34, and 99.3%/74.2% AP at the image/pixel level using a simple UNet, surpassing all zero-shot SOTA methods. The code is available at \urlhttps://github.com/sssjmmm/AnomalyAgent. \endabstract

Abstract:
Few-shot anomaly detection (FSAD) has recently benefited from vision-language models such as CLIP, which enable anomaly detection by aligning visual features with text descriptions of normal and abnormal states. However, existing methods typically rely on static text prompts that are applied uniformly across the entire feature hierarchy and spatial dimensions. This rigid global-to-local matching fails to capture the highly localized and scale-dependent physical variations of industrial defects. To address this, we propose DriftAD, a FSAD framework built on three key modules. First, an Anomaly Signal Amplification (ASA) module enhances subtle defect signals through spatial and frequency branches before text-visual matching. Second, Visually-Guided Text Drift (VGTD) dynamically transforms frozen CLIP text embeddings, steering them into layer-wise, spatially-adaptive anomaly descriptors conditioned on local visual context at each encoder depth. Third, Drift-Guided Spatial Gating (DGSG) uses the drifted abnormal descriptor as a spatial probe to selectively enhance anomaly-relevant visual features. Additionally, a drift separation loss prevents representational collapse of the drifted descriptors, and a gate supervision loss enforces spatially discriminative gating in DGSG. Extensive experiments on MVTec-AD and VisA demonstrate state-of-the-art performance across all 1-, 2-, and 4-shot settings on both image-level and pixel-level metrics. Code is available at https://github.com/wenyang001/DriftAD.

Abstract:
Industrial anomaly inspection is severely hindered by the extreme scarcity of defective training samples. Although diffusion models offer a promising solution for synthetic anomaly generation, they struggle in few-shot settings due to an inherent conflict between faithfully generating normal regions of anomalous images and accurately synthesizing diverse, localized defects. To address this, we propose ResAnoDiff, a decoupled generative framework that reformulates anomaly synthesis into a structured two-stage process. First, our approach leverages a pre-trained diffusion backbone with LoRA adaptation to generate structurally faithful images containing coarse anomaly patterns. Second, a Residual Anomaly Enhancement module injects high-frequency defect details by modeling anomaly-specific residuals. To confine this enhancement to the anomaly regions, an Adaptive Mask Generation module uses visual context and text to determine the anomaly's exact shape and location. Additionally, a Reference-Guided Sampling strategy anchors normal region generation to real samples, preventing overfitting. Extensive experiments on the MVTec AD and VisA datasets confirm that ResAnoDiff achieves state-of-the-art anomaly generation performance, consistently boosting downstream anomaly detection performance. Code is available at https://github.com/getmoons/ResAnoDiff.

Abstract:
Text-Motion Retrieval (TMR) aims to retrieve 3D motion sequences semantically relevant to text descriptions. However, matching 3D motions with text remains highly challenging, primarily due to the intricate structure of the human body and its spatiotemporal dynamics. Existing approaches often overlook these complexities, relying on general encoding methods that fail to distinguish different body parts and their dynamics, limiting precise semantic alignment. To address this, we propose WaMo, a novel wavelet-based multi-frequency feature extraction framework. It fully captures joint-specific and time-varying motion details at multiple resolutions for individual joint trajectories, extracting discriminative motion features to achieve fine-grained alignment with texts. WaMo has three key components: (1) Trajectory Wavelet Decomposition decomposes motion signals into frequency components that preserve both local kinematic details and global motion semantics. (2) Trajectory Wavelet Reconstruction uses learnable inverse wavelet transforms to reconstruct original joint trajectories from extracted features, ensuring the preservation of essential spatiotemporal information. (3) Disordered Motion Sequence Prediction reorders shuffled motion sequences to improve learning of inherent temporal coherence, enhancing motion-text alignment. Extensive experiments demonstrate WaMo's superiority, achieving 17.0% and 18.2% relative improvements in Rsum on HumanML3D and KIT-ML datasets, respectively, outperforming existing state-of-the-art (SOTA) methods. Code is available at https://github.com/3DAgentWorld/WaMo/.

Abstract:
Recently, an increasing number of multimodal (text and audio) benchmarks have emerged, primarily focusing on evaluating models' understanding capability. However, exploration into assessing generative capabilities remains limited, especially for open-ended long-form content generation. Significant challenges lie in no reference standard answer, no unified evaluation metrics and uncontrollable human judgments. In this work, we take podcast-like audio generation as a starting point and propose PodEval, a comprehensive and well-designed open-source evaluation framework. In this framework: 1) We construct a real-world podcast dataset spanning diverse topics, serving as a reference for human-level creative quality. 2) We introduce a multimodal evaluation strategy and decompose the complex task into three dimensions: text, speech and audio, with different evaluation emphasis on ``Content" and ``Format". 3) For each modality, we design corresponding evaluation methods, involving both objective metrics and subjective listening test. We leverage representative podcast generation systems (including open-source, close-source, and human-made) in our experiments. The results offer in-depth analysis and insights into podcast generation, demonstrating the effectiveness of PodEval in evaluating open-ended long-form audio. This project is fully open-source to facilitate public use: \urlhttps://github.com/yujxx/PodEval.

Abstract:
Unrestricted adversarial transfer attacks are important for evaluating the black-box robustness of deep visual models. Diffusion-based attacks have shown promising transferability and visual imperceptibility by optimizing adversarial perturbations along denoising trajectories in latent space. However, existing methods are limited by two challenges: memory-intensive multistep backpropagation and frequency-agnostic perturbation over intermediate latents. To address these issues, we propose IDATA, a memory-efficient diffusion framework for unrestricted adversarial transfer attack. IDATA consists of two key components: an Invertible Diffusion Module (IDM) and a Low-Frequency Constraint Module (LFCM). Specifically, IDM reformulates adversarial optimization over diffusion trajectories as an invertible process, enabling constant-memory backpropagation through on-demand reconstruction of intermediate states instead of storing the full denoising chain. Moreover, LFCM leverages Discrete Wavelet Transform (DWT) to decompose latent variables into low- and high-frequency components, restricting perturbations to semantically stable low-frequency subspaces, thereby improving transferability while preserving visual imperceptibility. Extensive experiments on multiple benchmarks and diverse model architectures demonstrate that IDATA consistently outperforms state-of-the-art baselines in attack success rate, memory efficiency, and visual imperceptibility. These results suggest that IDATA is a promising tool for black-box robustness evaluation of deep visual models. Code is available at \urlhttps://github.com/colourful-pan/IDATA.

Abstract:
Although Multimodal Large Language Models have achieved strong performance across a wide range of vision-language tasks, they still suffer from hallucinations, where model outputs become inconsistent with the visual content, textual context, or commonsense knowledge. Existing studies primarily address this problem through coarse-grained detection. However, these approaches often provide insufficient diagnostic information for understanding hallucination types and supporting downstream hallucination mitigation. To bridge this gap, we propose fine-grained hallucination diagnosis for MLLMs, a new unified task that jointly performs hallucination detection, classification, and interpretable explanation generation. We develop an automated data generation pipeline and construct HalluScope-30K, a large-scale diagnostic dataset covering eight sources and five task categories. Based on this dataset, we design a multi-granular joint reward function and train two diagnosis models, HalluScope-4B and HalluScope-8B, which achieve state-of-the-art performance on both the MHALO benchmark and our fine-grained hallucination classification benchmark. Notably, detection and classification are mutually beneficial under joint optimization. Furthermore, diagnosis-driven feedback experiments show that the fine-grained diagnostic explanations produced by our model effectively guide target models to correct their hallucinations, with full diagnosis substantially outperforming all baselines on both Qwen3-VL-8B-Instruct and LLaVA-1.5-7B. Our code, data, and models are available at \urlhttps://github.com/wkinglin/HalluScope.

Abstract:
Rare concept generation focuses on synthesizing customized images conditioned on text prompts that describe objects with unusual attributes. Previous works failed to align the generated images with rare concepts, resulting in incorrect attribute rendering or inconsistent composition of concepts. Such failures, as we observed, stem from the inherent common knowledge bias in the training stage of diffusion models, where objects are strongly associated with their common attributes, making it difficult to break these associations when generating rare concepts. To address such challenges, in this paper, we propose a novel Counterfactual Inference-based Diffusion approach, dubbed CI-Diff. CI-Diff blocks the interference of the model's inherent common knowledge bias and utilizes the Natural Direct Effect to capture the independent influence of the text prompt of rare concepts on image generation so that decoupling the unusual attributes from the rare concepts. To this end, we reformulate the classifier-free guidance mechanism to highlight the atypical attributes. To the best of our knowledge, we are the first to introduce causal inference into the rare concept generation task. Extensive experiments on the RareBench benchmark validate the superiority of CI-Diff over state-of-the-art diffusion models. Our code can be accessed from \urlhttps://github.com/200204jzy/CI-Diff.

Abstract:
Predicting spatial gene expression from hematoxylin and eosin (H\&E)-stained images offers a cost-effective alternative to spatial transcriptomics (ST). However, existing methods treat H\&E images as generic visual inputs and ignore their intrinsic biological hierarchy, where spatially organized cell types collectively form functional tissue microenvironments that govern local gene expression programs. To bridge this gap, we formulate H\&E-to-ST prediction as a cross-modal semantic translation task and propose Path2ST, a hierarchically grounded autoregressive framework featuring three key components: (i) a Hierarchical Cell-Tissue Conditioning mechanism that fuses explicit and implicit cellular features with tissue-level semantic representations to construct hierarchical conditioning signals; (ii) a Scale-Adaptive Autoregressive Generation process over a hierarchical semantic vocabulary, enabling coarse-to-fine, biologically consistent expression synthesis; and (iii) SpectraLoss, a full-spectrum objective that jointly enforces ordinal fidelity, models transcriptional bursts, and aligns semantic structures with cell types. Extensive experiments on three datasets demonstrate state-of-the-art performance, validating that Path2ST generates highly accurate and spatially coherent transcriptomic profiles. The related code is released at https://github.com/RuochenLiu23/Path2ST.

Abstract:
Masked Image Generation Models (MIGMs) have achieved great success, yet their efficiency is hampered by the multiple steps of bi-directional attention. In fact, there exists notable redundancy in their computation: when sampling discrete tokens, the rich semantics contained in the continuous features are lost. Some existing works attempt to cache the features to approximate future features. However, they exhibit considerable approximation error under aggressive acceleration settings. We attribute this to their limited expressivity and the failure to account for sampling information. To fill this gap, we propose learning a lightweight model that incorporates both previous features and sampled tokens, and regresses the average velocity field of feature evolution. The model has moderate complexity that suffices to capture the subtle dynamics while keeping lightweight compared to the original base model. We apply our method to two representative MIGMs and tasks. In particular, on the state-of-the-art Lumina-DiMOO, it achieves over 4x acceleration of text-to-image generation while maintaining quality, significantly pushing the Pareto frontier of masked image generation. The code and model weights are available at https://github.com/Kaiwen-Zhu/MIGM-Shortcut.

Abstract:
Referring Video Object Segmentation (RVOS) segments the object described by a natural language expression throughout a video. Despite steady progress, prevailing approaches often rely on excessive task-specific training and cross-modal fusion, yet remain vulnerable to language-vision mismatch: CLIP hallucination, background co-occurrence, and tracking bias cause identity switches and temporal drift under distractors, occlusions, or large motions. We introduce POLAR (Polarized Text Differentials), a training-free, plug-and-play versatile vision-language bridge that provides a robust language-consistency signal to reliably identify the referred target and guide its propagation across frames. Without fine-tuning any backbone, POLAR establishes a precise mask-aware, facet-conditioned language-vision relation by contrasting facet-consistent and facet-inconsistent evidence; incorporates Top M evidence aggregation, logistic calibration, and facet gating with adaptive weighting; and thereby suppresses hallucination, discounts co-occurrence, and mitigates tracking bias to stabilize segmentation across long videos and challenging scenes. POLAR achieves a new state-of-the-art performance on Ref-YouTube-VOS, Ref-DAVIS17 and MeViS, all while remaining simple, efficient, and easy to deploy. Code is available at https://github.com/Louis-zxLiu/POLAR.

Abstract:
Open-Vocabulary Audio-Visual Semantic Segmentation (OV-AVSS) aims to perform pixel-level segmentation of sound-emitting objects from an open set of categories. The previous method relies on a class-agnostic foreground definition, which groups semantically diverse objects into a heterogeneous positive set, causing the model to learn unstable sounding patterns and produce unreliable proposals. To address this, we reformulate the objective to be category-specific and propose a novel Acoustically Grounded Cost Learning (AGCL) framework to transform the static, audio-agnostic visual-text priors into dynamic, audio-grounded cost representations. For intra-category soundingness discovery, we devise Audio-Modulated Cost Generation (AMCG) and Audio-Guided Temporal Aggregation (AGTA) modules to enable both frame-level sounding region highlighting and video-level temporal refinement with a low-intrusive audio injection mechanism. For inter-category distractor discrimination, we introduce a Synergistic Distractor Mining (SDM) strategy, which selectively penalizes acoustically and semantically confusing negative categories to learn more discriminative decision boundaries. Extensive experiments on the AVSBench-OV dataset demonstrate that our method significantly outperforms previous state-of-the-art approaches, particularly on unseen categories. Code is available at \urlhttps://github.com/spyflying/AGCL.

Abstract:
Achieving physically accurate object manipulation in image editing is essential for its potential applications in interactive world models. However, existing visual generative models often fail at precise spatial manipulation, resulting in incorrect scaling and positioning of objects. This limitation primarily stems from the lack of explicit mechanisms to incorporate 3D geometry and perspective projection. To achieve accurate manipulation, we develop PhyEdit, an image editing framework that leverages explicit geometric simulation as contextual 3D-aware visual guidance. By combining this plug-and-play 3D prior with joint 2D--3D supervision, our method effectively improves physical accuracy and manipulation consistency. To support this method and evaluate performance, we present a real-world dataset, RealManip-40K, for 3D-aware object manipulation featuring paired images and depth annotations. We also propose ManipEval, a benchmark with multi-dimensional metrics to evaluate 3D spatial control and geometric consistency. Extensive experiments show that our approach outperforms existing methods, including strong closed-source models, in both 3D geometric accuracy and manipulation consistency. Code is available at https://github.com/nenhang/PhyEdit.

Abstract:
Binarized Neural Networks (BNNs) offer remarkable computational efficiency, yet their strict symmetry assumptions and limited representational capacity lead to severe performance degradation in complex tasks such as Underwater Image Enhancement (UIE). In this paper, we present a binarized architecture that achieves high-fidelity restoration while preserving the efficiency advantages of BNNs. The proposed framework is built upon a Calibrate-and-Compensate Block (CCB). Specifically, a pre-quantization Adaptive Distribution Shifter (ADS) dynamically aligns asymmetric activation distributions caused by underwater color distortions, while a post-quantization Statistical Compensation Module (SCM) restores the magnitude information lost during 1-bit quantization. Furthermore, we introduce a BNN-Friendly Sampling (BFS) strategy to eliminate mixed-precision bottlenecks and enable highly binarized end-to-end inference. Experiments show that our method achieves performance comparable to full-precision state-of-the-art models and outperforms existing binarized approaches, while maintaining high computational efficiency. This work provides a promising direction for ultra-efficient image restoration in complex visual environments. The code is publicly available at \urlhttps://github.com/Sy-ling/BU-Net.

Abstract:
Monocular 4D head reconstruction remains challenging as monocular videos provide only view-limited supervision. Although leveraging generative priors to synthesize pseudo-ground truth (pseudo-GT) for unseen views has become a common strategy to expand supervision, directly applying such priors introduces two compounding challenges. The inherent hallucination of diffusion models produces unreliable pseudo-GT that may corrupt faithful identity details in well-observed regions. Meanwhile, the geometric fragility induced by monocular overfitting causes Gaussians to drift from the true geometric surface, producing degraded avatars that in turn trigger worse pseudo-GT generation. Therefore, we propose ConfAvatar, a novel confidence-aware self-bootstrapping framework that addresses both challenges jointly. Its key idea is to modulate prior-based completion according to the confidence from current target view, so that reliable observed content is preserved and under-constrained regions receive more plausible completion. To make such confidence well-defined, we first introduce SDF-Opacity Co-Regularization (SOC), which anchors the Gaussian avatar to a coherent surface by coupling Gaussian opacity with a signed distance field. Based on this representation, we formulate view-dependent Gaussian confidence and render confidence maps for novel views, which guide a pretrained diffusion prior through a confidence-guided multi-scale injection mechanism. This enables identity-preserving completion of weakly constrained regions while preventing reliable content from being overwritten. Extensive experiments demonstrate that ConfAvatar achieves the state-of-the-art performance for monocular 4D head reconstruction. Our code is available at https://github.com/fan2goa1/ConfAvatar.

Abstract:
In point-cloud representation learning, the encoder maps raw data to high-dimensional feature manifolds and projects them onto low-dimensional semantic manifolds. Corruptions distort this geometric consistency, requiring robust models to calibrate manifold shift. To this end, we propose a novel robust point-cloud representation framework, termed Probabilistic Geometry-based Manifold Calibration (PGMC), which leverages geometric unity on the probability simplex to restore distorted manifold mapping dynamically. Specifically, our proposed PGMC method integrates three key components: (1) A Manifold Shift Degree Quantification module that filters out reliable manifold anchors through mapping different manifolds to the probability simplex. (2) A Hybrid Calibration Memory module that discretely stores global discriminative anchors and continuously refines local geometry. (3) A Dual Manifold Calibration strategy, which employs a geometry-preserved manifold to anchor the ideal semantic structure and a shift-aware manifold to capture distortions. By combining different pre-training models, we evaluate our method on three challenging benchmarks for point-cloud corruption and domain shift. Extensive experiments demonstrate that our PGMC method outperforms existing state-of-the-art methods, significantly enhancing the robustness of pre-trained representation models. Our code and supplementary materials are available at \urlhttps://github.com/ConstantineWayne/PGMC.

Abstract:
Multimodal Sarcasm Detection (MSD) identifies sarcastic intents by jointly comprehending visual and textual modalities, providing deeper contextual understanding than conventional text-only detectors. However, it remains challenging as it relies on capturing subtle, fine-grained pragmatic incongruities across heterogeneous modalities. Existing methods mainly adopt traditional feature fusion networks or Vision Large Language Models (VLLMs). Nevertheless, they encounter two key problems: (1) conventional representation learning focuses on explicit feature alignment but lacks the deductive logic necessary for deep pragmatic prediction, and (2) mainstream LLM-based generative approaches suffer from prohibitive computational latency due to autoregressive generation, limiting their viability under real-time constraints. To address this, we propose MIRD, a novel Multimodal Implicit Reasoning Distillation framework that balances profound cognitive reasoning with high efficiency requirements. Our method comprises three core designs. First, a specialized cross-modal alignment network is developed to realize early-stage, fine-grained feature interleaving, providing a robust perception foundation. Second, we introduce a Multi-Perspective Implicit Preference Distillation (MIPD) strategy, which utilizes structurally defective negative trajectories to explicitly internalize the powerful deductive logic of teacher LLMs into a compact implicit latent state. Third, through Implicit Reasoning-Driven Adaptation (IRDA), MIRD directly maps the converged implicit reasoning representations to task-specific classification boundaries, thereby completely bypassing the computational bottleneck of autoregressive decoding. Extensive experiments on three benchmark datasets (MMSD, MMSD2.0, and RedEval) demonstrate that MIRD achieves state-of-the-art accuracy while consistently delivering significant improvements in efficiency over native autoregressive baselines. Code is available at \urlhttps://github.com/possion0/MIRD.

Abstract:
Multimodal large language models (MLLMs) combine linguistic reasoning with visual perception, yet their ability to perform visual spatial planning under explicit or previously unseen rule constraints remains underexplored. This setting requires models to jointly understand spatial layouts, interpret natural-language rules, and plan valid actions accordingly. To address this gap, we introduce RuleMaze, a controllable benchmark in which MLLMs must navigate mazes while obeying natural-language rules of varying complexity. RuleMaze isolates rule-compliant spatial planning by requiring accurate perception, rule interpretation, and constrained action planning. To enable scalable and systematic rule construction, we propose LanguageLogicFunction Hybridization, which automatically generates natural-language rules and translates them into logical representations and executable validators, eliminating manual rule engineering. To improve rule following and generalization, we introduce Disentangled Multimodal Planning (DMP), which separates perception, execution, and rule verification through interpretable reasoning primitives. By disentangling these components, DMP facilitates systematic generalization to more complex and previously unseen rules, while providing transparent intermediate planning traces. Experiments demonstrate that DMP substantially improves rule compliance and planning success compared to end-to-end textual planning baselines. Overall, RuleMaze establishes a principled benchmark for studying grounded and interpretable rule-based spatial planning in MLLMs. Code is available at https://github.com/oceanflowlab/RuleMaze.

Abstract:
InfRared Small Target Detection (IRSTD) is a prominent and challenging task in computer vision. In recent years, text-guided methods have significantly improved detection performance. However, they still suffer from two key limitations. First, a single text description simultaneously modeling both background and target leads to semantic entanglement, which contradicts the objective of background suppression and target enhancement. Second, reliance on image-specific textual prompts (requiring additional external models such as CLIP during inference) results in deployment constraints. To address these issues, we propose a novel Dual-knowledge Guided Network (DGNet) based on multiple generalizable texts. Specifically, we design a Prior-knowledge Wavelet Modulation (PWM) module, which leverages dual textual priors that separately characterize large-scale backgrounds and sparse targets to effectively disentangle and modulate entangled semantics in the frequency domain. Furthermore, we introduce a Consensus-knowledge Directional Alignment (CDA) loss, which models the initial state and the ideal target across samples as `complex background' and `bright target', respectively, thereby constructing a clear and unified directional optimization trajectory for the model. Extensive experiments on three public datasets demonstrate the superior performance of DGNet and the effectiveness of each component. The source code is available at https://github.com/iLearn-Lab/MM26-DGNet.

Abstract:
Sequential Recommendation (SR) models infer user preferences from interaction histories. While transferable Multi-modal SR models outperform traditional ID-based approaches, existing methods struggle with slow fine-tuning convergence due to complex optimization requirements and negative transfer effects. We propose MMM4Rec (\underlineMulti-\underlineModal \underlineMamba for Sequential \underlineRecommendation), a novel Multi-modal SR framework that incorporates a dedicated algebraic constraint mechanism for efficient transfer learning. By combining State Space Duality (SSD)'s temporal decay properties with a globally-aware temporal modeling design, our model dynamically prioritizes key modality information, overcoming limitations of Transformer-based approaches. The framework implements a constrained two-stage process: (1) sequence-level cross-modal alignment via shared projection matrices, followed by (2) temporal fusion using our newly designed Cross-SSD module and dual-channel Fourier adaptive filtering. This architecture maintains semantic consistency while suppressing noise propagation. By incorporating algebraic structural constraints aligned with SR priors, MMM4Rec employs a simple and consistent cross-entropy objective across both pre-training and fine-tuning, enabling rapid fine-tuning convergence, substantially improving multimodal recommendation accuracy, and preserving strong transferability. Extensive experiments demonstrate MMM4Rec's state-of-the-art performance, achieving strong multi-modal retrieval capability and exhibiting 10× faster average convergence speed when transferring to large-scale downstream datasets. The implementation is available at link \urlhttps://github.com/AlwaysFHao/MMM4Rec.

Abstract:
Existing methods for thermal radiance field reconstruction treat the dynamic temperature simply as a function of time. However, the change of temperature is essentially driven by environmental factors. In outdoor scenes, surface temperatures are driven by an interplay between variable environmental excitations (\eg, solar radiation and air temperature) and intrinsic properties (\eg, terrain and material properties). To address this, we propose Eco-GS, an \underlineEnvironmental-fa\underlinect\underlineors-driven 3D \underlineGaussian \underlineSplatting framework for thermal radiance field reconstruction. First, Eco-GS establishes a high-fidelity geometric foundation via RGB-assisted 3DGS and geometry regularization. Then, it generates per-frame dynamic excitations based on timestamps, using scene geometry and meteorological data. Finally, a dual-branch temperature regressor predicts both the overall temperature trend and local temperature details, ultimately combining them to produce thermal images. To validate the proposed method, we collect a diurnal RGB-T drone dataset from multiple outdoor scenes. This dataset spans different time periods, featuring precise geospatial and meteorological information. Experimental results show that Eco-GS achieves a PSNR of 33.30 dB and an MAE of 0.769 C under seen environmental conditions, surpassing the second-best method by 3.97 dB and 0.524 C. Benefiting from the introduction of environmental factors, Eco-GS also achieves a PSNR of 32.22 dB and an MAE of 1.040 C under unseen conditions, surpassing the second-best method by 4.85 dB and 0.735~C. Our data and code are available \hrefhttps://github.com/XT5un/Eco-GS\textcolormagentahere.

Abstract:
With the growth of gaming, animation, and virtual reality industries, the demand for efficient generation of stylized 3D assets is rapidly increasing. However, existing approaches still struggle to jointly preserve style fidelity, geometric consistency, and generation efficiency, as most of them still rely on indirect 2D-to-3D stylization pipelines. This motivates a native 3D stylization framework that can explicitly disentangle style from geometry while remaining efficient. To this end, we propose DreamStyle3D, an efficient framework for stylized 3D asset generation built on a Decoupled Dual Cross-Attention mechanism. Our method explicitly separates geometric and stylistic features to enable efficient style injection while preserving structural consistency, and further adopts a lightweight training strategy to enhance style consistency and model generalization. In addition, we build an automated data pipeline and construct a dataset of about 15K content-style-stylized triplets for training and evaluation. Extensive experiments demonstrate that our DreamStyle3D can generate high-fidelity, geometrically consistent stylized 3D assets within 10 seconds, substantially improving efficiency while maintaining superior style quality and offering a new solution for 3D content creation. The project is available at https://github.com/NK-JittorCV/nk-3D/tree/main/models/DreamStyle3D.

Abstract:
Current visual tokenizers in Multimodal Large Language Models (MLLMs) predominantly rely on patch-based partitioning, which causes severe semantic mixture and object fragmentation in remote sensing imagery due to the irregular contours of geo-objects. Moreover, existing adaptive methods struggle to extract precise object-level tokens and lack dedicated geometric positional encodings for irregular regions. In this paper, we propose HeatTok, a semantic-aware tokenizer driven by thermodiffusion aggregation. Inspired by the physical principles of heat conduction, HeatTok adaptively merges adjacent homogeneous regions to generate semantically independent, object-aligned irregular tokens. To enable MLLMs to perceive these irregular shapes, we design the Gaussian Multimodal Rotary Positional Embedding (G-MRoPE), which models token spatial distributions via 2D Gaussians and explicitly injects center, scale, and orientation cues. Extensive evaluations on the VRSBench and EarthVQA datasets demonstrate that HeatTok effectively preserves object-level semantic integrity and achieves state-of-the-art performance under a reasonable token budget. The code is available: https://github.com/YingyingYan1/HeatTok.

Abstract:
Open-vocabulary semantic segmentation in remote sensing images is challenged by scale ambiguity. Although higher-resolution views can partially recover thin, safety-critical structures (e.g., roads and damage traces), they often amplify semantic uncertainty in repetitive textures, yielding noisy predictions, a counter intuitive Scale-Uncertainty Paradox. We introduce GUMix-RS, a training-free, zero-shot framework that resolves this via a pixel-wise geometryuncertainty mixture-of-scales, mimicking human perceptual strategies without target domain adaptation. Built on a frozen GeoRSCLIP backbone, we exploit SAM2 masks to derive a pixel-wise geometric complexity prior. The prior guides geometry-aware fusion of multi-level DINOv3-SAT features for GeoRSCLIP correlation refinement. Together with predictive entropy, it yields an analytic energy formulation that arbitrates scale selection through pixel-wise hard or soft routing. On 12 benchmarks, GUMix-RS sets new zero-shot state-of-the-art, boosting average mIoU on eight multi-class datasets from 35.78 to 41.97 (+17.3% rel.). Code is publicly available at: https://github.com/akuan1234/gumix_rs.

Abstract:
Despite yielding higher visual quality than image-to-image approaches, masked generation paradigms for video face editing fundamentally lack attribute consistency (e.g., illumination, background). We introduce LumaID, a novel framework that explicitly disentangles identity and expression representations from environmental contexts, enabling high-fidelity, fine-grained video head editing while strictly preserving these crucial attributes. At its core, LumaID employs an Omni-Disentangled Diffusion Transformer (OD-DiT) that leverages 3D proxy representations to thoroughly isolate the source and target facial features, fundamentally preventing identity leakage and illumination degradation. To further overcome the distributional drift caused by proxy estimation noise and the lack of explicit consistency supervision, we propose Consist-GRPO. This post-training reinforcement learning mechanism formulates multi-dimensional reward signals (spanning identity, expression, pose, and lighting) to continuously steer the generative process toward strict spatiotemporal alignment. Extensive evaluations demonstrate that LumaID serves as a highly competitive baseline, exhibiting strong performance over prior approaches in both attribute consistency and overall visual quality. Project page: https://wyy27.github.io/LumaID/.

Abstract:
We introduce KooCity, an initial large-scale real-world dataset with two large scenes to advance omnispace perception research. Existing datasets are limited to a single viewpoint, task or are synthetic, hindering progress in applications like autonomous driving and augmented reality.KooCity integrates aerial and ground-view imagery with LiDAR, accompanied by a new automated pipeline for precise cross-view alignment. This provides accurate six Degree-of-Freedom (6-DoF) poses and 3D ground truth, supporting diverse tasks such as cross-view retrieval, matching, 6-DoF localization, neural rendering, and 3D reconstruction. Benchmarks of state-of-the-art methods reveal significant challenges, establishing KooCity as a foundational benchmark to drive future research. Dataset is available \hrefhttps://github.com/Liumouliu/KooCity\colorred here.

Abstract:
We propose FlowAnchor, a training-free framework for stable and efficient inversion-free, flow-based video editing. Inversion-free editing methods have recently shown impressive efficiency and structure preservation in images by directly steering the sampling trajectory with an editing signal. However, extending this paradigm to videos remains challenging, often failing in multi-object scenes or with increased frame counts. We identify the root cause as the instability of the editing signal in high-dimensional video latent spaces, which arises from imprecise spatial localization and length-induced magnitude attenuation. To overcome this challenge, FlowAnchor explicitly anchors both where to edit and how strongly to edit. It introduces Spatial-aware Attention Refinement, which enforces consistent alignment between textual guidance and spatial regions, and Adaptive Magnitude Modulation, which adaptively preserves sufficient editing strength. Together, these mechanisms stabilize the editing signal and guide the flow-based evolution toward the desired target distribution. Extensive experiments demonstrate that FlowAnchor achieves more faithful, temporally coherent, and computationally efficient video editing across challenging multi-object and fast-motion scenarios. The project page is available at https://cuc-mipg.github.io/FlowAnchor.github.io/.

Abstract:
Multimodal large language models (MLLMs) are rapidly expanding from general video understanding to finer-grained understanding such as spatio-temporal video grounding (STVG) and reasoning. In these tasks, an MLLM must localize the user-queried target in time and space and take the results as evidence for reasoning. Existing MLLM methods mainly follow two paradigms. ? Direct Localization, which outputs STVG results with extra alignment modules or specialized decoders. ? Candidate-based Selection, which first constructs tube-level candidates and then selects the relevant one using an MLLM. However, both suffer from a serious efficiency bottleneck: the former incurs decoding costs that grow linearly with the queried temporal span, while the latter relies on costly candidate construction. To break this bottleneck, we propose DEViL, a detector-empowered Video-LLM with a simple key idea: offloading dense spatial grounding from the MLLM to a fully parallelizable, well-trained detector. Specifically, DEViL distills the query into a detector-compatible reference-semantic token, which replaces the detectors text embedding to enable spatial grounding in a single pass. Then, we design temporal consistency regularization to match objects across frames and enforce their coherence over time. In this way, DEViL avoids long coordinate decoding and heavy candidate pipelines. Extensive experiments show that DEViL achieves strong performance (43.1% m_vIoU on HC-STVG v1) with superior efficiency (14.33 FPS), while preserving the general reasoning capacity of the MLLM backbone. The code is available at https://github.com/gaostar123/DeViL.

Abstract:
Tensorized anchor graph-based multi-view clustering (MVC) has shown promising performance on high-dimensional data. However, existing methods generally underexploit latent data structures, while conventional tensor low-rank constraints cannot simultaneously capture intra-view smoothness and inter-view high-order correlations. To address these limitations, we propose SMVCAL, a smooth multi-view clustering framework that jointly performs unified anchor representation learning and latent representation learning. This framework captures underlying structures and alleviates information loss caused by limited anchors. Furthermore, we introduce a tensor total variation Schatten-p regularizer to jointly promote tensor low-rankness and intra-view smoothness, thereby preserving high-order cross-view correlations and local geometric structures. An efficient optimization algorithm is developed for the resulting problem. Extensive experiments on benchmark datasets demonstrate the effectiveness and computational efficiency of SMVCAL. The code is publicly available at: https://github.com/xd-wyp/SMVCAL.

Abstract:
Understanding visual attention is fundamental to video analysis, as it reveals where viewers focus and provides cues for interpreting dynamic scenes. However, existing eye-tracking benchmarks primarily provide saliency maps as supervision, lacking explicit signals that explain why attention is drawn to specific regions. This limitation hinders the development of explainable video saliency models. To address this gap, we present SGI-ADV (Subjective Gazing Indicators for Attention-Diverse Videos), a large-scale benchmark for learning video attention from reasoning-annotated eye-tracking supervision. SGI-ADV contains eye-tracking data collected from 4,600 participants, along with multi-frame gaze annotations paired with textual explanations that associate attention patterns with salient visual content and temporal dynamics. Building upon this dataset, we propose SalDAR (Saliency Decoding with Attribution and Reasoning), a novel framework that integrates a reason-guided visual decoder into multimodal large language models (MLLMs). SalDAR jointly predicts pixel-level saliency maps and generates human-understandable rationales from multi-frame video inputs, enabling unified modeling of where attention occurs and how it can be explained. Extensive experiments on SGI-ADV demonstrate that SalDAR consistently outperforms existing video saliency methods. Furthermore, evaluations on two widely used public benchmarks, DHF1K and MVS, verify its strong generalization capability. Codes and dataset are available at https://github.com/daisysunset/SalDAR.

Abstract:
Federated multi-view clustering (FMVC) enables collaborative data grouping across distributed silos while preserving privacy. However, existing FMVC methods typically rely on statistical correlations between observed features and cluster labels, making them vulnerable to spurious correlations caused by hidden confounders (e.g., selection bias), which are often amplified by the non-IID nature of federated views. In this paper, we propose \methodname (\underlineSpurious-\underlineFree \underlineFederated \underlineMulti-View \underlineClustering), the first framework to incorporate causal inference into FMVC. By leveraging a structural causal model (SCM), we identify the ``backdoor paths'' that mislead clustering. We then develop a privacy-preserving confounder balancing mechanism that learns optimal sample weights to minimize distributional discrepancies, effectively isolating causal features (e.g., essential semantic attributes) from non-causal noise without exchanging raw data. Integrated into a communication-efficient matrix factorization framework, \methodname eliminates spurious associations while maintaining low bandwidth overhead. Extensive experiments on eight real-world datasets demonstrate that \methodname significantly outperforms existing state-of-the-art baselines. The source code is available at \urlhttps://github.com/Xianjie-Guo/SFFedMC.

Abstract:
Artificial intelligenceassisted ultrasound scanning enhances diagnostic reliability and efficiency by providing real-time guidance for standardized image acquisition and reducing operator dependence. However, existing reinforcement learning and learning-assisted ultrasound scanning methods typically rely on carefully designed reward functions or extensive interaction data, which limits their generalization ability and stability across different devices, patient populations, and complex clinical scenarios. To address these challenges, we propose an ultrasound vision-language-action model (US-VLA) for automated ultrasound scanning that explicitly encodes clinical semantic goals and generates sequential probe manipulation actions under real-time ultrasound feedback. In particular, we first design an ultrasound-aware expert fusion module to jointly integrate ultrasound observations with auxiliary contextual information, enabling semantic ultrasound feedback to effectively guide the scanning process. Then, we construct US-VLA-Data, a real-world dataset covering liver and kidney examinations, which includes five clinically defined standard planes and comprises 320 expert scanning trajectories with approximately 80,000 synchronized timesteps. Extensive experiments demonstrate that US-VLA achieves competitive performance in ultrasound probe manipulation tasks, indicating its effectiveness and promising generalization within the evaluated abdominal ultrasound setting. The source code is available at https://github.com/VMVLab/US-VLA.

Abstract:
Low-resolution medical volumes can lead to missed diagnoses, making medical volumetric super-resolution (MVSR) essential for reliable clinical decision-making. However, real-world scenarios, such as multi-center studies and large-scale medical emergencies, require rapid model development, whereas traditional MVSR methods rely on time-consuming training procedures, thereby increasing development time and delaying deployment. In this paper, we present ESA-one, a foundation model-assisted MVSR framework that significantly accelerates training. The key idea is to leverage the strong shape representation capability of the Segment Anything Model (SAM) to guide early-stage reconstruction, thereby avoiding inefficient shape learning from low-resolution inputs. In addition, tri-axial frequency features are introduced to enhance texture refinement, and a shape-prior distillation strategy enables the removal of SAM during inference for efficient deployment. Experiments demonstrate that ESA-one achieves high-resolution synthesis in only one training epoch73 seconds on an NVIDIA A800 or 95 seconds on an NVIDIA RTX 3090outperforming state-of-the-art methods by 10.06 dB at 2 upsampling on the T2 modality. Our code is available at: https://github.com/caoluyang0830/ESA-one.git.

Abstract:
General image fusion aims to integrate complementary information from different modalities. We observe that existing general fusion methods, despite sharing encoders to extract common features, perform well only on some tasks while struggling on others. We attribute this phenomenon to two interrelated key factors: distribution confusion caused by lack of explicit constraints on the encoder and task characteristic imbalance for general fusion methods. Inspired by the recent Contrastive Language-Image Pre-Training model CLIP, this paper discovers the task awareness of CLIP that the feature clusters of different tasks extracted from the frozen CLIP ResNet image encoder are naturally separable in unified space. The task-aware property of CLIP compensates for the task confusion of feature distributions in existing general-purpose encoders. Based on above observations, we propose a general image fusion network based on encoder-decoder architecture, called CLIPFuse, which integrates dense features extracted by the frozen CLIP ResNet Encoder into learnable decoder. In the decoder design, to fully utilize differences in task distributions, we propose a novel Task-Adaptive Bank Module(TABM), aiming to store the common features of each task and achieve differentiated fusion utilizing prior-separable distributions from CLIP. Extensive experiments demonstrate the superiority of our method on IVIF, MFF, MEF and PAN tasks. Code is available at https://github.com/liuuuuu777/CLIPFuse.

Abstract:
Image restoration seeks to recover clean images from degraded inputs. While significant progress has been made on natural images, restoration of fine-structured images remains underexplored. In such images, even small local errors can cause severe structural distortions and adversely affect downstream analysis. To address this gap, we introduce CurveBench, a benchmark for fine-structured image restoration with style-consistent reconstruction. CurveBench comprises 10,453 paired samples derived from real industrial well-logging blueprints, featuring dense, overlapping curves and authentic degradations accumulated during long-term storage and scanning. These characteristics necessitate simultaneous degradation removal, structural preservation, and style-consistent reconstruction. To provide a strong baseline, we propose Structure-Cycle, an end-to-end framework that integrates restoration and style translation within a unified latent space, augmented with explicit structural priors and consistency constraints. Extensive experiments demonstrate that existing restoration, style transfer, and two-stage cascade methods struggle with the composite challenges posed by CurveBench. By offering realistic fine-structured data and a standardized evaluation protocol, CurveBench serves as a valuable testbed for future research in structure-sensitive image restoration.\footnoteThe dataset and source code are available at \urlhttps://github.com/zyzzzz777/CurveBench.

Abstract:
\beginabstract Thermal infrared novel view synthesis (TIR-NVS) is important for nighttime robotics, infrastructure inspection, aerial monitoring, and adverse-weather perception. However, sparse-view TIR-NVS remains highly under-constrained. Existing thermal reconstruction methods often become unstable under sparse views, while directly transferring RGB-centric diffusion priors to TIR is also problematic because TIR synthesis requires radiometric faithfulness rather than appearance-driven edits. We propose ThermalDiffusion, a two-stage diffusion framework for sparse-view TIR-NVS. To bridge the domain gap between RGB-centric video diffusion and thermal generation, we construct filtered pseudo thermal trajectories using multiple nearby-view synthesis strategies, and combine them with real thermal trajectory clips to adapt a pretrained latent video diffusion backbone to the TIR domain via lightweight low-rank adaptation (LoRA) tuning. At inference time, we reconstruct a coarse 3D Gaussian Splatting (3DGS) proxy from the sparse inputs and derive trajectory-aligned controls, including thermal intensity, depth, and a reliability confidence map. These controls enable confidence-aware dual guidance during reverse diffusion: depth-aware screened-Poisson guidance regularizes uncertain regions, while radiometric evidence anchoring preserves reliable thermal evidence in high-confidence regions. The generated trajectory is further distilled into a refined 3DGS representation for efficient free-view rendering. Experiments on TI-NSD show that ThermalDiffusion consistently outperforms the strongest baseline under sparse-view settings, improving PSNR by up to 2.30 dB while also achieving better SSIM and LPIPS. The pseudo dataset and code will be released at \urlhttps://github.com/JingyuanGao0919/ThermalDiffusion. \endabstract

Abstract:
Text-to-image (T2I) generation has achieved remarkable progress in recent years. However, existing research has largely focused on English-only settings, leaving cross-lingual performance gaps and language-specific effects insufficiently explored. To fill this gap, we introduce LingT2I, a benchmark covering 10 widely used languages with 33K prompts, designed to evaluate cross-lingual effects in both content generation and text rendering. Building on this benchmark, we conduct a comprehensive cross-lingual analysis, uncovering linguistic inequality and language-dependent trade-offs across evaluation dimensions. Beyond quantitative evaluation, we further reveal a range of language-dependent generation patterns, highlighting how linguistic factors and their corresponding cultural contexts systematically impact model outputs. Our benchmark and analysis provide a foundation for studying cross-lingual behavior in T2I generation and facilitate the development of more robust and inclusive models. Code and dataset are available at https://github.com/RISys-Lab/LingT2I.

Abstract:
Multi-modal multi-label emotion recognition (MMER) aims to identify multiple co-occurring emotions from text, audio, and visual modalities. However, because emotional expressions are often subtle, asynchronous, and distributed across modalities, effectively capturing complementary cues across different temporal granularities within each modality and across modalities is crucial for accurate MMER. Existing methods typically rely on single-granularity intra-modal representations or limited cross-modal fusion, which restricts their ability to fully exploit these complementary signals. To address these issues, we propose M3-Fusion, a framework that jointly models complementarity across temporal granularities within each modality and inter-modal progressive complementarity through two stages. First, Temporal Granularity-Adaptive Fusion (TGAF) captures temporal patterns at multiple granularities within each modality and aggregates them in a time-step-wise adaptive manner, so that each temporal position can dynamically focus on the granularity that best matches its emotional expression. Second, Multi-Round Modality Communication (MRMC) iteratively refines cross-modal representations through shared/specific decomposition and gated updates, enabling progressive alignment of cross-modal consensus while preserving modality-specific cues. Experiments on CMU-MOSEI and MED show that M3-Fusion achieves state-of-the-art among reproduced baselines across multiple metrics. The ablation studies further confirm that both temporal granularity-adaptive fusion and iterative inter-modal communication contribute to the performance gains. The code is available at https://github.com/JianZhang-chick/M3-fusion.

Abstract:
Human-interpretable computer-aided diagnosis is crucial for clinical decision making. Concept-based models excel by providing transparent reasoning and enabling post-hoc, clinician-in-the-loop interventions. However, their rigid dataset-specific adaptation inherently restricts cross-site generalization. Applying them across diverse modalities, such as dermoscopic and clinical photographs, is challenging due to heterogeneous concept taxonomies varying in availability, granularity, and semantics across cohorts. Consequently, adapting Foundation Vision-Language Models (FVLMs) demands costly label engineering and repeated post-training. Existing intervention mechanisms remain rigidly tied to predefined concepts, lacking adaptability and hindering scalable dermatology CAD deployment. To address these bottlenecks, we propose UniCon, an open-linguistic unified concept learning framework for multimodal interpretable vision-language diagnosis. UniCon resolves these challenges through three contributions: (1) A shared semantic representation space via a unified concept prototype codebook, seamlessly coordinating heterogeneous concept systems across modalities without dataset-specific retraining. (2) Open-linguistic based multi-faceted semantic specifications to overcome sparse textual label limitations, improving boundary sensitivity in uncertain clinical contexts. (3) A robust, cross-site adjustable intervention interface powered by reliability-gated bottleneck aggregation, enabling consistent reasoning and transferable clinician corrections. Extensive experiments demonstrate that beyond securing top-tier diagnostic accuracy, UniCon successfully bridges disparate clinical taxonomies, unlocking unprecedented cross-site intervention capabilities. Code is available at https://github.com/wuchengyu123/UniCon.

Abstract:
Multi-modal learning combining medical images and clinical text is promising for disease diagnosis. However, standard multi-modal training leads to shortcut learning: models exploit the easier modality (e.g., diagnostic cues in text) while neglecting harder-to-learn features (e.g., subtle visual patterns). We propose UniMod, a framework that mitigates shortcut learning by requiring each modality to predict the diagnosis on its own. It supervises image-only, text-only, and multi-modal classification simultaneously, so each modality must extract diagnostic features. We add cross-modality alignment for knowledge transfer and within-modality supervised contrastive alignment over same-diagnosis patients. On Harvard-Glaucoma, UniMod reaches 0.850 AUC, outperforming OGM-GE and Gradient Blending by 1.6-1.8%; on CheXpert Plus, it reaches 0.966 AUC, surpassing them by over 5%. UniMod also extends to 5-class multi-label diagnosis without architectural change, improving mean AUC by 0.097 over CGGM.

Abstract:
Knowledge-Based Visual Question Answering (KB-VQA) relies on retrieving external information to answer queries involving long-tail entities. However, existing retrieval pipelines predominantly employ CLIP-style dual encoders, which prioritize surface-level visual similarity over entity-level semantic alignment. This paradigm often fails when semantically identical concepts exhibit large visual variations or when distinct entities appear visually similar. To address this, we propose KBMR, the first MLLM-based embedding retriever tailored for KB-VQA. Leveraging the robust autoregressive capabilities of MLLMs, KBMR maps images into a semantic space that better preserves concept identity. To tackle the challenge of noisy supervision in Wikipedia-scale retrieval, we introduce an MLLM-based semantic discriminator that generates continuous entity-consistency weights. These weights guide a novel continuous semantic distillation objective, enabling effective hard negative sampling and soft supervision beyond rigid binary labels. Extensive experiments demonstrate that KBMR significantly outperforms CLIP baselines, yielding up to a 14.7% improvement in retrieval Recall@1 and a 9.4% gain in end-to-end VQA accuracy. Code is available at https://github.com/realHarryX/KBMR.

Abstract:
Fine-grained incremental remote sensing object detection requires models to learn novel fine-grained targets while avoiding catastrophic forgetting of prior knowledge. However, when directly migrating existing class-incremental object detection paradigms from natural images to the remote sensing domain, the wide-swath characteristic of remote sensing imagery and high visual similarity of fine-grained targets cause past, current, and future classes to densely co-occur with complex backgrounds in the same scene, leading to severe mutual confusion. Consequently, this forms a tight coupling between complex backgrounds and cross-temporal classes. Failure to effectively decouple them results in pseudo-label collapse, feature contamination, and background shift. To address these challenges, we propose the Internal-External Collaborative Decoupling (IECD) method, which innovatively synergizes the historical memory of an internal teacher model with the physical prior of an external Vision-Language Model (VLM). Specifically: (1) Internal Memory Modeling (IMM) fits class-specific distributions to reconstruct internal memory, decoupling old-class targets from complex backgrounds to generate high-quality pseudo-labels; (2) Collaborative Frequency Distillation (CFD) leverages VLM-driven external physical priors and the teacher model's internal memory of old classes to suppress spatial neighborhood interference in the frequency domain, preserving the discriminative fine-grained features of old classes; (3) Synergistic Future Penalty Attenuation (SFPA) cross-verifies internal and external perceptions to decouple latent future targets from complex backgrounds, mitigating background shift. Extensive experiments demonstrate that our method achieves state-of-the-art performance on the FAIR1M and ShipRSImageNet datasets, overcoming performance bottlenecks in fine-grained incremental remote sensing object detection. Our code is available at \hrefhttps://github.com/zhangdoeg/IECDhttps://github.com/zhangdoeg/IECD.

Abstract:
Monocular RGB-based hand pose estimation has emerged as a critical research frontier in computer vision. The local hand pose estimation methods predict hand poses relative to the wrist, while global hand pose estimation also requires estimating the wrist's position in the camera coordinate system. However, this camera-space estimation confronts two fundamental challenges: (1) depth ambiguity in monocular settings, and (2) the coupling effect of hand local poses and global wrist positions in the perspective projections. In particular, this coupling reflects that the projections are jointly determined by local hand poses, wrist positions, and camera intrinsics. To overcome these challenges, our framework proposes two key innovations: Transformation-Isomorphism Supervision for hand-depth information extraction and Perspective Information Embedding for resolving above coupling effect of local pose and wrist position, both integrated within the mainstream encoder-decoder architecture. Besides, we propose a novel framerate-aware multi-dataset training strategy for sequential pose refinement. Our fully integrated approach achieves at most 37.1% superiority in CS-MJE over SOTA on HO3D. Project page: https://github.com/Mine268/CS-ViT.

Authors: Xianhui Meng (University of Science and Technology of China), Yuchen Zhang (Georgia Institute of Technology), Zhijian Huang (Xiaomi EV), Zheng Lu (Xiaomi EV), Ziling Ji (Fudan University), Yandan Lin (Fudan University), Yaoyao Yin (Xidian University), Hongyuan Zhang (University of Hong Kong), Wei Zhou (Cardiff University), Guangfeng Jiang (University of Science and Technology of China), Li Zhang (University of Science and Technology of China), Long Chen (Xiaomi EV), Hangjun Ye (Xiaomi EV), Jun Liu (University of Science and Technology of China), Xiaoshuai Hao (Xiaomi EV)

Abstract:
Driving safety in assisted driving systems requires holistic understanding of both the external traffic environment and the internal driver state. However, existing research typically treats out-cabin scene understanding and in-cabin driver monitoring as separate problems, resulting in fragmented safety assessment and limited support for real-world warning applications. Moreover, conventional task-specific models often perform well on narrowly defined tasks but struggle to generalize to the diverse and long-tail hazards encountered in open-world driving scenarios. To address these limitations, we propose a new research setting, termed \emphdriving safety understanding for assisted driving, which aims to jointly perceive, reason about, and warn against risks by integrating information from both inside and outside the vehicle. To support this task, we construct a large-scale driving safety dataset containing 98K vision-language annotations that cover diverse in-cabin and out-cabin safety-critical scenarios, and further curate a 3K benchmark named DSBench(Driving Safety Benchmark) set for systematic evaluation. Based on this dataset, we develop DSVLM, a domain-specialized vision-language model for driving safety. Unlike end-to-end assisted driving models, DSVLM is designed as an in-vehicle safety assistant that identifies risk factors, infers potential danger, and provides timely warnings to help prevent accidents. Extensive evaluations against a wide range of mainstream open-source and proprietary VLMs show that existing models suffer substantial performance degradation in complex safety-critical situations, revealing a clear gap in reliable driving-safety understanding. In contrast, DSVLM establishes a new state of the art, achieving an average score of 68.4, which surpasses the strongest commercial baseline, Seed-1.6 (49.52), by 18.88 points. Its superiority is particularly pronounced on highly challenging tasks such as cockpit understanding, where DSVLM attains 80.1, dramatically outperforming the runner-up score of 29.49. These results demonstrate that DSVLM not only improves overall performance, but also delivers markedly stronger robustness and safety-awareness in complex real-world driving scenarios. The benchmark toolkit, source code, and model checkpoints are publicly available at https://github.com/mengxh20/DSBench.

Abstract:
Autonomous biomedical laboratories increasingly rely on visual perception to recognize, localize, and manipulate transparent plasticware, yet high-quality real-world datasets for this setting remain limited. The scarcity of domain-relevant data is particularly restrictive in cluttered multi-object scenes, where mutual occlusion and view-dependent appearance changes remain challenging even for contemporary visual foundation models. Existing transparent-object datasets have advanced segmentation, depth, and pose estimation, but they usually do not evaluate the combined setting of multi-object clutter, occlusion, and calibrated multi-view capture that characterizes real laboratory manipulation scenes. To address this gap, we present TransBiolab, a real-world RGB-D dataset of cluttered transparent biomedical objects captured as calibrated multi-view sequences. TransBiolab contains 161,315 frames from 98 scenes and 1.03M instance annotations over 15 laboratory object types, including 6D poses, full and visible masks, depth, and per-frame camera calibration. The dataset is organized along three axes that reflect operational difficulty: object category, the total number of objects in a frame, and camera viewpoint. We further define dataset-centric benchmarks for segmentation, depth estimation and completion, and 6D pose estimation, and report a system-level robot manipulation evaluation enabled by the released annotations and calibrations. By focusing on repeated transparent instances, clutter, and multi-view laboratory capture, TransBiolab provides a resource for segmentation, depth estimation, 6D pose estimation, and multi-view reasoning in autonomous laboratory manipulation. Project page: https://dualtransparency.github.io/TransBiolab/.

Abstract:
Most research on hallucinations in Large Vision-Language Models (LVLMs) focuses on factual description tasks that prohibit any output absent from the image. However, little attention has been paid to hallucinations in voluntary imagination tasks, such as story writing, despite this human-like cognitive ability being essential for real-world generative applications. To address this limitation, we introduce Voluntary-imagined Object Presence Evaluation (VOPE\footnoteOur benchmark is available at https://github.com/qqwsad5/VOPE.)---a recheck-based evaluation benchmark for assessing LVLMs' grounding behavior in voluntary imagination tasks. Specifically, VOPE poses recheck-based questions to evaluate how an LVLM interprets the presence of the imagined objects in its own response. Rather than penalizing the imagined content itself, VOPE identifies hallucinations based on the correctness of the model's presence judgments for the generated objects. Built on this idea, we construct a dataset covering captioning, reasoning, and writing tasks with different levels of voluntary imagination. We apply VOPE to several mainstream LVLMs and hallucination mitigation methods, revealing two key findings: (1) most LVLMs hallucinate heavily during voluntary imagination, and their performance in presence evaluation is notably poor on imagined objects; (2) existing hallucination mitigation methods show limited effect in voluntary imagination tasks, making this an important direction for future research.

Abstract:
Computer-use agents extend language models from text generation to persistent action over tools, files, and execution environments. Unlike chat systems, they maintain state across interactions and translate intermediate outputs into concrete actions. This creates a distinct safety challenge: harmful behavior may emerge through sequences of individually plausible steps that collectively lead to unauthorized operations, sensitive data access, or destructive execution. We present AgentHazard, a benchmark for evaluating harmful behavior in computer-use agents. AgentHazard contains 2,653 instances spanning 10 risk categories and 10 attack strategies. Each instance pairs a harmful objective with a sequence of operational steps that appear locally legitimate but jointly induce unsafe behavior. The benchmark evaluates whether agents can recognize and interrupt harm arising from accumulated context and repeated tool use across multi-step trajectories. We evaluate AgentHazard on Claude Code, OpenClaw, and IFlow using diverse open and openly deployable models from the Qwen, Kimi, and GLM families. Our experimental results indicate that current systems remain highly vulnerable: when powered by Qwen3-Coder, Claude Code exhibits an attack success rate of 73.63%, suggesting that model-level alignment does not reliably guarantee agent-level safety. To facilitate future research, we open-source the dataset, supplementary materials, and benchmark construction pipeline at \urlhttps://yunhao-feng.github.io/AgentHazard/.

Abstract:
We introduce RDVSv2, a large-scale benchmark for RGB-D video salient object detection (RGB-D VSOD) with dense frame-level annotations. Existing datasets in this emerging field are often limited in scale and annotation quality, while also relying on less geometry-consistent depth cues. To address these limitations, RDVSv2 is built from publicly accessible stereoscopic online videos and contains 249 video sequences with 29,077 annotated frames. It includes depth maps derived from stereoscopic videos, together with frame-wise salient object masks annotated with eye-tracking guidance. Compared with existing datasets, RDVSv2 is much larger in scale and covers more diverse and challenging scenarios. In addition, we establish a strong baseline for RGB-D VSOD based on Segment Anything Model 2 (SAM2). Specifically, we employ a parameter-efficient fine-tuning (PEFT) strategy to adapt the SAM2 encoder to jointly encode RGB, depth, and optical flow cues. Extensive experiments show that RDVSv2 is substantially more challenging for existing RGB-D VSOD methods. Meanwhile, the proposed baseline achieves state-of-the-art results on RDVSv2 and existing RGB-D VSOD benchmarks. We hope that RDVSv2 and the provided baseline will serve as useful resources for future research on RGB-D VSOD and related multi-modal video understanding tasks. Our dataset and code are available at https://github.com/ltynick/RDVSv2.

Abstract:
While multimodal large language models (MLLMs) excel in medical applications, most of them favor static images or short-term signals. In the critical field of dynamic electrocardiograms (ECG), models struggle with complex temporal reasoning and diagnostic report generation due to a lack of high-quality datasets and benchmarks. To address this, we introduce (i) Holtercare-23K, a large-scale multimodal dynamic ECG dataset comprising 22,980 QA pairs derived from 788 clinical Holter records and featuring a novel signal--video--text tri-modal alignment. Based on this dataset, we present (ii) Holtercare-Bench, a multimodal benchmark that evaluates models on temporal localization, clinical diagnosis, and global summarization. Zero-shot evaluations of leading MLLMs reveal a significant performance gap in processing ultra-long pathological sequences. However, fine-tuning representative models yields substantial improvements. This work illuminates the limitations of current MLLMs in electrophysiology and provides a foundational benchmark for long-term medical MLLMs. Our project is available at \urlhttps://github.com/ZJU4HealthCare/Holtercare-Bench.

Abstract:
4D Gaussian Splatting (4DGS) has recently become a leading representation for photorealistic dynamic scene rendering. However, streaming 4DGS in real-time presents a fundamental trade-off between visual fidelity and data volume. Existing methods suffer from a decoupled training-compression pipeline and employ coarse-grained, uniform bitrate allocation across Gaussian attributes during compression. This leads to delivery inefficiency, yielding suboptimal rate-distortion performance. To bridge this gap, we propose PrismGS, a coordinated training and compression framework for 4D Gaussian streaming that realizes compression-aware dynamic scene reconstruction and attribute-level fine-grained bitrate allocation. First, we propose keyframe regularization constraining Gaussian scale and anisotropy during initialization to mitigate error propagation in dynamic training and quantization-induced artifacts in later compression stage. We then adopt a residual-based dynamic training pipeline to ensure temporal coherence and high compressibility. Finally, we propose a dynamic programming algorithm to identify the Pareto-optimal bitrate allocation by leveraging the heterogeneous sensitivity of Gaussian attributes to compression. In addition, we introduce a parallel decoding scheme in client side, significantly reducing decoding latency. Extensive experiments demonstrate that PrismGS outperforms existing methods by up to 1.3 dB in PSNR, 30% size reduction, 15% faster rendering speed and 4 times faster decoding speed in various datasets. Code and additional material are available at \hrefhttps://touringtrain.github.io/PrismGS/https://touringtrain.github.io/PrismGS/.

Abstract:
\beginabstract Multimodal Large Language Models (MLLMs) encounter significant computational and memory bottlenecks from the massive number of visual tokens generated by high-resolution images or multi-image inputs. Previous token compression techniques are often constrained by heuristic rules that risk discarding critical information. They may suffer from biases, such as attention sinks, that lead to sharp performance drops under aggressive compression ratios. To address these limitations, we reformulate token compression as a lightweight plug-and-play framework and turn it into an end-to-end learnable decision process. Specifically, we propose VisionSelector, a scorer module decoupled from the MLLM backbone that incorporates a differentiable Top-K mechanism and a curriculum annealing strategy to bridge the traininginference gap, enabling efficient and adaptive token selection across various compression rates. Remarkably lightweight with only 12.85M trainable parameters, VisionSelector demonstrates generalization across various compression rates and adaptively identifies critical tokens. This leads to superior performance across the evaluated compression budgets, evidenced by preserving 100% performance on MME with a 30% retention budget, outperforming representative heuristic baselines by 12.14 percentage points at a 10% retention budget, and doubling prefill speed. Our code is available at \urlhttps://github.com/JulietChoo/VisionSelector. \endabstract

Abstract:
Extending vision-language models (VLMs) to long-form video understanding is bottlenecked by the massive volume of visual tokens that incur quadratic prefill cost in the LLM decoder. Existing compression methods either discard tokens before the question is known (query-agnostic) or passively observe post-hoc attention patterns (passive query-aware), risking premature loss of question-relevant details. We propose Question-in-the-Loop (QitL), which actively interrogates the visual stream by structurally interleaving the actual text query tokens between temporal chunks of video frames inside the LLM. Each query copy uses causal self-attention but is visually restricted to its assigned chunk, simultaneously serving three roles: a pruning guide that scores visual token relevance, a local attention sink that mitigates attention dilution in long sequences, and a persistent summary that carries compressed, question-conditioned information into deeper layers. A lightweight prune-merge step then removes low-relevance tokens while folding residual information into the summaries. We further provide empirical evidence that shallow-layer attention is semantically uncalibrated for pruning and that local masking accelerates the emergence of meaningful cross-modal attention, which motivates our aggregation boundary design. Entirely training-free, QitL requires no learnable parameters or weight modifications. Experiments on Video-MME, MLVU, LongVideoBench, and MVBench demonstrate competitive accuracyefficiency trade-offs among the reported training-free baselines. Code will be released at https://github.com/NihiI-obstat/QitL.

Abstract:
\beginabstract Multimodal Emotion Recognition in Conversation (MERC) aims to identify speakers' emotions by integrating conversational context with textual, acoustic, and visual cues. Existing methods remain limited in capturing emotion-relevant contextual dependencies and modeling cross-modal discrepancies during multimodal fusion. To address these challenges, we propose a Conflict-Aware Dual-Path Context Propagation framework (CDCP) for MERC. Specifically, CDCP first constructs dynamic semantic neighborhoods for each utterance via Adaptive Neighbor Selection, and then models contextual dependencies through speaker-aware Dual-Path Dynamic Multi-Hop Propagation to learn higher-order unimodal representations. To better handle cross-modal discrepancies, we further design a Conflict-Aware Expert Reconstruction module. It decomposes modality features into high- and low-frequency components in the frequency-domain, characterizes cross-modal consensus and modality-specific discrepancies through component interaction and fusion, captures potential cross-modal conflicts, and adaptively reconstructs emotion representations. Experiments on MELD and IEMOCAP demonstrate the effectiveness of CDCP, and further analyses validate the design of the proposed framework. Our code is available at \urlhttps://github.com/MKMaS-GUET/CDCP. \endabstract

Abstract:
3D Gaussian Splatting achieves photorealistic reconstruction within training view distribution, yet it degrades on out-of-distribution novel views, exhibiting holes in unobserved regions and artifacts in observable areas. Recent works formulat this task as extrapolation and interpolation and tried to address it with generative models, but remain limited in extrapolation scale and quality. They repeat a generatereconstructshift cycle to progressively build a scene, which introduces accumulated errors with every step conditioning on previous outcomes. In this work, we propose a holistic framework for extrapolation and interpolation. We devise an independent camera view detection mechanism to enable parallel conflict-free extrapolation, circumventing the reliance on aforementioned error-prone cycle. Building upon this, we design a hierarchical pipeline that extrapolates independent and dependent camera views separately. Additionally, previous methods overlook inconsistency between generated and original images, resulting in compromising well-reconstructed areas. We propose a plug-and-play Quality-Aware Mask (QA-Mask) module, enabling selective utilization on generated data. By calibrating learning weights with pixel-wise rendering quality, it prevents generation-induced degradations on well-constructed areas. Extensive experiments demonstrate the superior performance of our framework, with QA-Mask generalizing on multiple generative reconstruction models (\hrefhttps://vulab-ai.github.io/filling-the-unseen/project page).

Abstract:
While Large Audio Language Models (LALMs) achieve strong performance on short audio, they degrade on long-form inputs. This degradation is more severe in temporal awareness tasks, where temporal alignment becomes increasingly inaccurate as audio duration grows. We attribute these limitations to the lack of data, benchmarks, and modeling approaches tailored for long-form temporal awareness. To bridge this gap, we first construct LAT-Chronicle, a 1.2k hour long-form audio dataset with temporal annotations across real-world scenarios. We further develop LAT-Bench, the first human-verified benchmark supporting audio up to 30 minutes while covering three core tasks: Dense Audio Caption, Temporal Audio Grounding, and Targeted Audio Caption. Leveraging these resources, we propose LAT-Audio, formulating temporal awareness as a progressive global-to-local reasoning paradigm. A global timeline is first constructed as an aligned temporal-semantic context, and the Think-With-Audio Chain-of-Thought (TWA-CoT) is then introduced to perform iterative reasoning by incorporating local audio information via tool use. Experiments show that LAT-Audio surpasses existing models on long-form audio temporal awareness tasks and improves robustness to input duration. We release the dataset, benchmark, and model to facilitate future research at https://github.com/alanshaoTT/LAT-Audio-Repo.

Abstract:
Remote sensing offers an unparalleled vantage point for observing the Earth's long-term surface evolution, yet it demands that a model not only perceive land cover at isolated moments, but also track changes, memorize evolution histories, and reason across time and space. However, existing studies lack a systematic evaluation that dissects these distinct competencies. To fill this gap, we introduce ChronoBench, a multidimensional benchmark that decomposes this task into four progressive cognitive levels (i.e., Land Cover Perception, Temporal Recognition, Long-Term Memory, and Spatio-Temporal Reasoning). The ChronoBench comprises 12 sub-tasks and 17,689 rigorously validated QA (Question-Answer) pairs. Extensive evaluations reveal that mainstream MLLMs fall drastically behind human experts, with Long-Term Memory emerging as the most critical bottleneck. Motivated by this finding, we further propose GeoChrono, an MLLM with enhanced capabilities for tracing, memorizing, and reasoning about long-term geographic evolution. Leveraging the physical prior that geographic parcels remain spatially fixed while their semantics evolve, we design a Temporal Trajectory Encoder~(TempEnc) that constructs per-location temporal trajectories for dedicated land cover evolution modeling, and we introduce a Coarse-to-Fine Token Compressor~(C2FComp) that adaptively preserves dynamic regions while compressing the static background. To support training, we also construct ChronoInstruct, a 104K-sample instruction-tuning dataset spanning all competency levels for training. GeoChrono achieves state-of-the-art performance on ChronoBench, surpassing the leading commercial MLLMs by over 20%, while C2FComp reduces visual tokens by over 56% while retaining GeoChrono's 94.6% performance. The code and data will be available at \urlhttps://github.com/IntelliSensing/GeoChrono.

Abstract:
Agricultural visual question answering (VQA) requires fine-grained reasoning to distinguish similar symptoms and recognize rare diseaseschallenges exacerbated by imbalanced training data. Recent work Agri-r1~\citezhang2026agri reveals that while Group Relative Policy Optimization (GRPO) enables open-ended reasoning, its direct application triggers frequency-induced class collapse: frequent classes dominate gradient updates, causing rare classes to be forgotten. Potential solutions such as frequency-aware reward weighting attempt to mitigate collapse during Reinforcement Learning (RL), yet these reactive approaches still face a fundamental issue: the initialization deficitrare classes enter RL with unstable, under-trained representations. We propose Agri-Cure, a novel framework that unifies curriculum learning across both SFT and RL. First, our proposed curriculum learning paradigm captures sample difficulty, enabling progressive training from easy to hard to establish stable representations before RL. Second, tier-dependent reward weighting adapts optimization objectives to sample difficulty: format compliance for easy samples, reasoning quality for hard samples.Evaluated on CDDMBench, our method consistently outperforms existing GRPO-based approaches, achieving substantial gains in rare-class recognition while establishing an effective paradigm for RL-based VLM adaptation. Our code and data are available at \textcolorblue\urlhttps://github.com/CPJ-Agricultural/Agri-Cure.

Abstract:
Despite rapid advances in generative models, achieving pixel-level precision in sketch-based image editing remains a persistent challenge, particularly for fine-grained local deformations. This gap stems primarily from the critical shortage of high-quality, publicly available benchmark datasets that jointly provide geometric constraints and semantic instructions. To address this issue, we first introduce SI-Data, a high-quality dataset specifically designed for instruction-guided local sketch editing. We develop an automated pipeline leveraging Multimodal Large Language Models (MLLMs) to synthesize comprehensive quadruplets comprising original images, local geometric sketches, semantic instructions, and corresponding edited images. By providing both reliable spatial anchors and explicit semantic intent, SI-Data uniquely enables collaborative spatial-semantic learning. Building upon this, we propose a collaborative framework called SI-Edit that integrates semantic instructions with precise geometric constraints. Furthermore, to address the lack of standardized evaluation, we establish a comprehensive set of metrics designed to measure both structural fidelity (e.g., sketch-to-edge alignment) and semantic adherence. Experimental results demonstrate that SI-Edit provides more reliable structural control than baselines for sketch-based image editing, and achieves precise, pixel-level local refinements aligned with user intent. The data and code are released on the \hrefhttps://github.com/ywxsuperstar/SIEditproject page.

Abstract:
EEG emotion recognition typically follows the standard signal-to-label paradigm. It maps EEG signals to discrete labels via feature engineering and machine learning, yet neglects the rich semantic information inherent in the task. This purely data-driven approach easily overfits to subject-specific physiological patterns, severely degrading its generalization on new subjects. Here, we propose a Semantically Guided Domain Adaptation (SGDA) method for cross-subject EEG emotion recognition. Specifically, SGDA projects non-stationary EEG features into an emotional semantic space, leveraging the discriminative power of structured linguistic embeddings to enhance class separability. To construct this unified space, a pretrained language model is utilized to extract linguistic embeddings from emotion text labels, which serve as invariant semantic prototypes to guide cross-domain feature alignment. Moreover, to facilitate robust inference, an adaptive distribution-aware feature fusion strategy dynamically aggregates multi-source knowledge based on the distances between target samples and source centroids. Extensive experiments on five benchmarks demonstrate that SGDA achieves state-of-the-art performance in cross-subject scenarios. The source code is publicly available at \urlhttps://github.com/EDH777/SGDA.

Abstract:
Recent advancements in Speech Large Language Models have demonstrated remarkable capabilities in understanding complex audio tasks. Despite this progress, their long-context inference remains severely bottlenecked by prohibitive KV cache memory demands. Existing text-centric compression methods struggle here, often disrupting speech continuity or discarding crucial semantic cues. To address this, we propose VoxZip, a train-free, two-stage semantic-anchored KV cache compression framework. The first stage uses automatic speech recognition (ASR) transcriptions as explicit semantic anchors to temporally align, compress, and fuse audio tokens, significantly reducing the initial KV cache while elevating token information density. To further improve the compression ratio, the second stage employs a dynamic filtering strategy based on temporally decayed accumulated attention to evict non-essential tokens while mitigating early-token bias. Comprehensive evaluations on Qwen3-Omni across six diverse audio benchmarks demonstrate the superiority of our approach. VoxZip excels in long-audio reasoning and consistently maintains high-fidelity perception on short-form tasks. Notably, it sustains over 90% of the uncompressed baseline performance even under an aggressive 20x KV cache compression in long-context scenarios. Furthermore, at a 4x compression ratio, VoxZip yields a 1.9x increase in inference throughput alongside a 3.3x reduction in peak memory overhead. Code and models will be available at https://github.com/MM-Speech/VoxZip.

Abstract:
\beginabstract Spatial transcriptomics (ST) characterizes gene expression and spatial tissue architecture, and some ST technologies further provide histology images. To more comprehensively exploit ST data, multimodal modeling is necessary to learn more informative representations. Moreover, since downstream ST analyses, including spatial domain identification, are closely tied to clustering quality, the clustering-friendliness of learned representations is particularly important. However, existing methods often overlook modality imbalance during multimodal fusion, and lack perspective-aware constraints tailored to clustering tasks, making it difficult to learn stable and clustering-friendly representations. To address these issues, we propose \method, a multimodal rebalancing and perspective-aware clustering learning framework for ST data. It jointly learns multimodal representations through modality-specific graph encoding on a shared spatial graph, dual-belief modality calibration mechanism, and the integration of self-attention and cross-attention, while leveraging perspective-aware clustering constraints in Euclidean space and subspace together with spatial topology and expression reconstruction to improve representation quality. Extensive experiments and downstream analyses on multiple benchmark datasets show that \method achieves better overall performance than a range of advanced baseline methods, demonstrating its effectiveness and robustness. The code is released on \urlhttps://github.com/youweiyu/stMRPC. \endabstract

Abstract:
To reduce deployment cost and retraining overhead, adapting pretrained learned image compression (LIC) models to downstream machine vision tasks has attracted growing attention. However, existing methods typically insert fine-tuning modules independently into frozen backbones, lacking explicit mechanisms for cross-layer coordination. To address this limitation, we propose a novel framework named CrossMambaTuning, which integrates State Space Models with cross-layer interaction mechanisms for parameter-efficient fine-tuning. Specifically, we design an efficient Mamba adapter equipped with task-specific prompts and multi-scale branching to precisely capture both local features and global dependencies. Furthermore, we introduce a Scale-Invariant Cross-Layer Adapter (SICA) utilizing a parameter-sharing strategy to fuse task information across different scales and reduce redundancy. Extensive experiments demonstrate that CrossMambaTuning achieves state-of-the-art (SOTA) performance on multiple machine vision tasks, reducing parameter overhead by 72% compared to SOTA methods. Code is available at \urlhttps://github.com/rsr1123/CrossMambaTuning.

Abstract:
Contrastive Language-Image Pretraining (CLIP) learns representations for both text and images in a shared embedding space. While CLIP is originally designed for short text, recent research has attempted to extend the length of its input text when more detailed descriptions are available. However, CLIP still exhibits significant inconsistencies when processing texts of different lengths; that is, the retrieved images based on semantic similarity can vary across texts of different lengths that describe the same image without ambiguity. In this work, we study this under-explored problem and define long-short text consistency. To improve consistency, we propose an objective function that aligns the representations of long and short texts. This objective function, named Multi-Granularity Text Alignment (MGTA-CLIP), introduces long-short text alignments in detailed semantics while retaining CLIP's original text-image semantic alignments. MGTA-CLIP finetunes pre-trained CLIP models and demonstrates significant consistency improvement on zero-shot retrieval tasks. In addition, this improvement in longshort text consistency also leads to enhanced retrieval performance for both long and short texts, achieving state-of-the-art results. The code is available at \urlhttps://github.com/lq-blackcat/MGTA-CLIP.

Abstract:
Current mobile GUI agent benchmarks systematically fail to assess memory capabilities, with only 5.2-11.8% memory-related tasks and no cross-session learning evaluation. We introduce \ourbench, a comprehensive memory-centric benchmark with pass@k and staged LLM-as-judge evaluation. Our contributions include: (1) a systematic memory taxonomy analyzing 11 agents across 5 architectures; (2) 128 tasks across 26 applications where 89.8% challenge memory through cross-temporal and cross-spatial retention; (3) \oureval, an automated pipeline with Progressive Scrutiny and 7 hierarchical metrics; and (4) RQ-driven assessment of 11 state-of-the-art agents. Our experiments reveal significant memory deficits across all evaluated systems, identify 5 distinct failure modes, and synthesize 5 actionable design implications. Core code is available at \urlhttps://memgui-bench.github.io/.

Abstract:
With the proliferation of media data across diverse sources, multimodal information retrieval has emerged as a trending research. However, it faces inherent challenges arising from heterogeneous data sources and complex cross-modal alignment. Although previous studies have revealed modal gaps in feature spaces, a systematic approach to address these challenges remains unexplored. In this work, we introduce UNITE, a universal multimodal embedding framework that jointly supports text, image, and video modalities. We systematically investigate two critical yet underexplored aspects: data curation and modality-aware training configurations. Specifically, we provide a comprehensive analysis of how modality-specific data properties (e.g., cross-modal distributional discrepancies, varying levels of semantic granularity) systematically influence downstream task performance across diverse scenarios. Furthermore, we propose Modal-Aware Masked Contrastive Learning (MAMCL), which selectively masks non-target-modal instances based on modality-aware signals, enabling the model to learn more discriminative representations and effectively mitigating cross-modal competition. Extensive experiments on diverse multimodal retrieval benchmarks show that UNITE achieves state-of-the-art performance across coarse-grained, fine-grained, and instruction-based retrieval tasks. These results demonstrate that strategic modality-specific data curation and tailored training protocols are pivotal for robust universal multimodal representation learning.

Abstract:
Audio Large Language Models (ALLMs) enable spoken interaction but introduce new jailbreak vulnerabilities. Existing perturbation-based jailbreaks do not explicitly control which frequency bands carry the perturbation. Although such perturbations can elicit unsafe responses, repeatedly applying a universal perturbation across diverse inputs may also degrade utility on normal tasks, leaving a conspicuous behavioral footprint that may expose the attack to users or automated monitoring systems and thereby compromise its stealthiness. To determine whether full-band perturbation is necessary, we vary coverage from partial-band to full-band. Jailbreak Success Rate (JSR) varies non-monotonically, while utility degradation grows with coverage. This mismatch shows that selected bands can yield stronger attacks with less utility degradation than full-band perturbations. Based on this observation, we propose GRM, a utility-aware, frequency-selective jailbreak framework that ranks Mel bands by the ratio between jailbreak contribution and transcript sensitivity, confines a universal perturbation to selected bands, and regularizes deviations from the intended request semantics. Experiments on four ALLMs show that GRM achieves an average JSR of 88.46% while substantially reducing utility degradation across benign transcription and response tasks relative to baselines. Our code is available at \hrefhttps://github.com/159753Fetter/GRMthis repository. Warning: This paper contains potentially sensitive content.

Abstract:
Occlusion control is essential for maintaining visual hierarchy and structural coherence in multi-object image generation. However, existing methods often struggle with fine-grained occlusion, where coarse spatial supports can incorrectly suppress visible regions, leading to over-occlusion and structural distortion. We identify inaccurate object support as a key source of such failures, as support errors directly bias transmittance and visibility composition. Based on this insight, we propose Support-Aware Transmittance Refinement (SATR), a training-free framework that couples support estimation with latent transmittance modeling. SATR follows a two-stage inference strategy: it first generates drafts under relaxed occlusion to recover more reliable object support, and then re-injects the refined support into latent rendering to improve transmittance and visibility composition. Moreover, we introduce CompOcc, a benchmark for complex occlusion scenes, together with a structure-aware evaluation protocol covering occlusion correctness, object presence, contact plausibility, and structural fidelity. Experimental results demonstrate that SATR achieves more accurate occlusion control and better structural preservation across diverse complex scenes. Our code and dataset are available at \urlhttps://github.com/jiangw-0/SATR.

Abstract:
Cinematographic captioning aims to describe how a video is filmed using professional film-language concepts such as camera movement, shot size, depth of field, composition, and shooting angle. This capability is important for fine-grained video understanding and controllable movie-quality video generation, yet remains underexplored in existing multimodal large language models. Unlike question-answering-based evaluation of cinematic understanding, cinematographic captioning requires a unified open-form description over multiple cinematographic dimensions. This task is challenging for two main reasons: the model must infer professional cinematographic concepts from subtle visual evidence, and it must generate captions that are both comprehensive and accurate. Accordingly, we propose CineCap, a framework that combines structured reasoning with spatio-temporal anchors and reinforcement learning with comprehensiveness, accuracy, and gated coverage rewards. The former grounds professional cinematographic descriptions in explicit visual evidence and organizes them into compact atomic reasoning for supervised fine-tuning, while the latter improves the balance between descriptive completeness and factual correctness. In addition, we construct CineCap Bench, a benchmark of 472 manually annotated video-caption pairs for systematic evaluation. Extensive experiments show that CineCap consistently outperforms strong proprietary and open-source baselines, establishing a new state of the art for cinematographic captioning. The code, model checkpoint, and benchmark are publicly available in our https://github.com/Hectormxy/CineCap.git.

Abstract:
Utilizing auxiliary outlier datasets for model regularization has shown potential in OOD detection.\@ Given the labor-intensive data collection and cleaning, automated outlier generation has become an attractive alternative.\@ Recent methods address this issue by synthesizing virtual outliers, but often struggle with inferior quality or geometric mismatch due to limited sampling strategies.\@ This paper proposes a novel framework, \textscFantasy-OOD, which formulates the synthesis process as a boundary-outward Markov chain for generating diverse outliers on the hypersphere and imagining their underlying semantics via diffusion models.\@ Specifically, \textscFantasy-OOD first learns a text-conditioned hyperspherical space, then leverages k-NN geometry to induce a cluster-wise OOD-ness density, and finally uses Hamiltonian Monte Carlo (HMC) to instantiate boundary-outward Markov chains that progressively expand from the ID class boundary into the broader OOD region for sampling diverse outliers. These outliers are then decoded into images by the diffusion model.\@ By comparing with SOTA baselines on standard and large-scale benchmarks, we validate the superiority of \textscFantasy-OOD and show that it can be readily extended to generate ID data as \textscFantasy-ID, supporting both ID and OOD generalization.\@ Code is available at: \urlhttps://github.com/736381914/Fantasy-OOD-Official.

Abstract:
InfRared Small Target Detection (IRSTD) is a challenging task. Relying solely on pixel-level information, vision-only methods struggle to distinguish targets from clutter. Current multimodal methods typically describe both targets and backgrounds with a single textual prompt. Such an approach lacks dedicated regional guidance and ignores infrared semantic asymmetry. Consequently, it provides insufficient background suppression information and introduces severe feature optimization conflicts, overwhelming small targets with noise. To address these issues, we propose a novel Asymmetric Dual-text Guided Network (ADGNet). Specifically, accounting for the infrared semantic asymmetry, we first design the Asymmetric Dual-text Prompt (ADP), comprising an image-agnostic abstract target prompt and an image-specific detailed background prompt. To leverage these prompts, we introduce an Asymmetric Dual-Branch Interaction (ADBI) module to separately guide visual features with their respective text priors, protecting targets from noise while suppressing background clutter. Subsequently, we introduce an Adaptive Feature Aggregation (AFA) module to dynamically fuse features from the two branches. Furthermore, we construct a multimodal Asymmetric Image-Text Infrared (AITIR) dataset by providing asymmetric text annotations for three public datasets (IRSTD-1K, NUDT-SIRST, and SIRST). Extensive experiments demonstrate that ADGNet outperforms 21 state-of-the-art (SOTA) methods. Code is available at https://github.com/iLearn-Lab/MM26-ADGNet.

Abstract:
Spatio-temporal forecasting (STF) supports traffic, energy, and climate applications, but modeling temporal dynamics, spatial topology, and inter-node dependence can be costly on large networks. We propose ViST, which transforms the observed signals into a fixed-resolution visual representation rather than relying on external imagery. Three image channels encode topology, temporal evolution, and dynamic node correlations. A conditional reconstructor injects graph and data-driven text context; block-wise cross-modal attention and softmax weighting fuse the reconstructed features with a node-identity representation. Experiments on six datasets cover graph-STF and long-horizon forecasting. On SD, ViST improves the strongest baseline by 2.39% in average MAE and 7.62% in average MAPE. Under the long-horizon protocol, it reduces DMMV's average MSE and MAE by 7.06% and 3.81%, respectively. ViST also trains faster than the evaluated graph baselines on the two largest traffic graphs under the measured settings. Code and sanitized configurations are available at https://github.com/RWLinno/ViST.

Authors: Hao Li (Shanghai Jiao Tong University), Han Fang (China Telecom Artificial Intelligence Technology (Beijing) Co., Ltd.), Zixin Pan (The Hong Kong University of Science and Technology (Guangzhou)), Xin Wei (China Telecom Artificial Intelligence Technology (Beijing) Co., Ltd.), Hongbo Sun (China Telecom Artificial Intelligence Technology (Beijing) Co., Ltd.), Jinglin Xu (University of Science and Technology Beijing), Zhiyu Lin (China Telecom Artificial Intelligence Technology (Beijing) Co., Ltd.), Ye Yuan (China Telecom Artificial Intelligence Technology (Beijing) Co., Ltd.), Zhongjiang He (China Telecom Artificial Intelligence Technology (Beijing) Co., Ltd.), Yu Yu (Shanghai Jiao Tong University), Hao Sun (China Telecom Artificial Intelligence Technology (Beijing) Co., Ltd.)

Abstract:
Although multimodal large language models (MLLMs) have achieved remarkable progress, understanding 3D spatial relationships from 2D images remains a critical challenge. Existing methods primarily rely on symbolic text tokens, which inherently lack the fidelity to represent continuous geometric information. While recent methods use latent representations to enhance reasoning, relying on a single latent type cannot adapt to the diversity of spatial tasks, leading to misalignment in complex geometric scenarios. To address these limitations, we propose GeoAnchor, an interleaved text-latent reasoning framework. GeoAnchor decomposes 3D spatial information into three complementary components: position latents for object grounding, direction latents for relational orientation, and geometry latents for scene structure. These components are recombined in a structured space to construct local evidence while capturing global context, enabling dynamic and interpretable reasoning. Furthermore, we introduce a collaborative training strategy that guides the model from local spatial perception to comprehensive 3D understanding. Extensive experiments on diverse and complex 3D reasoning tasks demonstrate that GeoAnchor outperforms the state of the art, validating its effectiveness and generalization capabilities. Code is available at: https://github.com/JerryPW/GeoAnchor

Abstract:
Collaborative perception breaks through single-view limitations via multi-agent information exchange. However, multi-source noise such as pose errors and communication delays degrades fusion feature quality, constraining perception performance. Joint training of detection and BEV segmentation provides a natural remedy, where segmented road regions help constrain target distributions and detection bounding boxes help recover ambiguous segmentation boundaries. To this end, we propose a robust Collaborative perception framework with expert-driven Detection and bev Segmentation (CoDS). To address spatial inconsistency in fusion quality, we first introduce the Collaborative Reliability Map (CoRM) to explicitly quantify feature quality distribution. Based on CoRM, we design the Semantic Mixture-of-Experts (S-MoE) module to extract differentiated features for inconsistent feature demands. Finally, to further mitigate feature noise degradation, the Bidirectional Task Complementary Interaction (BTCI) refines task-aware features through bidirectional injection. Extensive experiments on OPV2V and V2V4Real datasets show that our CoDS surpasses existing baselines on both tasks and maintains stable robustness under multi-source noise. Code: https://github.com/JinlongW128/CoDS and https://openi.pcl.ac.cn/OpenAIDriving/CoDS.

Abstract:
As an agent-level reasoning and coordination paradigm, Multi-Agent Debate (MAD) orchestrates multiple agents through structured debate to improve answer quality and support complex reasoning. However, existing research on MAD suffers from two fundamental limitations: evaluations are conducted under fragmented and inconsistent settings, making fair comparison difficult, and are largely confined to text-only scenarios, leaving its effectiveness in multimodal settings underexplored. To address these gaps, we introduce M3MAD-Bench, a unified and extensible benchmark for evaluating MAD methods across Multi-domain tasks, Multi-modal inputs, and Multi-dimensional metrics. M3MAD-Bench establishes standardized protocols over five core task domains, including Knowledge, Mathematics, Medicine, Natural Sciences, and Complex Reasoning, covering a total of 13 datasets, and systematically includes both pure text and vision-language data, enabling controlled cross-modality comparison. We evaluate MAD methods on 9 base models spanning different architectures, scales, and modality capabilities. Beyond accuracy, M3MAD-Bench incorporates efficiency-oriented metrics such as token consumption and inference time, providing a holistic view of performance--cost trade-offs. Through extensive experiments, we derive nine key insights, revealing that MAD is not uniformly effective: collaborative methods are generally more robust than adversarial ones, especially on reasoning-intensive and multimodal tasks, but often incur substantial efficiency costs. These findings provide practical guidance for selecting and designing MAD strategies in real-world applications. We believe M3MAD-Bench offers a reliable foundation for future research on standardized and reproducible MAD evaluation. The code is available at ~\urlhttps://github.com/liaolea/M3MAD-Bench.

Abstract:
Most daily activities are inherently procedural. However, existing evaluations for egocentric video understanding seldom address procedural understanding and largely overlook complex key-step-level reasoning under the widely used video question answering (VQA) paradigm for MLLMs. Such capabilities are crucial for building procedural AI assistants deployable on wearable devices. To bridge this gap, we introduce the Egocentric Procedural Understanding VQA task (EgoProceVQA), which systematically evaluates egocentric procedural reasoning abilities of current MLLMs and agents through six types of key-step-centric questions. Furthermore, we develop EgoProceGen, a data generation platform that efficiently constructs QA data tailored to different question types. Based on this platform, we build a benchmark with 3,600 questions, four common procedural scenarios, and 31 everyday procedural tasks. Evaluations on EgoProceVQA show that existing MLLMs and agents still have substantial room for improvement in procedural understanding. Therefore, we further propose EgoProceAgent, a self-skill-exploration agentic framework. We design a generic tool library for procedural understanding and a standardized sub-skill library shared across tools and models, enabling self-exploration without ground-truth supervision. By exploring how to compose and select sub-skills, the agent discovers effective skill strategies for diverse problems, and attains state-of-the-art performance among open-source models on multiple tasks. Together, our benchmark, generation platform, and agentic framework establish a unified foundation for EgoProceVQA. Project page: \urlhttps://z1oong.github.io/EgoProceVQA/.

Abstract:
Single-image 3D hand avatar reconstruction is fundamentally ill-posed and particularly challenging due to limited visual evidence under severe self-occlusion and the complex pose-dependent deformation of highly articulated hands. Existing methods predominantly rely on implicit NeRF-style representations, whose volumetric fitting is computationally expensive and often struggles to preserve fine-grained hand details. In this work, we present OASIS, a tailored 3D Gaussian Splatting framework for single-image hand avatar reconstruction. To faithfully encode sparse image-specific appearance cues in single-view reconstruction, we construct geometry-aligned visual evidence tokens by explicitly aligning input image observations with 3D hand geometry and context-adaptively tokenizing the resulting visual evidence. Since severe self-occlusion makes the reliability of image evidence inherently visibility-dependent, we introduce a visibility-conditioned point-image attention to reliably transfer visual evidence to geometric tokens, yielding occlusion-aware Gaussian features for faithful and robust reconstruction. To further capture non-rigid deformation of articulated hands, we introduce a Feature-on-Mesh representation to enable Gaussian deformation to be guided by local surface stretching. Under this framework, we adopt a one-shot adaptation scheme that learns a shared hand prior from multi-identity training data and then fits it to a target image for target-specific reconstruction. Extensive experiments show that OASIS outperforms existing baselines in both visual fidelity and efficiency across challenging poses and in-the-wild scenarios, and further demonstrates strong versatility in downstream applications such as text-to-avatar generation and texture editing. Code will be released on the \hrefhttps://mova-hand.github.io/MOVA/\textcolormagentaproject page.

Abstract:
Token pruning has emerged as a mainstream approach for developing efficient Video Large Language Models (Video LLMs). This work revisits and advances the two predominant token-pruning paradigms: attention-based selection and similarity-based clustering. Our study reveals two critical limitations in existing methods: (1) conventional top-k selection strategies fail to fully account for the attention distribution, which is often spatially multi-modal and long-tailed in magnitude; and (2) direct similarity-based clustering frequently generates fragmented clusters, resulting in distorted representations after pooling. To address these bottlenecks, we propose Tango, a novel framework designed to optimize the utilization of visual signals. Tango integrates a diversity-driven strategy to enhance attention-based token selection, and introduces Spatio-temporal Rotary Position Embedding (ST-RoPE) to preserve geometric structure via locality priors. Comprehensive experiments across various Video LLMs and video understanding benchmarks demonstrate the effectiveness and generalizability of our approach. Notably, when retaining only 10% of the video tokens, Tango preserves 98.9% of the original performance on LLaVA-OV while delivering a 1.88x inference speedup. Code: \hrefhttps://github.com/xjtupanda/Tangogithub.com/xjtupanda/Tango

Abstract:
Multimodal knowledge graphs (MKGs) enrich entities with modality-specific features. Multimodal knowledge graph reasoning (MKGR) is a core task in MKGs that aims to infer missing facts from existing ones. Existing methods for MKGR are typically embedding-based with limited interpretability. Additionally, they usually fail to generalize to the inductive setting where missing facts involve unseen entities. In this paper, we propose \emphMultimodal Predicate Rule Learning (MPRL), an end-to-end rule learning approach that offers high interpretability while supporting the inductive setting. Specifically, we introduce a novel logical rule formalism, named \emphMultimodal Predicate Rules (MP-rules), which captures modality-specific features associated with entities to improve reasoning accuracy. To effectively learn MP-rules, we parameterize a neural model in MPRL to simulate MP-rule inference, thereby enabling fully end-to-end learning. Experimental results on six datasets demonstrate that MPRL achieves state-of-the-art performance for the inductive setting and remains competitive for the transductive setting. Our case study further shows that MPRL can learn MP-rules with interpretable multimodal predicates. Our code is publicly available at \urlhttps://github.com/majh3/MPRL.

Abstract:
Estimating reliable cross-modality association is crucial to unsupervised visible-infrared person re-ID. While optimal transport is shown to be a practical solution for cross-modality association, it suffers from the rigidness of hard label assignment without considering the impact of cluster noise. Moreover, enforcing only cross-modality contrast is also suboptimal, as it fails to jointly optimize the similarity relation within and across modality. In this paper, we propose a novel framework for cross-modality learning by well exploitation of prototypes: First, instead of contrasting with cross-modality prototypes, we show that modality-unified prototypical contrast facilitates better modality invariance by jointly and simultaneously optimizing similarity relation within and across-modality. Taking self-prototype as a steady teacher, we further refine the instance-prototype online relation through prototype-guided self-distillation. The two components are optimized in a unified framework, leading to a simple yet effective model. On standard VI-ReID benchmarks, we perform extensive comparison and analysis, validating the effectiveness of our proposed method. Code is available at: https://github.com/Terminator8758/PoSeD.

Abstract:
Self-supervised learning (SSL) encoders are vulnerable to backdoor attacks, posing threats to both visual SSL encoders and vision-language encoders. Existing defenses are typically designed for only one of these paradigms and rely on restrictive assumptions such as access to uninfected in-distribution data or precomputed pseudo-labels, which are difficult to satisfy in practice. To address these limitations, we propose DEFUSE, a generalizable backdoor detection framework for SSL encoders. Inspired by Bayesian posterior inference, we reformulate backdoor detection as a representation-conditioned image likelihood estimation problem parameterized by a conditional diffusion generative model. Uninfected representations tend to yield semantically consistent reconstructions, whereas backdoored ones are more likely to be mapped to the attacker's target class or semantically meaningless images, deviating from the original semantics and thereby exposing the backdoor. However, we find that the exact likelihood is intractable, because highly abstracted representations discard the low-level information necessary for pixel-faithful reconstruction. We therefore relax the objective to semantic reconstruction and evaluate it in a well-separated representation space provided by a reference encoder. Rather than training from scratch, we fine-tune a pretrained diffusion model, leveraging its generative prior to map data onto the natural image manifold while preserving semantic content. Extensive experiments demonstrate that DEFUSE substantially outperforms existing detectors across diverse attack settings, generalizing to both visual SSL and vision-language encoders. Notably, our method greatly reduces the reliance on prior knowledge about the victim encoder or the attack strategy. The source code is available at \urlhttps://github.com/jsrdcht/DEFUSE.

Abstract:
As AI video generators achieve cinematic realism, reliable detection becomes essential for safeguarding digital trust. We identify cross-scale coupling mismatch as a new forensic signal, where scale refers to the level of abstraction (semantic dynamics vs. pixel-level residuals): in natural videos, macro-level temporal dynamics and micro-level residual patterns are intrinsically coupled by the unified imaging physics pipeline, whereas AI generators, whose training objectives do not explicitly preserve this joint distribution, systematically violate this coupling. Detecting such mismatch is challenging because it requires independently extracting information at both scales while simultaneously quantifying their cross-scale relationship. We propose RIFT (Representation Inconsistency Forensics on Trajectories), an orthogonal forensic framework that addresses this through three interlocking components: a macro stream that builds a dynamic baseline of expected temporal evolution via differential geometry and persistent homology on learned manifold trajectories, a micro stream that acts as a sensitive forensic probe via steganalytic filtering and temporal modeling, and a coupling divergence module that measures the conditional dependency between the two streams. Gram-Schmidt orthogonality guarantees the information-theoretic validity of this measurement. Experiments on two benchmarks (VidProM, 120K videos, 7 generators; GenVidBench, 68K videos, 4 generators) demonstrate that RIFT achieves 99.33% and 99.72% F1-score respectively, with 97.87% unseen-generator detection rate in leave-one-out evaluation, while exhibiting encoder agnosticism: scaling from ViT-S/14 (22M) to ViT-L/14 (300M) changes F1 by less than 0.1%, and switching to a different encoder family (DINOv1) reduces F1 by only 0.73 pp. Code is available at https://github.com/Litsay/RIFT.

Abstract:
Sparse-view reconstruction with 3D Gaussian Splatting (3DGS) is fundamentally ill-posed due to insufficient geometric supervision, often leading to severe overfitting and the emergence of structural distortions and translucent haze-like artifacts. While existing approaches attempt to alleviate this issue via dropout-based regularization, they are largely heuristic and lack a unified understanding of artifact formation. In this paper, we revisit sparse-view 3DGS reconstruction from a new perspective and identify the core challenge as the unobservability of Gaussian primitive reliability. Unreliable Gaussians are insufficiently constrained during optimization and accumulate as haze-like degradations in rendered images. Motivated by this observation, we propose a unified Dual-domain Observation and Calibration framework for Sparse-View Gaussian Splatting (DOC-GS) that models and corrects Gaussian reliability through the synergy of optimization-domain inductive bias and observation-domain evidence. Specifically, in the optimization domain, we characterize Gaussian reliability by the degree to which each primitive is constrained during training, and instantiate this signal via a Continuous Depth-Guided Dropout (CDGD) strategy, where the dropout probability serves as an explicit proxy for primitive reliability. This imposes a smooth depth-aware inductive bias to suppress weakly constrained Gaussians and improve optimization stability. In the observation domain, we establish a connection between floater artifacts and atmospheric scattering, and leverage the Dark Channel Prior (DCP) as a structural consistency cue to identify and accumulate anomalous regions. Based on cross-view aggregated evidence, we further design a reliability-driven geometric pruning strategy to remove low-confidence Gaussians. Extensive experiments on multiple benchmarks demonstrate that DOC-GS consistently outperforms existing methods for sparse-view reconstruction, suppressing haze-like artifacts and improving geometric fidelity in three representative datasets. The source code is available at \urlhttps://github.com/lht65/DOC-GS.

Abstract:
Point cloud compression (PCC) is critical for efficient storage and transmission of 3D data. While recent learning-based PCC methods achieve good rate-distortion (R-D) performance, they generally rely on ideal transmission conditions. In practice, packet loss is a common issue and can severely distort latent features, causing coordinate drift and geometric degradation. To address this challenge, we present ResPCC, the first end-to-end neural point cloud codec designed to offer intrinsic resilience against data loss. Our framework is loss-rate-aware and adapts to diverse packet loss conditions. At the encoder, we introduce a Condition-Adaptive Latent Modulation (CALM) module to adjust latent feature distributions according to the perceived loss rate, as well as a Spatial-Channel Interleaving (SCI) mechanism that transforms channel-wise data extinction into spatially scattered element-wise missing patterns. At the decoder, we develop a Mask-Aware Graph-based Latent Restoration (MGLR) module, followed by a Dictionary-based Refinement (DBR) stage to recover corrupted features and align them with canonical priors. Evaluations on ShapeNet and SemanticKITTI under 5% to 30% packet loss rates show that ResPCC consistently delivers superior stability and R-D performance over baselines. Our framework maintains high reconstruction fidelity under lossy conditions, providing a reliable solution for 3D data transmission over practical networks. Code is available at \urlhttps://github.com/starrynight314/ResPCC.

Abstract:
As large language models (LLMs) evolve into autonomous agents capable of acting in open-ended environments, ensuring behavioral alignment with human values becomes a critical concern. Existing benchmarks, focused on static, single-turn prompts, fail to capture the interactive and multi-modal nature of real-world conflicts. We introduce ConflictBench, a benchmark for evaluating humanAI conflict through 150 multi-turn scenarios derived from prior alignment queries. ConflictBench integrates a text-based simulation engine with a visually grounded world model, enabling agents to perceive, plan, and act under dynamic conditions. Results show that while agents often act safely when human harm is immediate, they frequently prioritize self-preservation or adopt deceptive strategies in delayed or low-risk settings. A regret test further reveals that aligned decisions are often reversed under escalating pressure, especially with visual input. These findings underscore the need for interaction-level, multi-modal evaluation to surface alignment failures that remain hidden in conventional benchmarks. Beyond evaluation, we show that ConflictBench can also support alignment improvement by generating high-quality interactive trajectories for targeted fine-tuning, significantly enhancing agent behavior in complex conflict scenarios. Code and data are available at: https://github.com/Yranes/ConflictBench.

Abstract:
Robust message embedding in images is important for multimedia security applications such as copyright protection and content tracing. Existing methods are largely developed under a \emphdistortion robustness paradigm, where the embedded signal remains spatially present but is degraded by noise, blur, or compression. Severe cropping poses a fundamentally different challenge because it removes part of the carrier itself, causing partial payload disappearance rather than mere signal corruption. In this paper, we revisit robust message embedding from an \empherasure-resilience perspective and present CREST, a proof-of-concept framework for severe-cropping-robust embedding. CREST combines coding-theoretic redundancy with neural embedding and recovery by expanding a compact QR message into a redundant spatial payload via LT fountain coding and coupling it with cropping-aware embedding and fragment recovery. Experiments on COCO, DIV2K, and VOC2012 show that CREST improves recovery under severe cropping while maintaining competitive visual quality. Under mixed distortions with an area retention ratio of 0.7, CREST improves TRA from 18.52% to 68.45% and reduces EMR from 13.88% to 4.21% over the strongest baseline. On COCO2017, CREST still achieves 48.55--65.12% TRA when only 30--50% of the image area is retained, whereas all compared baselines fail to recover the message\footnoteCode: https://github.com/ZPZhangLab/CREST-ACMMM2026. These results suggest that severe cropping is better understood as an erasure problem rather than a conventional distortion problem, motivating the joint design of neural embedding and coding-based recovery.

Abstract:
Existing co-speech gesture generation methods are predominantly studied in offline settings, where gestures are synthesized from complete speech segments. However, interactive digital humans in real-world scenarios are required to generate speech-synchronous gestures online, using only currently available response audio under strict latency constraints. As a result, prior methods are unsuitable for real-time interaction, as they either rely on future speech information or incur substantial inference delay. In this paper, we formulate online co-speech gesture generation for interactive digital humans and propose a real-time interactive framework that couples a streaming speech response module with an online gesture generation module. Specifically, the gesture generator is designed as a causal multimodal autoregressive model that predicts body motion from streaming response speech and motion history, enabling low-latency and speech-aligned gesture synthesis without access to future speech. To support this setting, we further propose an offline data synthesis pipeline tailored to virtual companion scenarios, which leverages topic- and emotion-aware subject corpora to construct diverse human-agent dialogues and then generates co-speech gestures conditioned on the agent responses. Moreover, to bridge the gap between offline data construction and online deployment, we establish a self-evolving training loop by incorporating user feedback collected during online interaction into the data generation process, enabling continual adaptation to user preferences. Extensive experiments demonstrate that our framework achieves superior better latency-quality trade-off, stronger speech-motion synchronization, and higher user preference than competitive existing baselines. Project Page: \urlhttps://super-star-2026.github.io/

Abstract:
Music-to-humanoid dance is commonly implemented as a two-stage pipeline in which a music-conditioned generator produces a reference motion and a fixed whole-body controller (WBC) executes it. In this setting, reference-space quality is only a proxy for the quality of the executed motion. We focus on two upstream decisions that strongly affect this proxy gap: how temporal planning units are defined and how generated candidates are selected before execution. ChoreoPlan introduces beat-snapped, variable-length planning segments with hybrid discrete and continuous motion attributes, together with an \emphEmbodied Selector trained on offline rollouts of the fixed controller. The planner provides beat-aligned choreography guidance in humanoid token space, while the selector reranks candidates using predicted execution quality and semantic compatibility. Across AIST++- and FineDance-derived humanoid tracks in IsaacGym and MuJoCo, ChoreoPlan improves rollout success, tracking accuracy, executed beat alignment, and semantic retention over retrained baselines. Qualitative Unitree G1 demonstrations further illustrate coherent and executable dance motions. Code and video results are available at \urlhttps://vkgo.github.io/ChoreoPlan.

Abstract:
Multimodal anomaly detection benefits from complementary RGB and 3D evidence, yet auxiliary RGB reconstruction is not equally reliable across product categories and class-wise test-time policy selection is usually unavailable. We propose TC-MAF, a base-anchored multi-evidence fusion design that combines a multimodal detector, complementary Dinomaly evidence, and a small cross-modal consistency cue under one fixed pixel-level fusion formula. A lightweight training-dispersion confidence (TDC) term scales auxiliary participation using only normal training statistics. On MVTec-3D, TC-MAF reaches 0.979 image-level AUROC and 0.990 pixel-level AUPRO, achieving the best mean results on both detection and localization among the compared multimodal methods. Systematic ablations show that the fusion structure itself is the dominant factor, while TDC provides a smaller but reproducible calibration gain over no calibration or arbitrary calibration. Additional experiments show that the same design remains effective under a pooled-statistics variant, auxiliary-branch and backbone substitutions, few-shot settings, a missing-3D setting, and cross-dataset evaluation on Eyecandies. Code is available at https://github.com/Chranos/TC_MAF.

Abstract:
Zero-shot skeleton-based action recognition (ZSAR) aims to recognize unseen action categories by aligning skeleton features with textual semantics. However, existing methods rely on text-derived prototypes that inherently lack geometric structure and physical constraints, resulting in a pronounced semantic-kinematic gap. To bridge this gap, we propose GenPrior, the first framework to exploit generative priors from pre-trained Text-to-Motion (T2M) models for ZSAR. Specifically, we introduce Dispersion-Gated Feature Fusion, which distills kinematic prototypes and intra-class dispersion from generative motion sequences and employs a learned gating network to adaptively inject reliable structural cues into textual embeddings while suppressing synthetic artifacts. Furthermore, we propose Generative Prototype Refinement, which leverages these generation-enhanced prototypes as anchors to mine high-confidence unseen samples, calibrating class prototypes toward the true distribution and thereby unleashing strong performance gains. Extensive experiments on NTU-60, NTU-120, and PKU-MMD demonstrate that GenPrior achieves state-of-the-art performance under both zero-shot and generalized zero-shot settings. Code is available at https://github.com/jidongkuang/GenPrior.

Abstract:
Conversational Speech Synthesis (CSS) aims to synthesize speech with human-like emotional expression and contextual consistency in user-agent interactions. Existing CSS methods struggle to render authentic human emotions due to limited predefined emotion label spaces (e.g., seven emotion categories), while redundant multimodal tokens in multi-turn dialogue history interfere with context understanding. To address these issues, we propose AuEmoChat, a CSS framework for authentic emotion understanding and rendering. First, we develop AuEmoCodec, which learns a discrete authentic emotion token space from large-scale emotional speech via finite scalar quantization, enabling a more authentic emotion representation than limited basic emotion categories. Furthermore, we propose AuEmoToMe, an authentic-emotion-guided token merging algorithm that merges redundant tokens in multimodal dialogue history while preserving emotion-relevant context. We integrate it into an autoregressive text-speech model to predict the target authentic emotion token and speech tokens. Finally, we propose Authentic Emotion Flow Matching, which renders speech by jointly conditioning on merged dialogue context, target authentic emotion, and acoustic priors. Extensive experiments on the NCSSD-EmCap dataset demonstrate that AuEmoChat outperforms state-of-the-art CSS baselines and generates more expressive and authentic emotional speech. The code and speech demos will be available at: https://github.com/AI-S2-Lab/AuEmoChat.

Abstract:
Knowledge-Based Visual Question Answering (KB-VQA) requires retrieving entity knowledge from external sources to answer visually grounded questions. Existing retrieval-augmented systems suffer from two critical limitations. First, relying on a single retrieval modality creates a Single-Source Retrieval Bottleneck, missing ground-truth entities that are only accessible through complementary sources. Second, dual-tower pointwise rerankers suffer from Retrieval-Source-Blind Reranking, as they overlook retrieval origins and candidate-level retrieval priors, leading to redundant modality reliance. To address these challenges, we propose UniHEAR, a unified lightweight framework for heterogeneous-source entity retrieval and reranking. UniHEAR constructs a Coarse Retrieval Descriptor for each candidate entity, and introduces Retrieval-Guided Attentive Modality Gating to condition modality attention weights on this descriptor, complemented by Entropy-Weighted Source Fusion of coarse retrieval priors. A hybrid training strategy combining contrastive learning with an auxiliary modality-preserving loss unifies entity-level and section-level retrieval within a single model. Extensive experiments on E-VQA and InfoSeek demonstrate that UniHEAR achieves state-of-the-art retrieval and VQA performance, improving Recall@1 by 6.7 and 1.2 points over the strongest baselines while maintaining a lightweight reranking architecture. Code and model are available at https://github.com/iven-luo/UniHEAR.

Abstract:
Generative retrieval reformulates text-to-image retrieval as autoregressive generation of discrete image tokens, offering a promising alternative to conventional dense-retrieval paradigms. However, existing approaches typically quantize a single global visual representation, struggling to capture the hierarchical semantics of images spanning scene-level context and entity-level details. Furthermore, the standard training objective maximizes only target token likelihood, lacking discriminative ability against semantically similar but irrelevant candidates. To address these, we propose HiToken, which decomposes visual semantics into scene-level and entity-level representations quantized by separate residual codebooks for structured hierarchical tokens, and leverages an LLM to decompose captions into scene and entity descriptions for level-specific cross-modal supervision. A dual hard negative discriminative training strategy further mines image-side negatives via token prefix overlap and text-side negatives via semantic similarity, jointly optimized with the generation objective through ranking losses to enhance discriminative power. Extensive experiments on MS-COCO and Flickr30K demonstrate that HiToken outperforms existing generative retrieval methods. The code is available at https://github.com/immc-lab/HiToken.

Abstract:
Monocular depth estimation (MDE) predicts per-pixel depth from a single RGB image to recover the 3D structure of a scene. When trained with LiDAR supervision, MDE models can learn to predict metric-scale depth directly. However, LiDAR annotations projected onto the image plane are spatially sparse and discontinuous and may provide limited supervision near depth discontinuities and thin structures. This issue can be more challenging for lightweight networks due to limited representational capacity and repeated downsampling operations. To mitigate this issue, we propose a feature-decoupled lightweight dual-path architecture that explicitly separates structural and contextual representations during feature encoding. Based on a decouplerefineintegrate design principle, the encoder forms role-specific structural and contextual features, refines the two representations independently, and finally re-integrates contextual information under structural guidance. The proposed model can be trained from scratch on a single GPU without external pre-training while achieving competitive performance with only 2.9M parameters. Experiments on KITTI and NYU Depth v2 demonstrate competitive performance against representative lightweight backbones and dedicated MDE models. In addition, zero-shot cross-dataset evaluation from KITTI to DDAD demonstrates improved transferability across different camera configurations and outdoor driving data distributions without additional training or fine-tuning. Our code and models are publicly available at https://github.com/starsleep/FD-Depth.

Abstract:
Interleaved-modal Chain-of-Thought (I-MCoT) has enhanced the capabilities of Vision-Language Models (VLMs) in Visual Question Answering (VQA) by grounding the reasoning process in inserted visual evidence. However, existing approaches often operate under a largely unexamined assumption that visual insertions are advantageous. In contrast, our analysis reveals a fundamental trade-off: while the introduced visual information aids reasoning, the insertion action could impair reasoning fluency and model performance, with the impairment scaling with token quantity and insertion frequency. Ultimately, the detrimental effects can outweigh the benefits of visual insertion. In this paper, we propose ADROIT, a training-free I-MCoT framework that explicitly models and resolves this trade-off via two synergistic components: (1) Visual Sufficiency Trigger (VST), which helps a VLM avoid superfluous insertions by identifying insertion moments where the gain from visual information is theoretically guaranteed. (2) Optimal Transport Planner (OTP), which retrieves the compact, most reasoning-relevant, and semantically coherent visual evidence for a VLM by formulating evidence selection as an entropy-regularized optimal transport problem. Extensive experiments on five benchmarks and four backbones confirm the consistent superiority of our proposed approach. Our code is available at \urlhttps://github.com/HsipingLi/ADROIT.

Abstract:
Model extraction attacks pose a severe threat to the intellectual property of proprietary MLaaS models. While well-studied for classification, extracting object detectors in black-box settings remains challenging due to task complexity and data scarcity. Prior methods often rely on public datasets related to the problem domain, an assumption which is often impractical in strict black-box scenarios. However, the emergence of multimodal generative models capable of high-fidelity text-to-image synthesis has fundamentally altered this landscape. These advancements enable the execution of model extraction attacks against unimodal models under conditions previously considered unfeasible due to the lack of data. Leveraging this shift, we introduce \method, a framework that utilizes diffusion models to synthesize a diverse, purely synthetic training set, refined by an Oracle-guided filtering pipeline. Experiments demonstrate that our approach outperforms state-of-the-art baselines, including those utilizing real data, without accessing a single problem-domain image. Notably, \method achieves robust generalization in cross-architecture scenarios, recovering over 83% of the Oracle's performance. These findings expose a critical vulnerability, proving that data scarcity is no longer a defensive barrier against extraction attacks driven by multimodal models. Our code is available at \urlhttps://github.com/jhonatanleao/OD-Copycat

Abstract:
Pre-trained diffusion models provide strong natural-image priors for all-in-one image restoration (AiOIR), but most existing diffusion-based AiOIR methods still inherit the standard generative pipeline by starting from pure Gaussian noise and performing multi-step denoising conditioned on degraded inputs. Such a design is computationally expensive and also suboptimal for restoration, where degraded images already preserve substantial recoverable content. Although one-step distillation improves efficiency, it typically forces the diverse restoration behaviors associated with different diffusion timesteps into a single shared mapping, making it insufficiently adaptive for AiOIR. In this paper, we present MoDEIR, a Mixture-of-Diffusion-Experts framework that reformulates a sparse set of preset diffusion timesteps as a selectable pool of timestep-specific one-step restoration experts. Each expert projects the degraded input to the intermediate latent state associated with its preset timestep and restores it in one denoising step, thereby preserving a distinct restoration tendency inherited from diffusion priors. To better retain reliable structures from degraded inputs, we introduce a Timestep-aware Structural Compensation Module (TSCM) that adaptively modulates latent perturbation and early denoising features using degradation-aware structural cues. We further design a Timestep-aligned Router (TAR) to estimate the restoration demand of each degraded input and select suitable experts from the expert pool. Finally, lightweight Fidelity Reconstruction Modules (FRMs) are inserted into the VAE decoder to improve reconstruction fidelity after latent fusion. Extensive experiments show that MoDEIR achieves strong overall performance in restoration quality and efficiency across multiple AiOIR benchmarks, while also generalizing well to unseen degradations. Code will be released at \hrefhttps://github.com/WYu-Feng/MoDEIR\textttMoDEIR.

Abstract:
Latent Flow Models have revolutionized compressed-space image synthesis, yet their application to high-fidelity inverse problems remains bottlenecked. In this paper, we trace this dilemma to a fundamental geometric limitation of pre-trained autoencoders, which we term \emphFirst-Order Manifold Blindness. Severe decoder compression (e.g., retaining only ～\!2% of the original degrees of freedom) produces a rank-deficient Jacobian, rendering high-frequency measurement residuals in its orthogonal complement invisible to latent gradients even when the decoder can represent the target image. To overcome this bottleneck, we propose Hybrid-Domain Posterior Sampling (HDPS), a decoupled inference framework that disentangles physical measurement consistency from semantic prior modeling. HDPS diverges into the pixel space, leveraging Langevin dynamics to absorb precise orthogonal measurement gradients, and subsequently projects these structural corrections back onto the generative manifold. An optimization-based latent alignment is introduced to filter pixel-space artifacts while avoiding the semantic drift of direct encoding. Extensive experiments on diverse inverse problems demonstrate that HDPS establishes a new state-of-the-art, successfully recovering the high-frequency structural precision that latent-only solvers inherently discard. The code is available at https://github.com/74587887/HDPS.

Abstract:
While Text-to-Image (T2I) diffusion models have achieved remarkable success, precise spatial and orientational control in multi-object scenes remains a persistent challenge. Existing methods either rely on computationally expensive dense 3D maps or suffer from severe attribute leakage and ``cut-and-paste'' artifacts. To address these limitations, we propose PoseAdapter, a lightweight framework for high-fidelity 2.5D controllable image generation. Instead of dense spatial maps, it establishes precise spatial-angular anchors using an efficient condition layout: individual object captions, 2D bounding boxes, and 3D angles. To resolve the generative trade-off between strict instance isolation and global coherence, we introduce a Context-Aware Dual-Stream Representation. By injecting local object tokens and relation-enriched scene tokens into the visual stream of modern MM-DiT architectures via parallel masked and unmasked pathways, PoseAdapter eliminates attribute leakage while preserving natural inter-object relationships and scene-level coherence. To support this paradigm, we construct OrientLayout, a high-quality dataset featuring standardized 2.5D annotations and instance-level decoupled semantics. Extensive experiments demonstrate that PoseAdapter outperforms state-of-the-art baselines in spatial accuracy, orientational precision, and multi-object visual fidelity. Code and dataset will be available at \urlhttps://github.com/cyf23/PoseAdapter.

Abstract:
Multimodal learning holds significant promise for time series modeling by integrating both visual and textual data. However, existing methods primarily focus on textual modalities, overlooking the potential of visual information. In this work, we advocate for cross-modal visual augmentation as a powerful paradigm to enhance multivariate time series classification by leveraging rich, pretrained visual models. We identify three core challenges in bridging time-series and visual modalities: semantic mismatch, structural asymmetry, and divergent patching strategies. To address these, we propose TimesBook, a novel framework that achieves consistent semantic and structural alignment. Specifically, TimesBook utilizes (1)~Gromov Wasserstein Mapping to semantically align temporal and visual modality. (2)~Masked Variate Composer to map time series variables into a visual representation, ensuring structural alignment with visual data; (3)~Multimodal Codebook to combine a pretrained visual codebook and a trainable time series codebook with an integrated translator to enable fine-grained semantic alignment; (4)~Fusion Classifier to enable a robust and selective cross-modal information fusion. Extensive experiments on 22 multivariate time series classification benchmarks show that TimesBook achieves state-of-the-art performance, with improvements of up to 6.9% over the second-best method. Notably, TimesBook maintains comparable performance to baselines in few-shot regimes, demonstrating its superior generalization ability. Codes are available at \urlhttps://github.com/JJPQ/TimesBook.

Abstract:
Historical document analysis is crucial for cultural heritage preservation and low-resource language digitization. However, real historical documents often suffer from severe degradation, such as ink erosion and structural damage, causing mixed-visibility character sequences. Conventional OCR systems struggle in such evidence-limited scenarios, while contextual language models may over-correct visually clear characters. To address these challenges, we propose DDVR-Net, a degradation-aware decoupling and visibility-guided restoration framework for historical document recognition. DDVR-Net decouples intrinsic glyph structures from physical corruption at the feature level and predicts local visibility to estimate character reliability. A contextual decoder is selectively activated through a visibility-guided gate, restoring degraded characters while preserving clear visual recognition. Furthermore, we construct a low-resource multilingual benchmark covering Tibetan, Mongolian, and Uyghur scripts with 187000 document pairs, combining large-scale synthetic data and progressive training with real-world annotations to reduce the synthetic-to-real gap. Extensive experiments demonstrate that DDVR-Net consistently outperforms existing methods across three languages and diverse reading orders, providing an effective framework and benchmark for low-resource historical document recognition. Code and Benchmark: \urlhttps://github.com/Shajiu/DDVR-Net

Abstract:
Driver emotion recognition is a safety-critical capability for advanced driver-assistance systems. However, existing approaches face a persistent trade-off: unimodal visual methods lack discriminability, multimodal fusion introduces deployment complexity, and recent vision-language adaptations either underutilize semantic priors through naive zero-shot transfer or risk catastrophic forgetting via indiscriminate full fine-tuning. To address these limitations, we propose Prompt Aggregation and Calibration for Emotion Recognition (PACER), a parameter-efficient framework that reformulates driver affect recognition as semantic alignment between video representations and emotion-aware language prototypes within a frozen CLIP feature space. To address the limited expressiveness of single-prompt matching, PACER introduces Class-Adaptive Prompt Calibration to distill multiple driver-specific textual descriptions into class-representative prototypes via learnable per-class importance weighting. Since emotional states manifest as subtle temporal dynamics rather than isolated static cues, PACER incorporates Temporal-Aware Gated Aggregation to capture the progressive evolution of driver affect across video frames with gated pooling-augmented temporal transformer. By jointly enhancing prototype construction and temporal visual adaptation without modifying the pre-trained backbone, PACER achieves discriminative and parameter-efficient recognition with only 10.5 M trainable parameters (approximately 6.9% of the CLIP backbone) and a mere 0.0297 GFLOPs of additional computation. Experiments on our AIDE-Emotion and the RAVDESS benchmark demonstrate that PACER consistently outperforms state-of-the-art methods across both in-vehicle and unconstrained settings. Practical deployment on the NVIDIA Orin-X automotive computing platform further validates its efficiency, with PACER achieving approximately 110 FPS inference speed. Code, checkpoints, and the self-constructed AIDE-Emotion dataset are released at https://github.com/Xigua-Liangdao/PACER.

Abstract:
Hallucination remains a persistent challenge in generative super-resolution (GSR), where reconstructed results may contain visually plausible yet weakly supported content, structural deviations, or unnatural textures with respect to the low-resolution (LR) input. Existing GSR methods have extensively explored the trade-off between perceptual realism and reconstruction fidelity, but the division between preserving reliable coarse-scale information and restoring more uncertain fine details is often handled implicitly within the overall restoration process. Visual autoregressive (VAR) modeling provides a natural opportunity to revisit this issue, as its coarse-to-fine next-scale prediction offers an explicit scale-wise generation interface. However, existing VAR-based SR methods still inherit the original full 1-to-N autoregressive generation path, even though, for super-resolution, coarse-scale information in LR is often relatively more reliable, while long autoregressive chains may accumulate prediction errors. Motivated by these observations, we propose K2N, which reformulates VAR-based SR from full-path generation into a k-to-N detail continuation process. Specifically, early coarse-scale states are established directly from LR, while only the remaining finer scales are restored autoregressively. Experimental results show that K2N remains competitive with the VARSR baseline on standard SR metrics, while exhibiting clearer advantages on hallucination-focused evaluation. These findings suggest that explicitly rethinking the generation path in a scale-wise manner can be a promising direction for improving the reliability of generative super-resolution. Our code will be released soon at: https://github.com/BRL-SYSU/K2NSR.

Abstract:
Spiking Neural Networks (SNNs) offer superior energy efficiency over Artificial Neural Networks (ANNs). However, they encounter significant deficiencies in training and inference metrics when applied to Spiking Vision Transformers (S-ViTs). Existing paradigms including ANN-SNN Conversion and Spatial-Temporal Backpropagation (STBP) suffer from inherent limitations, precluding concurrent optimization of memory, accuracy and energy consumption. To address these issues, we propose GemS-T, a novel architecture implementing grouped computation across temporal, spatial and network structure dimensions. Specifically, we introduce the Grouped-Exponential-Coding-based IF (ExpG-IF) model, enabling lossless conversion with constant training overhead and precise regulation for spike patterns. Additionally, we develop Group-wise Spiking Self-Attention (GW-SSA) to reduce computational complexity via multi-scale token grouping and multiplication-free operations within a hybrid attention-convolution framework. Experiments confirm that our method can achieve superior performance with ultra-high energy efficiency on challenging benchmarks. To our best knowledge, this is the first work to systematically establish multi-dimensional grouped computation for resolving the triad of memory overhead, learning capability and energy budget in S-ViTs. Code is available at \urlhttps://github.com/hzc1208/Ge2mST.

Abstract:
Previous RGB-D semantic segmentation methods commonly employ dual encoders to separately process RGB and depth inputs, followed by dedicated modules for cross-modal feature fusion. However, such designs often inadequately capture depth representations and consequently limit effective cross-modal interaction, while the additional encoder branch introduces redundant computation that hinders lightweight execution. To tackle these challenges, we propose URNet, a Unified Reparameterized RGB-D Network that performs simultaneous multi-modal feature extraction and cross-modal fusion within a single encoder. Specifically, we adopt a reparameterization strategy to compact the network architecture and facilitate fast inference. Within each Reparameterized Block (RepBlock), a Linear Gated Attention (LGA) module is introduced to fully exploit complementary RGB and depth cues across different feature scales. Furthermore, considering that decoder design has been relatively underexplored in existing RGB-D segmentation models, we develop a concise yet effective universal decoder, termed the Pyramid Merging Decoder (PMD). Extensive experiments on multiple RGB-D segmentation benchmarks demonstrate that URNet achieves state-of-the-art performance while maintaining high efficiency. Code will be available at \urlhttps://github.com/Wild-Stephen/URNet.

Abstract:
Vision-Language Models (VLMs) suffer significant performance degradation when deployed in specialized target domains, and existing adaptation methods are hindered by noisy pseudo-labels. We observe that VLMs typically retain the correct class among their top candidates but fail to distinguish it from visually or semantically similar confounders. This suggests that discriminative knowledge is not lost but diluted: the feature dimensions useful for separating one class pair can differ from those useful for another, whereas standard cosine similarity treats all dimensions uniformly. We therefore propose \bf Pairwise Discriminative Comparison for Reliable Pseudo-Labeling (DisCo), which decomposes classification into pairwise comparisons in class-specific subspaces and aggregates their evidence for robust pseudo-label generation. An efficient average-mask score reduces per-sample computation from O(K^2d) to O(Kd) while emphasizing dimensions that are consistently discriminative across class pairs. The refined pseudo-labels guide iterative prompt learning through knowledge distillation and subspace-aware prototype alignment. Experiments on three domain-adaptation benchmarks and eleven general-recognition datasets show that DisCo consistently outperforms prior unsupervised baselines and remains competitive with few-shot supervised methods. Code is available at \urlhttps://github.com/ssbhss/DisCo.

Abstract:
Source-Free domain adaptive Object Detection (SFOD) addresses the practical challenge of adapting object detectors to unlabeled target domains without access to source data. Existing methods typically rely on consistency regularization between artificially constructed sample pairs, which often limits generalization under severe domain shifts and exacerbates error accumulation. We propose that effective domain adaptation should follow a curriculum driven by the model's own learning state: the model should consolidate what it has already mastered while actively exploring what remains unmastered, creating a self-sustaining loop of progressive improvement. To this end, a Self-Teaching framework (ST-SFOD) is proposed, comprising two synergistic components: (1) a Learning State Diagnoser (LSD) that intelligently partitions target data into mastered and unmastered subsets by analyzing gradient alignment against an evolving consensus; and (2) a Self-Teaching Loop (STL) that formulates adaptation as a bi-level optimization, where the model explores unmastered samples while being validated against mastered ones to ensure learned patterns generalize across the target distribution. Extensive experiments on five benchmark transfer scenarios demonstrate that ST-SFOD achieves state-of-the-art performance. The source code is available at https://github.com/liuukk/ST-SFOD.git.

Abstract:
Multimodal Emotion Recognition in Conversations (MERC) requires a sophisticated synthesis of linguistic, visual, and acoustic signals. Traditional methods predominantly rely on a direct mapping from activating events to emotional labels (A to C), often overlooking the internal cognitive appraisal process. Consequently, these "black-box" approaches struggle to distinguish surface-level stimuli from belief-driven emotional transitions. To address this, we reformulate MERC as a belief-mediated reasoning task and propose ABC-Emo, a framework inspired by the ABC emotion theory. Specifically, ABC-Emo mediates the transition from Activating Events (A) to emotional Consequences (C) by explicitly integrating vocal, facial, and contextual Beliefs (B). To emulate human cognitive flexibility, we introduce a Dynamic Belief Update (DBU) module that calibrates historical context with real-time perceptual cues via a gated cross-modal mechanism. By synergizing multi-source beliefs with refined contextual states, ABC-Emo consolidates evidence into a unified sequence for joint reasoning via Large Language Models (LLMs). Extensive experiments on AvaMERG, MELD, and IEMOCAP benchmarks demonstrate that ABC-Emo consistently outperforms state-of-the-art methods. Qualitative analyses further confirm that our belief-catalyzed approach provides a traceable and robust reasoning path, particularly in complex scenarios involving expectation violations. Code is available at https://github.com/x1bhm/ABC-Emo.

Abstract:
Open-Vocabulary Multimodal Emotion Recognition (OV-MER) aims to recognize human emotions from multimodal inputs beyond predefined label sets. Recent large multimodal models (LMMs) have shown strong open-vocabulary reasoning capabilities for this task, yet relying on a single LMM inevitably introduces inherent emotional bias. In this paper, we propose MAGIC-MER, a Multi-Agent Game-theoretic Consensus framework for open-vocabulary emotion recognition that formulates emotion inference as a multi-round imperfect-information game among heterogeneous LMMs. Each agent holds private emotional priors as hidden information, analogous to hole cards in poker, and engages in strategic interactions conditioned on its predictive confidence. To optimize this game, we develop an equilibrium-seeking game-theoretic solver that leverages trajectory-sampled counterfactual regret minimization to iteratively refine each agents decisions and drive competing agents toward a coarse correlated equilibrium. Extensive experiments across multiple benchmarks on MER-UniBench show that MAGIC-MER effectively mitigates emotional bias inherent in individual LMMs and achieves state-of-the-art (SOTA) performance. The code is available at https://github.com/AKAThoka/Magic-MER

Abstract:
Open-Vocabulary Object Detection (OVOD) aims to localize and recognize objects beyond predefined training categories by leveraging the shared semantic space of VisionLanguage Models (VLMs). Although recent VLM-based pseudo-labeling approaches have shown promise for OVOD, existing methods still suffer from two critical limitations: (i) the pseudo-labeled regions are typically generated by the Region Proposal Network (RPN) trained on base categories with annotations, limiting proposal diversity and generalization to novel classes; (ii) novel regions lack direct semantic supervision with annotations, hindering the learning of discriminative representations. To alleviate these limitations, we propose a novel method, termed Foreground-generalized RPN with Language-guided visual augmentation for open-vocabulary object Detection (FLDet). For limitation (i), we propose a Foreground-Generalized RPN (FG-RPN) trained with additional foreground cues from Grounding DINO to generate more diverse proposals, improving novel-region coverage in pseudo-label generation, and enhancing generalization to unseen categories. To address limitation (ii), we introduce a Language-Guided Visual Augmentation (LGVA) module that injects coarse- and fine-grained semantic information from an LLM into region-level visual features, strengthening the detectors semantic learning through knowledge distillation. Extensive experiments on OVOD benchmarks demonstrate that our proposed method achieves SOTA performance, i.e., 41.8% (+2.9%) in novel AP50 on OV-COCO and 26.3% (+2.1%) in rare AP on OV-LVIS. The code is publicly available at https://github.com/llqy123/FLDet.

Abstract:
Recent advances in video-to-audio (V2A) generation have enabled high-quality audio synthesis from visual content, whereas achieving robust and fine-grained controllability remains a fundamental challenge. In particular, existing methods suffer from two key limitations: weak textual controllability under visual-text semantic conflict, and imprecise stylistic control due to entangled temporal and timbre information in reference audio. Moreover, the lack of standardized benchmarks further hinders systematic evaluation of controllability. In this paper, we propose ControlFoley, a unified and controllable multimodal V2A framework that enables precise control across video, text, and reference audio. First, we introduce a joint visual encoding paradigm that integrates CLIP with a spatio-temporal audio-visual encoder, enhancing both audio-visual alignment and textual controllability under cross-modal conflict. Second, we propose a temporal-timbre decoupling strategy that suppresses redundant temporal information in reference audio while preserving discriminative timbre features, enabling accurate and interference-free stylistic control. Third, we design a modality-robust training scheme with unified multimodal representation alignment (REPA) and random modality dropout, aligning generated audio representations with aggregated multimodal conditions. Finally, we present VGGSound-TVC, the first benchmark specifically designed to quantify textual controllability under varying degrees of visual-text semantic conflict. Extensive experiments demonstrate that ControlFoley achieves state-of-the-art performance across multiple V2A tasks, including text-guided, text-controlled, and audio-controlled generation, outperforming both task-specific and unified baselines. Furthermore, ControlFoley exhibits superior controllability under cross-modal conflict while maintaining strong temporal synchronization and audio quality, establishing a unified framework for controllable V2A generation. Codes and full paper are available at: \urlhttps://github.com/xiaomi-research/controlfoley.

Abstract:
Large Audio-Language Models (LALMs) have set new benchmarks in speech processing, yet their deployment is hindered by the memory footprint of the Key-Value (KV) cache during long-context inference. While general KV cache compression techniques excel in LLMs, they often fail in the audio domain by overlooking the intrinsic temporal continuity of acoustic signals. To bridge this gap, we propose AudioKV, a novel framework that robustly prioritizes audio-critical attention heads through a hardware-friendly semantic-acoustic alignment mechanism. Specifically, we identify these modality-specialized heads by analyzing attention scores in ASR tasks and dynamically allocate KV cache budgets preferentially to them. Furthermore, we introduce Spectral Score Smoothing (SSS), an FFT-based global filtering strategy designed to suppress high-frequency noise and recover smooth global trends from importance scores, ensuring more balanced token selection with unprecedented precision. Extensive evaluations across multiple LALMs, including the Qwen and Gemma series, demonstrate that AudioKV significantly outperforms baselines while enhancing computational efficiency. Notably, at a 40% compression ratio, AudioKV maintains near-full accuracy on Qwen3-Omni-30B with only a 0.45% drop, whereas traditional methods suffer from catastrophic performance degradation and repetition. Our code is publicly available at \urlhttps://github.com/yxwang1215/Audio_kvcache.

Abstract:
The proliferation of fake news on social media motivates the development of multimodal fake news detection. While most existing methods focus on various fusion strategies, they can still yield unreliable results caused by epistemic and aleatoric uncertainties, which can arise from the restricted contextual horizon of isolated sample pairs or inherent noise in unimodal data. To address this, we propose a novel Decoupling Dual-level Uncertainties (DDU) framework for reliable multimodal fake news detection. First, for epistemic uncertainty, we retrieve semantically similar instances along with their veracity labels, distilling them into dynamic prompts to anchor and calibrate the perception of semantic relationships. Second, for aleatoric uncertainty, we employ a Mixture-of-Experts architecture to quantify feature-level uncertainty. A novel uncertainty-aware routing mechanism is then introduced to dynamically down-weight features with high uncertainty, ensuring reliable multimodal fusion. Extensive experiments are conducted on three real-world datasets spanning two languages, demonstrating significant performance improvements of our method. The code is available at https://github.com/njustkmg/MM26-DDU.

Abstract:
Recent advancements in image animation have utilized diffusion models to breathe life into static images. However, existing controllable frameworks typically rely on Lagrangian motion guidance, where optical flow is estimated relative to the initial frame. This paper revisits the same optical-flow primitive through a more local supervision design: we use adjacent-frame Eulerian motion fields to guide generation, where the motion signal always describes a short temporal hop. This shift enables parallelized training and provides bounded-error supervision throughout the generation process. To mitigate the drift artifacts common in adjacent frame generation, we introduce a Bidirectional Geometric Consistency mechanism, which computes a forward-backward cycle check to mathematically identify and mask occluded regions, preventing the model from learning incorrect warping objectives. Extensive experiments demonstrate that our approach accelerates training, preserves temporal coherence, and reduces dynamic artifacts compared to reference-based baselines. The code, model, and data have been made available at nguyentthong.github.io/eulerian.

Abstract:
Predicting spatial gene expression from histopathology images enables large-scale transcriptomic profiling without the cost of direct measurement. Existing methods decode the target gene set as a flat, unstructured vector, ignoring the inter-gene dependencies arising from shared biological pathways and regulatory programs. Without explicit structural guidance, models must infer these dependencies entirely from limited paired data, constraining prediction quality. We propose MSGR (Multi-Scale Gene Refiner), which bridges this gap by incorporating the Gene Ontology (GO), a curated functional hierarchy of genes, as an explicit structural prior. MSGR organizes target genes into a four-level GO tree. Its GO-guided decoder then progressively refines predictions from coarse functional domains to fine individual genes via residual corrections under scale-weighted supervision. Operating solely on the gene side, the GO-guided decoder serves as a seamless plug-in replacement that consistently improves existing architectures without requiring any image-side modifications. Extensive experiments on nine datasets from the HEST-1k benchmark provide empirical evidence for two central claims: GO-structured decoding consistently outperforms flat decoding, even against a state-of-the-art generative baseline, and the gain is attributable to biological ontology structure rather than hierarchical decomposition per se, as confirmed by a +0.027 margin over a structurally equivalent random hierarchy.

Abstract:
Precision molecular design aims to discover personalized drug candidates through joint control of multiple conditions, such as biological relevance and molecular design strategies. Biological relevance reflects cellular functional states under disease or perturbation conditions, while molecular design strategies provide complementary guidance in terms of structural intentions and property optimization. In this study, we propose JoPMol, a jointly controlled precision molecular generative model that integrates biological states encoded by gene expression profiles with molecular structure information expressed in text, and chemical properties quantified by numerical values within a unified modeling framework. This formulation enables coordinated generation and optimization of candidate molecules under joint condition control while preserving consistency among heterogeneous conditioning signals throughout the generation process. Experimental results show that JoPMol outperforms state-of-the-art methods across multiple evaluation metrics. Moreover, JoPMol demonstrates strong generalization ability in both transfer tasks and biologically grounded simulation scenarios, validating its effectiveness for precision molecular design. The source code is publicly available at https://github.com/hala-yh/JoPMol.

Abstract:
RGB-Thermal (RGBT) object detection enables robust perception in complex scenes by leveraging the complementary strengths of visible textures and thermal cues. However, existing methods mainly rely on dense cross-modal interactions over full-resolution features, which inevitably introduce background interference and hinder the learning of target-relevant representations. In this paper, we propose the Prototype HyperGraph Fusion Network (ProtoHGF-Net), a novel framework that redefines cross-modal fusion as prototype-level semantic interaction rather than the dense cross-modal interaction paradigm. Specifically, we design Prototype HyperGraph Fusion to perform cross-modal interaction in a compact prototype-level semantic space. This design enables more selective fusion among target-relevant prototypes. To support this prototype-level fusion, we propose Teacher-Mask Calibration Distillation, which calibrates modality features before fusion using modality-specific teachers and target-aware masks. This strategy suppresses backgrou- nd-dominant responses and produces more target-focused features. Extensive experiments on DroneVehicle, DVTOD, and FLIR demonstrate that ProtoHGF-Net achieves state-of-the-art performance with 85.9% mAP_50, 88.2% mAP_50, and 79.1% mAP_50, respectively.

Abstract:
Lensless imaging has emerged as a promising paradigm for ultra-compact, low-cost, and privacy-preserving visual sensing. However, extending recent advances in 2D lensless reconstruction to 3D novel view synthesis remains challenging. Existing methods mainly rely on neural-radiance-field-based representations, which suffer from slow rendering and limited practicality. In this work, we take the first step toward high-fidelity lensless novel view synthesis using 3D Gaussian Splatting (3DGS). This setting poses three major challenges: maintaining cross-view geometric consistency, suppressing ghosting in occluded regions, and restoring high-frequency textures under efficient optimization. To address these issues, we propose \underlineUncertainty-\underlineAware \underlineConsensus 3D \underlineGaussian \underlineSplatting (UACGS), a framework that combines structure-aware consensus preprocessing with mask-guided optimization. Specifically, our framework incorporates: (1) a structure-aware consensus module that leverages adjacent camera poses to filter out occlusion-induced erroneous projections; (2) an uncertainty-aware mask-guided supervision mechanism that effectively corrects reconstruction errors caused by physical limitations; and (3) a contextual loss design that restores fine-grained high-frequency textures while preserving geometric stability. Extensive experiments show that our framework achieves superior reconstruction quality across datasets, while also delivering strong cross-dataset generalization, real-time rendering, and highly efficient training. The code is publicly available at \urlhttps://github.com/YUNYU7472/UACGS.

Abstract:
World action models inherit the predictive capability of world models, enabling action generation to be guided by anticipated future observations. However, they rely primarily on vision and often fail in contact-rich manipulation, where critical cues arise from physical interaction. In this paper, we propose Dream-Tac, a unified Tactile-World Action Model that jointly models actions, future visual observations, and tactile dynamics. Specifically, Dream-Tac introduces (i) contact-gated visuotactile fusion to selectively integrate tactile signals and (ii) a contact-aware attention bias to better regulate cross-modal interactions during manipulation. To support real-time deployment, we further design a dual-level acceleration strategy, reformulating the contact-aware bias to preserve the fused attention path during training and introducing cache-based diffusion acceleration at inference, achieving up to 2.9× faster training and 1.8× faster inference. Across six contact-rich manipulation tasks, Dream-Tac improves action accuracy by 31.7% on average, demonstrating the effectiveness of unified visuotactile world modeling.Code is available at \urlhttps://github.com/LYFCLOUDFAN/Dream-Tac.

Abstract:
Diffusion Transformer (DiT)-based text-to-image models have recently achieved remarkable success in image synthesis. However, when directly applied beyond their training resolution, these models often suffer from severe semantic degradation, where meaningful semantic content disappears. Although existing training-free high-resolution generation methods can mitigate this issue, the inherent semantic degradation at high resolutions remains largely unresolved, leading to degraded fine-grained details. Furthermore, widely used structural guidance methods lack modeling of the structure and details in the intermediate trajectory, which degrades the quality of the generated results. In this paper, we propose a training-free framework for high-resolution image generation to address these challenges. First, we propose a textvisual scale-decoupled attention (TVSDA) module, which restores the models ability to generate semantically meaningful content at high resolutions by preserving the global receptive field of text tokens while aligning the receptive field of image tokens with the training resolution. Second, we propose a structuredetail guidance (SDG) strategy that explicitly models structural information and fine details through scale-wise and temporal-wise guidance, overcoming the limitations of conventional low-resolution guidance. Extensive experiments demonstrate that our method significantly improves both structural coherence and detail fidelity over existing training-free DiT-based high-resolution generation approaches. The code is available at \urlhttps://github.com/ljjcoder/Restoring-Semantic-Fidelity.

Abstract:
Infrared image super-resolution (IISR) mitigates the limitations imposed by low spatial resolution. Existing methods have recognized that IISR should preserve consistency in global distribution and structural information while enhancing image clarity. However, these methods are either insufficient or overly intrusive, a problem that becomes even more pronounced in diffusion-based models. To address these issues, we propose a dual-path diffusion-based framework for IISR, termed Shift-IISR. The proposed method is designed to improve the consistency of IISR results while preserving the generative capacity of diffusion models. Specifically, we develop a Global Representation Modulation (GRM) module to extract modality-specific information from infrared imagery and guide the global distribution of the diffusion model toward the ground truth. In addition, we introduce a Local Structure Refinement (LSR) module to encourage the model to focus on structural information at each step of the iterative denoising process. Extensive experiments demonstrate that the proposed method effectively improves distributional and structural consistency while maintaining competitive super-resolution performance. The source code of the proposed Shift-IISR can be available at https://github.com/Assassink8/Shift-IISR.

Abstract:
Extracting the photo-response non-uniformity (PRNU) fingerprints from images is a promising technique for source camera identification (SCI), a fundamental task in multimedia forensics. However, current state-of-the-art PRNU extractors built on CNN architectures remain sensitive to semantic noise from non?flat images and perform poorly when the number of available images is limited (e.g., fewer than ten). To fill this gap, we propose HF-Printor, a high-frequency enhanced camera fingerprint extractor for practical SCI. Our method is motivated by the fact that Transformer-based backbones are strong at suppressing complex scene semantics but weak at preserving PRNU-related high-frequency cues. We therefore retain their semantic modeling strength while explicitly enhancing high-frequency fingerprint extraction. Based on a Swin Transformer backbone with strong semantic modeling capability, HF-Printor incorporates patch-level fingerprint contrastive learning, high-frequency enhancement, and frequency-domain fingerprint regularization to improve the discriminability, stability, and high-frequency sensitivity of the extracted fingerprints. To evaluate the performance of the proposed HF-Printor, extensive validation experiments were conducted on the VISION and Dresden datasets, demonstrating that HF-Printor achieves over 90% closed-set identification accuracy with only three ordinary reference images, and consistently outperforms competing methods under practical limited-reference settings. Our code is available at \urlhttps://github.com/Hillthon/HF-Printor.

Abstract:
Multimodal brain networks integrate complementary structural and functional information to enhance the sensitivity of disease-specific features, thereby improving early diagnosis performance. However, existing methods typically model brain networks with predefined structures, which fail to capture task-specific interaction patterns, thereby limiting diagnostic performance and model explainability. To address this, we propose a Cross-Modal Relational Graph Learning (C-MRL) framework that transforms cross-modal connections into a learnable objective and employs a specially designed GNN learner to achieve bidirectional cross-modal interaction, enabling the construction of task-adaptive brain networks. C-MRL not only achieves superior performance but also provides new insights into the synergy of multimodal data. Specifically, C-MRL introduces two main designs: first, it constructs a Learnable Cross-modal Relational Graph (LCRG) based on neuroanatomical priors, where bidirectional cross-modal interactions are realized through two complementary directed acyclic graphs, each capturing unidirectional nonlinear relationships between brain regions. Second, based on the learned directed graph, we develop an Attentive Hierarchical Multimodal Fusion Learner (AHFL) that performs bidirectional message propagation and aggregation, outputting task-discriminating graph representations that integrate subject-specific multimodal information. Extensive experiments on four benchmark datasets demonstrate the superiority of C-MRL. In particular, it achieves state-of-the-art performance on brain disease classification, with the full model reaching 92.5%1.8 accuracy for Alzheimer's disease and exhibiting robust cross-dataset generalization (91.4%1.1 accuracy on the unseen AIBL dataset), substantially improving over existing methods. Our code is available at \urlhttps://github.com/LiuFei-AHU/CMRL.

Abstract:
Vision-language models (VLMs) have shown strong performance on static visual understanding, yet they still struggle with dynamic spatial reasoning that requires imagining how scenes evolve under egocentric motion. Recent efforts address this limitation either by scaling spatial supervision with synthetic data or by coupling VLMs with world models at inference time. However, the former often lacks explicit modeling of motion-conditioned state transitions, while the latter incurs substantial computational overhead. In this work, we propose WorldVLM, a training framework that distills spatial imagination from a generative world model into a vision-language model. Given an initial observation and a parameterized camera trajectory, we use a view-consistent world model to synthesize geometrically aligned future views and derive structured supervision for both forward (action-to-outcome) and inverse (outcome-to-action) spatial reasoning. We post-train the VLM with a two-stage recipe on a compact dataset generated by this pipeline and evaluate it on multiple spatial reasoning benchmarks. WorldVLM delivers consistent improvements over the base model across diverse benchmarks, including SAT-Real, SAT-Synthesized, VSI-Bench, and MindCube. It also outperforms the test-time world-model-coupled methods while eliminating the need for expensive inference-time generation. Our results suggest that world models can serve not only as inference-time tools, but also as effective training-time teachers, enabling VLMs to internalize spatial imagination in a scalable and efficient manner.

Abstract:
Video Large Language Models (Video-LLMs) have achieved strong performance on many video understanding tasks, yet remain vulnerable to temporal hallucination in time-sensitive reasoning scenarios such as scene transition understanding and event ordering. Existing solutions either require costly retraining or introduce inefficient multi-pass inference and external supervision at test time. In this work, we show that temporal evidence is already encoded in Video-LLM hidden representations but is not fully exploited during standard inference. Motivated by this observation, we propose Single-Pass Orthogonal Decomposition (SPOD), a training-free framework that recovers suppressed temporal evidence through analytical hidden-state manipulation in a single forward pass. SPOD consists of two components: Frame Distinctiveness Analysis (FDA), which identifies salient frames from middle-layer hidden-state geometry, and Temporal Delta Subspace (TDS), which constructs orthogonal temporal and prior subspaces from weighted inter-frame hidden-state differences. By decomposing final hidden states into temporal, prior, and residual components and rescaling them independently, SPOD enhances dynamic information, suppresses static appearance bias, and preserves complementary information. Experiments on VidHalluc, EventHallusion, and VideoMMMU show that SPOD consistently outperforms strong inference-time baselines across temporal reasoning tasks and model architectures while retaining near-vanilla inference efficiency. These results suggest that improving temporal reasoning in Video-LLMs may depend not only on additional supervision, but also on recovering temporal evidence already present in their internal representations. Our code is available at https://github.com/VioAgnes/SPOD.

Abstract:
Automatic mispronunciation detection and diagnosis (MDD) is a crucial component of computer-assisted language learning systems. However, adapting existing MDD paradigms to Mandarin faces critical challenges: the severe scarcity of Mandarin as a second language data and the lack of interpretability. To address these limitations, we leverage Mandarin's rigorous bottom-up phonological hierarchy, which provides a strong prior knowledge of phonological compositional structures and physical articulatory mechanisms, and propose the phonological prior-infused MDD (PPI-MDD) method. Specifically, we introduce a phonological representation learning module to map continuous speech onto physically meaningful articulatory representations and improve fine-grained interpretability at the underlying level. Furthermore, we design a phonology-based dual-stream network to explicitly model phonological structures, which effectively constrains the hypothesis space in low-resource scenarios. Finally, we propose a training-free text-fusion decoding algorithm based on the goodness of pronunciation to preserve explicit decision logic while effectively utilizing multimodal information, which also enables dynamic adjustment of MDD strictness. Extensive experiments show that PPI-MDD outperforms state-of-the-art methods with a 10.6% relative F1-score improvement without text prompts, and a 33.4% relative F1-score improvement with text prompts, while delivering multi-level and fine-grained diagnostic feedback. The implementation code is available at \urlhttps://github.com/BNU-ERC-ITEA/PPI-MDD.

Abstract:
Multi-modal medical image segmentation leverages complementary diagnostic information, yet fusion can underperform single-modality baselines when spatially aligned inputs differ in quality. Here, "corruption" primarily denotes resolution-induced degradation rather than misalignment or complete modality absence, while synthetic noise is evaluated only as an auxiliary setting. We identify a critical optimization-inference inconsistency: degraded modalities can receive weak training updates yet substantially affect predictions, indicating active interference with fusion. We attribute this failure to resampling-induced feature corruption and optimization bias, where noisy features propagate through skip connections and encourage unreliable modality selection. We therefore propose CoReFuse-Med, a Corruption-aware Rebalanced Fusion framework that suppresses corruption during feature transmission and rebalances modality contributions during high-level fusion. Experiments on EPVS, BraTS, and WMH, including multiple Z-axis slice-retention ratios and an auxiliary noise test, demonstrate improved accuracy and robustness under modality-quality discrepancies. Our code is available at https://github.com/lrever/CoReFuse.

Abstract:
Closed-loop simulation is central to autonomous driving (AD) safety evaluation, yet it remains difficult to turn natural-language requests into valid 3D test scenarios. Existing language-driven methods either operate in pixel space without explicit scene instantiation, or rely on fixed asset banks and coarse object placement that are insufficient for open-world, test-oriented scene construction. We present LATTE, a language-driven framework for constructing explicit 3D driving scenes from open-world scene requests. LATTE decomposes the problem into two coupled parts: Image-Grounded Asset Construction (IGAC), which acquires insertion-ready 3D assets for missing objects from visual evidence, and Structurally Coherent Layout Optimization (SCLO), which turns language-guided initial placements into simulation-suitable layouts under physical and semantic constraints. The resulting edits are written back to a real-scene-reconstructed simulator and evaluated through closed-loop AD rollouts. Experiments show that LATTE not only supports open-world asset insertion and more coherent scene layout, but also produces behaviorally effective test cases that measurably alter downstream AD performance. The project page is at \hrefhttps://zaiyihu.github.io/LATTE-ProjectPage/here.

Abstract:
Camera-controllable video generation aims to synthesize videos with flexible and physically plausible camera movements. However, existing methods either provide imprecise camera control from text prompts or rely on labor-intensive manual camera trajectory parameters, limiting their use in automated scenarios. To address these issues, we propose a novel Vision-Language-Camera model, termed CT-1 (Camera Transformer 1), a specialized model designed to transfer spatial reasoning knowledge to video generation by accurately estimating camera trajectories. Built upon vision-language modules and a Diffusion Transformer model, CT-1 employs a Wavelet-based Regularization Loss in the frequency domain to effectively learn complex camera trajectory distributions. These trajectories are integrated into a video diffusion model to enable spatially aware camera control that aligns with user intentions. To facilitate the training of CT-1, we design a dedicated data curation pipeline and construct CT-200K, a large-scale dataset containing over 47M frames. Experimental results demonstrate that our framework successfully bridges the gap between spatial reasoning and video synthesis, yielding faithful and high-quality camera-controllable videos and improving camera control accuracy by 25.7% over prior methods. The project page and code are available at: https://gulucaptain.github.io/Camera-Transformer-1/.

Abstract:
\beginabstract Semantic segmentation in 3D Gaussian Splatting (3DGS) is crucial for advancing 3D scene understanding. Existing methods predominantly rely on feature distillation, which incurs substantial per-scene training overhead and often yields blurred segmentation boundaries. We identify that these boundary artifacts are driven in part by insufficient viewpoint coverage and boundary overflow of anisotropic Gaussian primitives. To address these challenges, we propose VCAR, a training-free coarse-to-fine segmentation strategy based on View Completeness and Axis-aware Boundary Refinement. In the coarse stage, a visibility-based weighted multi-view voting scheme rapidly localizes the target. In the fine stage, an object-centric sphere derived from the coarse result generates supplementary viewpoints via Spherical Spiral Sampling (SSS), allowing multi-view voting on the augmented views to precisely refine object boundaries and suppress irrelevant 3D Gaussians. Moreover, we introduce Axis-aware Boundary Refinement (ABR) to mitigate artifacts from anisotropic primitives. By decomposing the projected 2D covariance into per-axis contributions, ABR identifies the dominant axis responsible for boundary leakage and applies targeted anisotropic compression exclusively along that axis. Extensive experiments on NVOS and LERF demonstrate that VCAR achieves state-of-the-art segmentation accuracy and efficiency without training. Our code is available at \urlhttps://github.com/DDKK0526/VCAR. \endabstract

Abstract:
Text-to-image synthesis for remote sensing (RS) lacks an accessible, high-performance generative foundation, as directly training diffusion models at large, high resolutions is computationally prohibitive. Training-free resolution promotion via Rotary Position Embedding (RoPE) extrapolation offers an efficient alternative, but existing algorithms apply static scaling rules tailored to natural scenes, whereas RS imagery is dominated by dense, minute instances that hinge on high-frequency structural integrity. We present a comprehensive framework for large-scale RS synthesis. First, we curate over 100,000 image-text pairs to train RS-FLUX, a domain-specialized generative prior. Second, we propose SHARP, a training-free extrapolation algorithm that introduces a Rational Decay Scheduler to modulate RoPE frequencies throughout denoising: strong positional extrapolation early on enforces coherent global layouts, and progressive relaxation later recovers dense high-frequency details. Extensive experiments show that SHARP consistently achieves state-of-the-art performance across multiple promoted resolutions with negligible (<4%) overhead, offering an efficient and structurally faithful solution for large-scale RS generation.

Abstract:
Camera control is essential for cinematic video generation, dictating shot composition and viewer perception. However, existing datasets are largely limited to basic camera translation and rotation. They lack support for expressive cinematic trajectories coordinated with dynamic optical control, and also suffer from limited scene and character diversity. These limitations hinder current models from reproducing professional cinematographic effects and generalizing to real-world applications. To address these gaps, we propose Hitchcock-300K, a large-scale, master-level dataset designed for camera-controlled video generation and editing. Built with Unreal Engine 5, the dataset contains 300,000 high-quality video-editing pairs, each accompanied by camera intrinsics, extrinsics, and a text editing instruction. Hitchcock-300K offers three main advantages: expressive camera motions, including complex cinematic trajectories; dynamic optical control, supporting effects such as rack focus and Hitchcock zoom; and broad visual diversity, comprising 1,200+ scenes and 250+ characters. Building on Hitchcock-300K, we further train a large-scale text-conditioned video-editing model. Experimental results show that models trained on Hitchcock-300K significantly outperform existing baselines across a range of camera editing tasks and exhibit promising generalization to real-world videos. Our dataset, model, and empirical findings provide a strong foundation for camera-controlled video generation:https://hitchcock300k.github.io/Hitchcock-300K.

Abstract:
Electroencephalography (EEG) is a promising neural signal for mental disorder assessment. However, existing EEG datasets mainly focus on single-disorder settings, limiting their use in realistic mental disorder differential diagnosis. To address this gap, we present a large-scale three-electrode EEG dataset for multi-disorder analysis, covering depression, anxiety and schizophrenia, with all labels clinically confirmed by professional psychiatrists. The dataset contains 729 subjects with both resting-state and stimulation-state EEG recordings, and enables flexible differential diagnosis tasks under arbitrary disease combinations. On top of this benchmark, we propose a prior-guided self-supervised pretraining framework for mental disorder differential diagnosis. By introducing prior-guided augmentation and prior-guided masking, the proposed method exploits physiological priors to mine richer disorder-discriminative cues from limited EEG electrodes. Extensive experiments verify the effectiveness of the proposed method. The dataset and code are publicly available at https://wushidiyishenqing.github.io/dataset-webpage/.

Abstract:
Current visible-light remote sensing dehazing (VLRSD) evaluation assumes improved visual quality guarantees better downstream performance. However, our experiments reveal a critical human-machine misalignment, where perceptual gains poorly correlate with machine task accuracy. To systematically investigate this, we introduce a comprehensive benchmark suite of 16 hazy datasets generated via our Task-aware Physically Constrained Adversarial (TPCA) framework. The suite provides full annotations for oriented object detection on DOTA and scene classification on AID. Guided by downstream models, the benchmarks center on two 1,000-sample diagnostic scenarios: StealthSet, featuring visually subtle yet feature-destructive haze, and FeatureSet, containing visually degraded yet task-preserving haze, alongside synthetic and diffusion baselines. Evaluating 13 SOTA methods demonstrates that high-fidelity restoration consistently fails to recover semantics in StealthSet and counterintuitively degrades task performance in FeatureSet. These benchmarks advocate a paradigm shift from pixel fidelity to task utility. Datasets and code are available \hrefhttps://github.com/Shan-rs/Diagnostic-Benchmark-for-VLRSD.git\textcolorACMBluehere.

Abstract:
Large multimodal models (LMMs) have demonstrated outstanding capabilities in various visual perception tasks, which has in turn made the evaluation of LMMs significant. However, the capability of video aesthetic quality assessment, which is a fundamental ability for human, remains underexplored for LMMs. To address this, we introduce VideoAesBench, a comprehensive benchmark for evaluating LMMs understanding of video aesthetic quality. VideoAesBench has several significant characteristics: (1) Diverse content including 1,804 videos from multiple video sources including user-generated (UGC), AI-generated (AIGC), compressed, robotic-generated (RGC), and game videos. (2) Multiple question formats containing traditional single-choice questions, multi-choice questions, True or False questions, and a novel open-ended questions for video aesthetics description. (3) Holistic video aesthetics dimensions including visual form related questions from 5 aspects, visual style related questions from 4 aspects, and visual affectiveness questions from 3 aspects. Based on VideoAesBench, we benchmark 23 open-source and commercial large multimodal models. Our findings show that current LMMs only contain basic video aesthetics perception ability, their performance remains incomplete and imprecise. We hope our VideoAesBench can be served as a strong testbed for explainable video aesthetics assessment. The proposed VideoAesBench is available at: https://huggingface.co/datasets/chris-123/VideoAesBench.

Abstract:
Dairy goats have attracted increasing attention recently due to their nutritional and economic importance, as well as their growing role in global livestock production. Meanwhile, computer vision and intelligent multimedia analysis have shown strong potential for improving dairy goat management by enabling automated monitoring, behavior understanding, and fine-grained decision support in practical farming environments. However, existing dairy goat datasets are typically small in scale, task-isolated, and collected under constrained conditions, making them insufficient for intelligent multimedia analysis in realistic farming environments. To address this gap, we present \confset, a large-scale benchmark suite for dairy goat vision, organized as multiple task-specific annotations under a standardized benchmark framework, containing 32,007 images and 1,459 videos collected from authentic farm environments, covering diverse conditions of illumination, occlusion, and inter-animal interaction. The benchmark supports eight representative tasks, including object detection, instance segmentation, semantic segmentation, object tracking, pose estimation, behavior recognition, individual identification, and image generation; it provides task-specific annotations for diverse dairy goat visual understanding scenarios. We further establish benchmark settings and report baseline results on six supervised tasks using representative models, confirming the practical application challenges and benchmark value of \confset, which provides a valuable resource for advancing future research in precision livestock farming, dairy goat scene understanding, and intelligent multimedia analysis. \confset\ is available: \urlhttps://y-f-wei.github.io/DGV-Suite/

Abstract:
Many multimedia language-generation tasks require grounding in both visible content and evolving, partially observable structured states. Professional esports commentary exemplifies this: in StarCraft: Brood War, meaningful commentary relies on short-term temporal context, observer camera motion, and observer-visible unit configurations. We introduce CASTER (Commentary Alignment and State Tracking for Esports Research), a multimodal dataset for observation-grounded commentary in partially observable environments. It aligns utterance-level observer-view clips, structured states from replay logs and viewport traces, and ASR+LLM-refined, manually verified commentary. CASTER comprises 239 matches and 28,734 aligned clipstatecommentary instances with strict game-level splits. We define two tasks for reproducible evaluation: (1) Clip-to-Observation, predicting structured observer-state events from clips, and (2) Clip+Observation-to-Commentary, generating commentary conditioned on synchronized visual and structured inputs. We detail the synchronization pipeline, report a 6,000-sample manual audit of alignment quality, and provide zero-shot baselines. Results indicate that recovering structured state from clips alone remains challenging, while Task 2 establishes reference points for commentary generation under various conditionings. The dataset, code, and benchmark are publicly available at https://kimd0.github.io/CASTER/.

Abstract:
Sycophantic hallucinations refer to the tendency of large language models (LLMs) to generate hallucinated responses by excessively aligning with human preferences. Unlike conventional hallucinations caused by missing knowledge, sycophantic hallucinations are harder to prevent. Even when the model already contains the relevant knowledge and is given explicit context, it may still produce plausible but incorrect answers. Existing datasets for sycophantic hallucination typically lack diverse induction settings and broad evaluation across multiple domains. To address this gap, we introduce SycoQA to evaluate sycophantic hallucinations across a broader spectrum of capability dimensions. SycoQA consists of a core and an extension subset to evaluate sycophantic hallucinations under internal-knowledge-based and context-grounded settings, respectively. Specifically, the Core subset adopts multiple induction paradigms with progressively increasing intensity, while the Extension subset constructs data through contextual corruption to simulate the corresponding scenarios. We believe that SycoQA can facilitate further analysis of sycophantic behavior and support comprehensive evaluation of advanced downstream tasks, particularly multi-level induction-aware and context-robust sycophantic hallucination detection. The dataset is publicly available at \urlhttps://github.com/hehebamei/SycoQA.

Abstract:
\beginabstract To fill the gap on benchmark datasets for vision language models (VLM) research on underground mining safety analysis, we introduce UMSBench-v1.0. As the first multimodal VQA benchmark dataset and also the first real-synthetic image dataset on mining safety, it consists of 10,000 real-synthetic underground images and 40,000 question-answer pairs correspondingly, and covers safety reasoning, captioning, understanding and perception over 30 underground mining site types. The synthetic data are generated with newly proposed Image-to-Text-to-Image(I2T2I)mechanism based on refined LLM-driven captioning capabilities, prompting engineering and visual content generation models. Moreover, benchmarking tasks are reformulized as Multiple-Choice Questions(MCQs) with advanced VLMs and human experts as references to archive higher flexibility, scalability and efficiency. State-of-the-art vision language models are benchmarked and finetuned to demonstrate the dataset's effectiveness and potential application. The dataset will be released at \urlhttps://github.com/BCU-AILAB/UMSBench. \endabstract

Abstract:
Large-scale visual instruction tuning (VIT) has become a key paradigm for advancing the performance of vision-language models (VLMs) across various multimodal tasks. However, training on the large-scale datasets is computationally expensive and inefficient due to redundancy in the data, which motivates the need for multimodal data selection to improve training efficiency. Existing data selection methods for VIT either require costly training or gradient computation. Training-free alternatives often depend on proxy models or datasets, instruction-agnostic representations, and pairwise similarity with quadratic complexity, limiting scalability and representation fidelity. In this work, we propose ScalSelect, a scalable training-free multimodal data selection method with linear-time complexity with respect to the number of samples, eliminating the need for external models or auxiliary datasets. ScalSelect first constructs sample representations by extracting visual features most attended by instruction tokens in the target VLM, capturing instruction-relevant information. It then identifies samples whose representations best approximate the dominant subspace of the full dataset representations, enabling scalable importance scoring without pairwise comparisons. Extensive experiments across multiple VLMs, datasets, and selection budgets demonstrate that ScalSelect achieves over 97.5% of the performance of training on the full dataset using only 16% of the data, and even outperforms full-data training in some settings. The code is available at https://github.com/ChangtiWu/ScalSelect.

Abstract:
Understanding humor in videos remains an underexplored challenge for Multimodal Large Language Models (MLLMs). It demands the integration of fine-grained multimodal perception, temporal narrative reasoning, and sociocultural knowledge, which existing works have not fully explored. Consequently, the absence of high-quality, targeted datasets has severely hindered quantitative evaluation and model development in this domain. To address this gap, we introduce ComedyQA, a comprehensive benchmark for video humor understanding. It comprises 2,526 comedy video clips and 12.6K fine-grained annotations covering multimodal humor cues, relevant plot context, external knowledge, and comprehensive humor explanations. It supports three tasks of increasing complexity: (1) Multimodal Humor Classification, (2) Multimodal Humor Grounding, and (3) Comprehensive Humor Explanation. Evaluations show that recent MLLMs still struggle with comprehensive video humor understanding. Our analysis identifies three key failure modes: visual humor misattribution, cultural unawareness, and contextual counter-intuitive reasoning deficits. Motivated by these findings, we propose VideoHumorAgent, a multi-agent framework that addresses each failure mode with specialized agents. Experiments demonstrate consistent improvements across all tested models. Our dataset and codes are released at https://github.com/wangxueyan666/ComedyQA.

Abstract:
Large language models and large multimodal models (LLMs and LMMs) deliver strong generative performance but suffer from slow decoding, a problem that becomes more severe when handling visual inputs, whose sequences typically contain many more tokens with lower information density than text. Speculative decoding accelerates LLM inference by letting a compact draft model propose candidate tokens that are selectively accepted by a larger target model, achieving speed-up without degrading quality. However, existing multimodal speculative decoding approaches largely ignore the structural characteristics of visual representations and usually rely on text-only draft models. In this paper, we introduce SpecFLASH, a speculative decoding framework tailored to LMMs that explicitly exploits multimodal structure when designing the draft model. We first mitigate redundancy in visual token sequences with a lightweight, latent-guided token compression module that compacts visual features while preserving semantics, and then leverage the co-occurrence and local correlations of visual entities via a semi-autoregressive decoding scheme that predicts multiple tokens in a single forward pass. Extensive experiments demonstrate that SpecFLASH consistently surpasses prior speculative decoding baselines, achieving up to 2.68× speed-up on video captioning and 2.55× on visual instruction tuning, relative to the original LMM.

Abstract:
Few-Shot Segmentation (FSS) aims to segment unseen categories using a few annotated examples. While most existing methods rely heavily on visual prototypes, they often suffer from matching instability and prototype bias under significant intra-class variations. Recent text-driven approaches attempt to incorporate semantic guidance, yet they typically integrate it in an indirect and loosely coupled manner, leaving the intrinsic gap between multimodal representation spaces unresolved. To address these limitations, this paper proposes SAM3-VLFS, a visionlanguage collaborative FSS framework built on the Segment Anything Model 3 (SAM3). Specifically, we introduce a dual-stage modality interaction mechanism to advance from implicit alignment to explicit interaction. At the feature level, we design cross-modal adapters to enable bidirectional interaction between visual and textual features, where semantic cues enhance visual representations and visual signals refine textual features for fine-grained dense prediction. At the decision level, we decouple FSS into two complementary pathways: a semantic-guided branch that leverages the inherent promptable segmentation capability of SAM3 to generate masks with strong category-level generalization, and a memory-based visual branch that repurposes the SAM3 tracker to capture instance-specific details through exemplar-driven segmentation. Predictions from both branches are combined through complementary pixel-wise maximum fusion. Extensive experiments on standard benchmarks show our method achieves state-of-the-art performance, with 71.1% and 78.4% 1-shot mIoU on COCO-20^i and PASCAL-5^i, respectively. Our code is available at https://github.com/zhouxinglun/SAM3-VLFS.

Abstract:
Source-Free Domain Adaptive Object Detection (SF-DAOD) aims to transfer detectors to an unlabeled target domain without access to source data. To mitigate noisy pseudo-labels in this setting, existing mean-teacher methods apply parameter smoothing but suffer from slow adaptation and error accumulation. To address two key challenges in pseudo-label learning, namely instantaneous optimization inconsistency and long-term error accumulation, Dynamic Discrepancy-aware Mean Teacher (D2-MT) is proposed to enhance robustness across both parameter and temporal dimensions. First, Dynamic Discrepancy-aware Parameter Grouping (DDPG) explicitly measures gradient discrepancies between the teacher and student at the parameter level, enabling discrepancy-aware updates that balance adaptation and stability. Second, Periodic Teacher-Guided Reset (PTGR) periodically reinitializes the student with EMA teacher weights and retrains it using updated pseudo-labels, effectively correcting accumulated errors and stabilizing optimization. These components are complementary and introduce no additional model complexity. Extensive experiments on three SF-DAOD benchmarks demonstrate that D2-MT consistently outperforms existing methods. Codes are available at: \urlhttps://github.com/zhiwenwang20122302/d2mt.

Abstract:
Knowledge Distillation (KD) offers a promising yet underexplored path for compressing large action recognition models. However, existing KD methods suffer from two key limitations: 1) reliance on fixed input samples leads to suboptimal feature alignment between the frozen teacher (larger model) and the learnable student (smaller model), and 2) applying a uniform distillation strength for all channels fails to account for their varying importance in capturing distinct knowledge (e.g., motion tempo or magnitude) across training epochs. This motivates us to develop an Adaptive Sample-aware Channel-wise Dynamic (ASCD) KD approach, which operates in two stages. First, we use an adaptive sample generation module to create updated samples by incorporating semantics from sample gradients, which are derived by minimizing a feature loss weighted by channel centroid frequency differences at each layer. Meanwhile, crucial motion-related details are preserved by applying a Gaussian mask to frequency features. Second, we employ a channel-wise dynamic distillation module to train student on these generated samples, guided by sample gradients and feature frequencies. For efficiency, samples are updated periodically rather than per epoch. Extensive experiments on three video benchmarks (UCF101, Kinetics-400, Something-Something-v2) and two image datasets (CIFAR-100, ImageNet) demonstrate the state-of-the-art performance of our method. Code is available at https://github.com/mlvccn/ASCD_KD_Action.

Abstract:
The application of Graph Convolutional Networks (GCNs) has significantly advanced skeleton-based action recognition in the field of computer vision in recent years. However, most existing studies overlook the issue of spatio-temporal interference under non-ideal condition in real-world scenarios, which is widespread and primarily manifested as spatial joint anomalies and temporal frame missingness, leading to substantial degradation in recognition performance. To address these challenges, we propose a Spatio-Temporal Complementary Repair Graph Convolutional Network (STCR-GCN), designed to enhance the robustness of skeleton-based action recognition. Specifically, a Static-Dynamic Synergistic Spatial Repair (SDSR) module is proposed, which integrates static posture structure anchoring and dynamic motion rheology perception to respectively preserve spatial structural integrity and motion continuity, thereby achieving spatial-level repair by synergistically compensating for missing joints and suppressing abnormal jitter. Furthermore, a Rhythm-Driven Temporal Repair (RDTR) module is introduced, which constructs rhythmic semantics through rhythmic primitive decomposition and activates rhythm expert compensation pathways for adaptive restoration of diverse action rhythms. Each expert incorporates dual-attention compensation to effectively address rhythm fragmentation and misalignment, thereby accomplishing temporal-level repair. Experimental results demonstrate that STCR-GCN achieves state-of-the-art performance on various non-ideal datasets, while also delivering excellent results on three mainstream benchmark datasets under ideal conditions. All source code are made available at https://github.com/XingLi1012/STCR-GCN.

Abstract:
Video-based social relationship recognition aims to identify the social relationships between characters, which can facilitate the construction of the multimodal knowledge graph and downstream video understanding tasks. Existing methods mainly focus on multimodal fusion and social graph modeling. However, a fundamental limitation of these methods is that they conduct relationship recognition on a single isolated clip, neglecting the rich historical context. Social relationships are inherently complex, information restricted to a single short clip is often insufficient to determine the relationship uniquely. To bridge this gap, we propose the Ambiguity-Aware Memory Network (AAMN), a framework that extends reasoning beyond the current clip by adaptively retrieving historical evidence when local inference is ambiguous. Specifically, we utilize the Jensen-Shannon Divergence to quantify the inconsistency between modalities as an indicator of ambiguity, which also serves as a trigger for memory activation. Furthermore, to handle the computational challenge of long videos, we construct a character-centric compressed memory bank that efficiently stores fine-grained spatiotemporal tokens. Extensive experiments on benchmarks demonstrate that AAMN achieves state-of-the-art performance, confirming the role of global context in reasoning complex social relationships. Our code is available at https://github.com/BoWang222/AAMN.

Abstract:
Curvilinear structure analysis is an important and fundamental task in multimedia. However, the controllable generation of images with precise curvilinear structure objects remains an open challenge. To address this, we propose CSGen, a hierarchical multimodal diffusion model that synthesizes high-fidelity images precisely aligned with multiple control conditions. The CSGen is built upon three key innovations: 1) We construct a multi-domain and multimodal dataset, including over 24K samples from 5 domains and 7 different types of annotations, to train the unified generation model. 2) We propose a novel hierarchical progressive control strategy that decouples topology clues from visual context by a phased signal injection, mitigating semantic drift while ensuring the topological integrity of sparse structures. 3) We design a sparsity-aware loss re-weighting mechanism to address the extreme sparsity of curvilinear structures, significantly enhancing the attention on thin and fragile structures during optimization. Extensive experiments demonstrate that CSGen generates images with superior structure accuracy and visual realism, significantly improving downstream segmentation performance while maintaining robustness across diverse prompts. Our results confirm CSGen as a scalable, data-centric paradigm for the analysis of complex curvilinear structures in diverse multimedia applications. Code and dataset are available at https://github.com/ShanZard/CSGen.

Abstract:
\beginabstract Existing research on object hallucination in multimodal large language models (MLLMs) predominantly attributes the problem to language priors such as over-reliance on textual co-occurrence statistics. We challenge this view by presenting quantitative evidence for a complementary, under-explored cause: \emphvisual-origin hallucination, where hallucinations arise from incorrect visual feature extraction and misalignment between image and text embeddings. Through cosine similarity analysis and Smooth Grad-CAM entropy measurements, we show that hallucinated samples exhibit systematically lower image-text similarity (average 0.158 vs.\ -0.122) and inverted attention patterns, where attention is dispersed when the target object is present but wrongly concentrated when it is absent. Guided by this diagnosis, we propose Adversarial Contrastive Fine-Tuning (ACFT). ACFT uses an Adversarial Hallucination Attribute Flipping (AHAF) procedure, involving minimal, targeted adversarial perturbations that flip an image's hallucination attribute, to construct perfectly aligned positive-negative pairs, which are then used for contrastive fine-tuning. AHAF simultaneously serves as a \emphdiagnostic probe, revealing that MLLM visual representations lie dangerously close to hallucination decision boundaries. Requiring only 0.9% of the COCO dataset and adding zero inference overhead, ACFT achieves state-of-the-art performance on POPE, MME, and four description-level hallucination benchmarks across LLaVA, MiniGPT-4, and Qwen2.5-VL. Code is available at \urlhttps://github.com/zxp555/ACFT_MM26. \endabstract

Abstract:
Remote sensing understanding inherently requires multi-resolution observation, since different targets and application tasks demand different levels of spatial detail. While low-resolution (LR) imagery enables efficient global observation, high-resolution (HR) imagery provides critical local details at a much higher acquisition cost and with limited coverage. This motivates a cross-scale sensing strategy that selectively acquires HR imagery guided by LR-based global perception to improve task performance under constrained cost. Existing HR sampling methods typically make selection decisions from isolated LR patches, thereby ignoring fine-grained intra-patch importance and cross-patch contextual interactions, leading to fragmented feature representation and suboptimal scene reasoning under sparse HR observations. To address this issue, we formulate cross-scale remote sensing understanding as a unified cost-aware problem that couples fine-grained HR sampling with cross-patch representation prediction, enabling more effective task reasoning with fewer HR observations. Furthermore, we present GL-10M, a high- and low-resolution dataset with nearly 100,000 scene pairs and 10 million images for large-scale cross-resolution pretraining. Extensive experiments on recognition and retrieval tasks show that our method consistently achieves a superior performance--cost trade-off. The code is publicly available at https://github.com/xzhacc/CrossSO.

Abstract:
Multimodal Large Language Models (MLLMs) have emerged as a promising foundation for universal multimodal embeddings. Recent studies have shown that reasoning-driven generative multimodal embeddings can outperform discriminative embeddings on several embedding tasks. However, Chain-of-Thought (CoT) reasoning tends to generate redundant thinking steps and introduce semantic ambiguity in the summarized answers in broader retrieval scenarios. To address this limitation, we propose Rewrite-drIven Multimodal Embedding(RIME), a unified framework that jointly optimizes generation and embedding through a retrieval-friendly rewrite. Meanwhile, we present the Cross-Mode Alignment (CMA) to bridge the generative and discriminative embedding spaces, enabling flexible mutual retrieval to trade off efficiency and accuracy. Based on this, we also introduce Refine Reinforcement Learning (Refine-RL) that treats discriminative embeddings as stable semantic anchors to guide the rewrite optimization. Extensive experiments on MMEB-V2, MRMR and UVRB demonstrate that RIME substantially outperforms prior generative embedding models while significantly reducing the length of thinking. The code is publicly available at \urlhttps://github.com/PeppaWu/RIME.

Abstract:
Recent studies highlight the critical role of retrieval-augmented mechanisms in social media popularity prediction (SMPP). Although such frameworks have improved SMPP performance by leveraging historical posts, existing methods still suffer from the low retrieval accuracy due to the oversight of relative relationships among UGC instances. To address this limitation, we propose a novel Relation-Enhanced Retrieval-Augmented framework (RE-Rag) that models UGC similarity as a continuous relation jointly driven by semantic content and social attributes. Specifically, RE-Rag employs a Semantic-Attribute Retriever (SAR) to obtain instances aligned in both semantic and social-attribute distributions. Subsequently, we design a Relation-Guided Predictor (RGP): first, cross-attention encodes multimodal features of retrieved instances; then, a relative relation graph is introduced to guide attention weight allocation, forming a Relation-Guided Transformer (RGTs) that dynamically modulate attention weights based on relative attribute relations to capture the interplay between semantics and various social attributes. The refined features are fused with the target instance for popularity prediction. Experiments on three public benchmarks show that RE-Rag consistently outperforms state-of-the-art methods in both prediction accuracy and retrieval efficiency. The code and data are available at \urlhttps://github.com/BBEC-opt/RE_RAG.

Abstract:
In Image-to-Video (I2V) generation, a video is created using an input image as the first-frame condition. Existing I2V methods concatenate the full information of the conditional image with noisy latents to achieve high fidelity. However, the denoisers in these methods tend to shortcut the conditional image, which is known as conditional image leakage, leading to performance degradation issues such as slow motion and color inconsistency. In this work, we further clarify that conditional image leakage leads to overfitting to in-domain data and decreases the performance in out-of-domain scenarios. Moreover, we introduce Fourier-Guided Latent Shifting I2V, named FlashI2V, to prevent conditional image leakage. Concretely, FlashI2V consists of: (1) Latent Shifting. We modify the source and target distributions of flow matching by subtracting the conditional image information from the noisy latents, thereby incorporating the condition implicitly. (2) Fourier Guidance. We use high-frequency magnitude features obtained by the Fourier Transform to accelerate convergence and enable the adjustment of detail levels in the generated video. Experimental results show that our method effectively overcomes conditional image leakage and achieves the best generalization and performance on out-of-domain data among various I2V paradigms. Our FlashI2V-14B outperforms CogVideoX1.5-5B-I2V and Wan2.1-I2V-14B on multiple metrics of VBench-I2V, achieving state-of-the-art performance.

Abstract:
Object removal aims to eliminate target objects specified by a mask while preserving visual consistency with the surrounding regions. Existing methods typically rely on contextual information from surrounding regions. However, in dense scenes where the surrounding regions contain instances visually similar to the removal target, such reliance often leads to semantic interference, resulting in incomplete removal. This problem arises from erroneous information propagation in the attention space, where masked queries tend to align with such instances due to global similarity matching in self-attention. To address this challenge, we propose a \underlineDiffusion-based \underlineObject \underlineRemoval framework for dense \underlineScenes, dubbed DORS, built upon a Dynamic Attention Routing mechanism comprising two complementary components: Instance-Filtered Attention, which suppresses misleading semantic information from similar instances through dynamically constructed mask-guided attention constraints, and Context-Guided Routing, which dynamically routes complementary scene information to maintain visual consistency. We further introduce DOR-Bench, a benchmark tailored for object removal in dense scenes. Extensive experiments demonstrate that DORS outperforms state-of-the-art methods, particularly in reducing incomplete removal and duplicate artifacts. The project page, which provides access to the source code and DOR-Bench, is publicly available at \urlhttps://httang1224.github.io/DORS/.

Abstract:
Cross-market recommendation (CMR) aims to enhance recommendation performance across multiple markets. Due to its inherent characteristics, i.e., data isolation, non-overlapping users, and market heterogeneity, CMR introduces unique challenges and fundamentally differs from cross-domain recommendation (CDR). Existing CMR approaches largely inherit CDR by adopting the one-to-one transfer paradigm, where a model is pretrained on a source market and then fine-tuned on a target market. However, such a paradigm suffers from CH1. source degradation, where the source market sacrifices its own performance for the target markets, and CH2. negative transfer, where market heterogeneity leads to suboptimal performance in target markets. To address these challenges, we propose FeCoSR, a novel federated collaboration framework for cross-market sequential recommendation. Specifically, to tackle CH1, we introduce a many-to-many collaboration paradigm that enables all markets to jointly participate in and benefit from training. It consists of a federated pretraining stage for capturing shared behavior-level patterns, followed by local fine-tuning for market-specific item-level preferences. For CH2, we theoretically and empirically show that vanilla Cross-Entropy (CE) exacerbates market heterogeneity, undermining federated optimization. To address this, we propose a Semantic Soft Cross-Entropy (S^2CE) that leverages shared semantic information to facilitate collaborative behavioral learning across markets. Then, we design a market-specific adaptation module during fine-tuning to capture local item preferences. Extensive experiments on the real-world datasets demonstrate the advantages of FeCoSR\footnoteOur code is available at https://github.com/jundongchen13/FeCoSR. over other methods.

Authors: Enzhen Liu (School of Computer Science and Technology Shandong University), ZHikun Hu (School of Computer Science and Technology Shandong University), Xiang Lei (Zhiyang Innovation Technology Co., Ltd.), Wenhong Duan (School of Artificial Intelligence Shandong University), Yutong Qi (Faculty of Science, The University of Hong Kong), Minghuan Liu (Department of Computer and Mathematical Sciences, University of Toronto), Junyi Wang (School of Computer Science and Technology Shandong University), Tian Gan (School of Computer Science and Technology Shandong University)

Abstract:
\beginabstract 3D Gaussian Splatting (3DGS) demonstrates exceptional potential in localization due to its realistic rendering. However, existing methods typically confine it to pose refinement, making their performance highly dependent on the quality of the initial pose estimate. This dependency limits their robustness when the initialization is inaccurate. In this paper, we introduce a new Render-to-Optimize relocalization paradigm, shifting the focus from optimizing camera poses to directly optimizing the scene coordinate regression network. Built on this idea, we present DiffGS-SCR, a differentiable Gaussian-based pipeline that directly optimizes the Scene Coordinate Regression (SCR) network for improving camera relocalization. The DiffGS-SCR pipeline proceeds in two stages. In the first stage, we propose a Render-Aware Geometric Optimization (RAGO) scheme to refine the coordinate regression network by leveraging differentiable rendering for photometric supervision. To ensure a robust geometric initialization, we further introduce a Geometry-Guided Sampling mechanism that filters out low-quality Gaussian primitives based on their predicted attributes, enabling more accurate pose estimation. In the second stage, we perform pose refinement using Feature Gaussian. By synthesizing dense feature maps and establishing dense feature matches with the query image, effectively optimize the initial pose. Extensive experiments demonstrate that DiffGS-SCR achieves state-of-the-art performance, significantly outperforming existing SCR and 3DGS methods in both accuracy and robustness. The code is available at \textcolormagentahttps://github.com/86057757comeon/DiffGS-SCR. \endabstract

Abstract:
Online error detection in egocentric procedural videos is critical for intelligent assistants, requiring action recognition and anomaly localization in streaming videos without access to future frames. However, existing methods fall short due to the inherent ambiguity of non-causal temporal modeling and the inability to handle the complex and varied nature of procedural sequences. To overcome these limitations, we propose a novel Procedural Probabilistic Graph Error Detector (PPGED) for egocentric error detection. It tackles these challenges through a dynamic memory action proposal for video streaming and constructs explicit probabilistic procedural constraints to capture the complexity among procedural sequences. Specifically, our PPGED method consists of two key components: (1) The Dynamic Memory Action Proposal module that harnesses a parametric memory to dynamically compress historical context and extract discriminative action features from streaming video. (2) The Procedural Probabilistic Graph module that combines an action transition graph to probabilistically model valid action sequences and execution distributions. Extensive experiments on three public egocentric datasets demonstrate that PPGED establishes new state-of-the-art performance for egocentric error detection under both online and offline settings. Our code is available at \urlhttps://github.com/SIG-MIPAL/PPGED.

Abstract:
Codebook-based blind face restoration (BFR) often suffers from ambiguous conditioning features and a fragile prediction mechanism under severe degradation. To address these challenges, we propose GeoMAR, a framework designed to unleash geometrically aligned features with masked autoregressive (MAR) refinement for robust face restoration. For feature conditioning, we introduce a dual-input extraction pipeline to extract component-based geometric descriptions with explicit, spatially faithful anchors. These textual priors are integrated with low-quality (LQ) features via an Aligned Geometric Priors Injector, which employs a KV-Q exchange strategy to generate geometrically aligned features. For prediction mechanism, we reformulate the one-step mapping into a multi-step MAR process. This coarse-to-fine generation progressively refines complex facial regions based on increasingly reliable context. Experiments on one synthetic and three real-world benchmarks demonstrate that GeoMAR achieves highly competitive perceptual quality and coherent visual structures compared with existing methods. The code is available at https://github.com/BRL-SYSU/GeoMAR.git.

Abstract:
Multimodal Emotion-Cause Pair Extraction (MECPE) in conversations aims to identify emotions and their corresponding causes from conversational data across text, audio, and video modalities. A key yet underexplored challenge in MECPE is that emotions are often associated with multiple correlated cues, while only a subset of them are causally necessary. Existing methods heavily rely on correlation-based learning and thus struggle to effectively distinguish statistical associations from genuine emotion causal effects, leading to spurious associations. To address this, we propose debiased CAusal REasoning HyperGraph named CARE-HG, a causality-aware framework for emotion-cause analysis. Specifically, we construct an LLM-guided heterogeneous hypergraph that leverages LLM priors as structural guidance to model interactions between multimodal emotion cues and potential causes. Furthermore, we introduce a novel prior-aware counterfactual intervention mechanism that masks salient modality nodes as potential causal factors, enabling robust learning of causal invariances. Extensive experiments on benchmark MECPE datasets demonstrate that CARE-HG consistently outperforms state-of-the-art methods, validating its effectiveness for debiased multimodal emotion-cause reasoning. Our code is available at \hrefhttps://github.com/xiyejing/CARE-HGhttps://github.com/xiyejing/CARE-HG.

Abstract:
Training-free unseen 6D object pose estimation leverages vision foundation model features to establish correspondences between query and template images and recover object pose without task-specific training or fine-tuning. However, existing methods mainly construct correspondences directly from feature similarity, while overlooking the retrieved template pose itself as an intrinsic spatial constraint; neglecting this geometric reasoning cue limits correspondence reliability and thus degrades pose estimation accuracy. To address this limitation, we propose HypoPose, which explicitly treats the retrieved template pose as a geometric hypothesis to guide correspondence construction. Under each pose hypothesis, rotation-aware similarity is combined with relaxed optimal transport to establish globally consistent correspondences. Subsequent geometric verification prunes inconsistent matches, and multi-hypothesis correspondence fusion aggregates verified matches before final pose estimation. Experiments on the RGB unseen BOP benchmark demonstrate that HypoPose significantly outperforms prior training-free baselines, achieving an Average Recall of 46.2% with a 9.0% absolute improvement. Real-world experiments further illustrate its practical applicability. Code is available at \urlhttps://github.com/jialeren/hypopose.

Abstract:
The rapid evolution of deepfake technology poses an unprecedented threat to the authenticity of Graphics Interchange Format (GIF) imagery, which serves as a representative of short-loop temporal media in social networks. However, existing proactive forensics works are designed for static images, which limits their applicability to animated GIFs. To bridge this gap, we propose GIFGuard, the first spatiotemporal watermarking framework tailored for deepfake proactive forensics in GIFs. In the embedding stage, we propose the Spatiotemporal Adaptive Residual Encoder (STARE) to ensure robustness against high-level semantic tampering. It employs a 3D convolutional backbone with adaptive channel recalibration to capture globally coherent temporal dependencies. In the extraction stage, we design the Deep Integrity Restoration Decoder (DIRD). It utilizes a spatiotemporal hourglass architecture equipped with 3D attention to restore latent features, allowing for the accurate extraction of watermark signals even under severe facial manipulation. Furthermore, we construct GIFfaces, the first large-scale benchmark dataset curated for GIF proactive forensics to facilitate research in this domain. Extensive results show that GIFGuard achieves high-fidelity visual quality and remarkable robustness performance against deepfakes. The related code and dataset are available at https://github.com/vpsg-research/GIFGuard.

Abstract:
Micro-video popularity prediction (MVPP) aims to forecast the future popularity of videos on online media, which is essential for applications such as content recommendation and traffic allocation.In real-world scenarios, it is critical for MVPP approaches to understand both the temporal dynamics of a given video (temporal) and its historical relevance to other videos (spatial).However, existing approaches suffer from limitations in both dimensions: temporally, they rely on sparse short-range sampling that restricts content perception; spatially, they depend on flat retrieval memory with limited capacity and low efficiency, hindering scalable knowledge utilization.To overcome these limitations, we propose a unified framework that achieves joint spatio-temporal enlargement, enabling precise perception of extremely long video sequences (e.g., up to 100 frames, far beyond the typical 20-frame window) while supporting a scalable memory bank that can infinitely expand to incorporate all relevant historical videos.Technically, we employ a Temporal Enlargement driven by a frame scoring module that extracts highlight cues from video frames through two complementary pathways: sparse sampling and dense perception. Their outputs are adaptively fused to enable robust long-sequence content understanding. For Spatial Enlargement, we construct a Topology-Aware Memory Bank that hierarchically clusters historically relevant content based on topological relationships. Instead of directly expanding memory capacity, we update the encoder features of the corresponding clusters when incorporating new videos, enabling unbounded historical association without unbounded storage growth.Extensive experiments on three widely used MVPP benchmarks demonstrate that our method consistently outperforms 11 strong baselines across mainstream metrics, achieving robust improvements in both prediction accuracy and ranking consistency.The code repository is available at https://github.com/AI4SS/STAP.

Abstract:
Visual Autoregressive (VAR) modeling has attracted attention for its coarse-to-fine generation via next-scale prediction. However, higher scale resolutions lead to rapid token growth and substantially increased latency. Existing methods accelerate the late high-resolution stages by reusing converged tokens, but may discard high-frequency details in unupdated regions. To address this challenge, we propose PruneVAR, a training-free VAR acceleration framework that reduces redundant model computations without decreasing token inputs to preserve fine-grained texture coherence. Our approach is motivated by the finding that the redundancy in VAR transitions from stage-level stability to layer-wise functional overlap across the multi-scale generation process. Based on this insight, we introduce a corse-to-fine layer pruning strategy. Specifically, we first prune all layers at specific stages by skipping model forward while restoring pruned outputs using cached image increments from previous scale steps. As generation proceeds to higher resolutions, we pre-identify redundant layers by analyzing the overlap of core update regions and skip only these layers in the late stages. Experiments show that PruneVAR achieves up to 2× speedup while maintaining competitive visual quality. More visualization results are available at \urlhttps://jiajianxie.github.io/prunevar/.

Abstract:
While Multimodal Large Language Models (MLLMs) excel in general video understanding, their capability in fine-grained and motion-centric tasks remains limited. This limitation is particularly critical in micro-gesture recognition (MGR), where micro-gestures (MG)subtle, short-duration, and spatially localized human movementsserve as key discriminative signals for implicit affective analysis, yet are easily neglected following common prompting practices. Although MGR has been intensively studied by many discriminative approaches, the use of MLLMs on MGR is underexplored, with notably poor performance. We hypothesize that the motion-sensitive representation ability of MLLMs is constrained by its inherent single-pass forward inference, which can be substantially enhanced through carefully designed test-time guidance. Motivated by this, building on our prior findings regarding temporal insensitivity in Video LLMs, we diagnose zero-shot MGR errors in the Negative Log-Likelihood (NLL) space. We observe that MLLMs suffer from two bottlenecks: 1) insufficient localized evidence and 2) severe score biases driven by language and motion-agnostic appearances. Thus, we propose a novel test-time evidence calibration framework that improves both reasoning details and prediction reliability. Specifically, we introduce a tree search mechanism to progressively acquire localized, fine-grained visual evidence, coupled with a test-time calibration module to mitigate score biases. The multi-cue fusion module then integrates evidence from multiple cues without relying on a single cue for final prediction. Our framework achieves a mean-class accuracy of 26.84% on iMiGUE and 22.10% on MA-52, significantly outperforming the Qwen2.5-VL baseline that produces 16.15% and 10.20%, respectively. The code will be available at https://zero-melo.github.io/Zero-MELO.

Abstract:
Vision-language models (VLMs) are increasingly deployed in autonomous driving (AD) systems, creating an urgent need for rigorous safety evaluation under rare yet safety-critical scenarios. Among these, interactions with vulnerable road users (VRUs), such as pedestrians, represent a major source of real-world failures. However, existing safety-critical scenario generation methods predominantly rely on simulator-based pipelines, which suffer from a substantial sim-to-real gap and often fail to capture realistic, diverse, and unforeseen humanvehicle interaction dynamics. To address this, we present \toolns, a goal-conditioned diffusion framework for safety-critical scenario generation in vision-language model-based autonomous driving systems (VLMADs). Our key insight is to formulate scenario generation as a goal-conditioned diffusion process, where a predefined catastrophic end-state (\eg, a collision with a pedestrian) serves as a strong supervisory signal, guiding the generation of temporally coherent video trajectories that naturally evolve toward safety-critical outcomes. Building on this formulation, we introduce \monens, which leverages VLMs to analyze benign driving contexts and infer latent vulnerabilities in humanvehicle interactions, producing structured end-state specifications that induce high-risk scenarios. Conditioned on these targets, we further propose \mtwons, which grounds semantic threats into physically plausible visual dynamics. Specifically, we instantiate high-risk agents (\eg, VRUs) within the scene via depth-aware geometric projection, followed by boundary-conditioned diffusion to generate intermediate frames with consistent motion patterns and temporal coherence. Extensive experiments across 3 VLMADs demonstrate that \tool increases the Judge Overall Score, a metric using a VLM judge to evaluate VLMADs' understanding and decision-making, by 24.25% on average compared to SoTA baselines. Furthermore, fine-tuning a VLMAD with \toolns-generated data improves performance in real-world driving scenes by an average of 15.9%. Our code is available at \hrefhttps://github.com/JoFrc/SafeGen\textcolorbluewebsite.

Abstract:
Long-context inference enhances the reasoning capability of Large Language Models (LLMs), but incurs significant computational overhead. Token-oriented methods, such as pruning and skipping, have shown great promise in reducing inference latency, yet still suffer from inherently insufficient structure optimization, outdated selection criteria, and redundancy interference, resulting in suboptimal speed-accuracy trade-off. To address these issues, we propose a novel training-free framework dubbed Self-Predictive Token Skipping (SPTS), for efficient long-context LLM inference. Specifically, motivated by probing the influence of target layers prior to skipping, we design two selective token skipping strategies for typical structures, including Partial Attention Probing (PAP) for multi-head attention and Low-rank Transformation Probing (LTP) for feed forward network. The former selects informative tokens via partial forward attention computation, while the latter constructs a low-rank proxy network to predict token transformations. In addition, a Multi-Stage Delayed Pruning (MSDP) strategy reallocates skipping budgets and progressively removes redundant tokens across layers. Extensive experiments display the effectiveness of our method, achieving up to 2.46× and 2.29× speedups for prefilling and end-to-end generation, respectively, while maintaining state-of-the-art accuracy. The source code is available at \urlhttps://github.com/Wuzimeng/SPTS.

Abstract:
Multimodal Empathetic Response Generation (MERG) integrates textual, visual, and acoustic signals to synthesize nuanced affective responses. While Large Language Models (LLMs) demonstrate strong capabilities in emotional reasoning, they still suffer from reasoning redundancy and contextual prolixity. Applying static, heavy-duty reasoning to all queries wastes computational resources, while uncompressed multimodal histories introduce excessive and noisy tokens into LLM inputs. To address these challenges, we propose the CoRE framework. Inspired by dual-process cognition, CoRE introduces a confidence-driven fast-slow reasoning mechanism that dynamically toggles between a lightweight fast path for intuitive fluency and an expanded slow reasoning chain for resolving complex emotional ambiguity. Additionally, a condensed-temporal dialogue encoder module distills noisy historical signals into compact representations, reducing the multimodal token burden. Extensive experiments on the large-scale AvaMERG benchmark demonstrate that CoRE achieves competitive performance, outperforming existing LLM baselines in emotion recognition, semantic alignment and human-evaluated empathy while optimizing inference efficiency. The source code is available at \urlhttps://github.com/Cathryn1103/CoRE_MERG.

Abstract:
Few-shot class-incremental learning (FSCIL) requires models to learn novel classes from limited examples while retaining previously acquired knowledge, leading to a fundamental stability--plasticity trade-off. Existing methods commonly freeze pretrained backbones, which restricts adaptation, or introduce prompts and adapters that may increase model size and inference cost across sessions. We study FSCIL from the perspective of directly reusing and adapting pretrained weights. To this end, we propose CKPD-FSCIL, a unified framework for knowledge-preserving adaptation at both the weight and layer levels. At the weight level, Continuous Knowledge-Preserving Decomposition uses past-class feature covariance to partition selected weights into preserved and learnable components, and recalibrates this partition across incremental sessions. At the layer level, Continuous Adaptive Layer Selection uses an Adapter Sensitivity Ratio to dynamically identify comparatively suitable layers for adaptation. After each session, the learned low-rank components are merged into the original weights, introducing no additional inference-time parameters or FLOPs. Experiments across multiple FSCIL benchmarks, backbone architectures, and evaluation settings demonstrate state-of-the-art or competitive performance, together with favorable knowledge retention, scalability, and efficiency. The code is available at https://github.com/xiaojieli0903/CKPD-FSCIL.

Abstract:
Hateful video detection has become increasingly important with the rapid growth of video-centric social media platforms, given the serious risks that hate speech poses to both individual well-being and social cohesion. Compared with text or static multimodal content, hateful video detection remains underexplored and significantly more challenging, as hateful meaning often arises from complex interactions among multimodal cues, including speech, audio, and visual content. Moreover, such signals are often brief, implicit, and temporally dependent, making them difficult to capture using conventional video-level representations. In this work, we propose CLARA\footnoteThe code for CLARA is available at \urlhttps://github.com/yuchenzhang-1/CLARA., a clip-level multimodal framework for hateful video detection. Instead of treating a video as a single instance, CLARA models it as a sequence of fine-grained clips, enabling more precise capture of temporally localized hateful signals. We introduce a Mixture-of-Experts clip encoder for adaptive multimodal alignment, a local-global segment contrastive objective to jointly model short-term cues and long-range temporal dependencies, and VLM-derived rationales integrated via a gated Transformer to provide high-level semantic guidance. Extensive experiments on three hateful video datasets demonstrate that CLARA consistently outperforms state-of-the-art methods. Further ablation studies and parameter analyses validate the effectiveness of each component.

Abstract:
Interpreting the emotional responses triggered by images is central to achieving emotional intelligence. Compared with natural images, visual art is intentionally created to elicit emotional responses from its viewers through abstract concepts and visual metaphors, making affective interpretation particularly challenging. However, most existing methods rely on general-purpose visual embeddings (e.g., CLIP), failing to capture the nuanced cues underlying artistic emotion. To address this gap, we propose ProFocus, a novel framework that models affective experience in artistic images via progressive visual focusing. The key idea is to model visual representation learning inspired by a hierarchical cognitive theory of human aesthetic appreciation. Technically, ProFocus contains two core components: a Hierarchical Art Critic (HAC) and a Progressive Hint Fusion (PHF) module. HAC leverages multimodal large language models to generate structured linguistic priors at three cognitive levels--atmospheric style, narrative subjects, and concrete details--thereby translating artistic perception into coherent semantic guidance. Building upon these priors, PHF departs from conventional cross-modal fusion by sequentially injecting the hierarchical hints into visual features, enabling a progressive focusing process that mirrors human perception. This design allows the model to capture subtle affective cues and produce more faithful explanations. Extensive experiments on the ArtEmis v1.0 and v2.0 datasets demonstrate that ProFocus consistently outperforms state-of-the-art methods in both emotion recognition and affective explanation. Project page: https://github.com/Zhang-Zhiyan/ProFocus.

Abstract:
Unsupervised video object-centric learning aims to decompose dynamic scenes into persistent, object-level representations without supervision. However, existing slot-based methods struggle to maintain stable object identity in challenging settings such as rapid motion and partial occlusion. First, they typically encode both the per-frame appearance of an object and its identity across frames in a single slot vector, creating an objective conflict that leads to \emphslot swapping: reconstruction requires sensitivity to transient visual changes, whereas temporal consistency requires invariance to them. Second, the token renormalization used in Slot Attention can amplify weakly attending slots, allowing them to absorb tokens from other objects and destabilize slot-to-object correspondence. We propose Dual-State Slot Attention (DSSA), a fully self-supervised framework that addresses these limitations by separating appearance from identity and by reducing spurious updates from weakly matching slots. DSSA decomposes each slot into a \emphlocal state for per-frame appearance and an \emphidentity state for temporally stable object information, thereby aligning reconstruction and temporal consistency with separate representations. The identity state is updated through a learned recurrent transition that acts as a temporal filter on the local state, while \emphcompetition-modulated aggregation (CMA) down-weights updates from weakly matching slots and prevents them from absorbing tokens from other objects. Experiments on MOVi-C, MOVi-D, and YouTube-VIS demonstrate that DSSA consistently improves segmentation quality and temporal consistency over prior methods, while also yielding stronger downstream object recognition and video dynamics prediction. Code is available at \hrefhttps://github.com/UARK-AICV/DSSA\textttgithub.com/UARK-AICV/DSSA.

Abstract:
Training-free open-vocabulary semantic segmentation (OVSS) has progressed rapidly with foundation models, yet most pipelines still treat inference as a one-shot dense semantic readout, remaining static even when auxiliary refinement or external model composition is applied. We observe that SAM3 already natively exposes decoupled predictions with complementary semantic, instance-level, and presence-aware signals that can be reinterpreted as internal evidence for test-time self-correction. Based on this observation, we propose SCOPE, a fully training-free framework that reformulates OVSS as a self-correcting inference process through three tightly coupled stages: a Self-Diagnostic stage that identifies unreliable queries and triggers conditional re-inference, a Self-Correcting stage that reconciles semantic and instance predictions through asymmetric cross-head refinement, and a Self-Stabilizing stage that suppresses view-sensitive variance via aligned prediction fusion. SCOPE operates entirely within the native SAM3 inference stack without parameter updates or auxiliary backbones, and establishes a new state-of-the-art among training-free methods on eight standard benchmarks. Code and model are available at https://github.com/akuan1234/SCOPE.

Abstract:
Brain-image alignment aims to establish correspondences between neural signals and visual representations, where Optimal Transport (OT) has emerged as one of the most effective approaches by modeling fine-grained matching between brain voxels and image tokens. However, existing OT methods treat all image tokens equally, ignoring the varying contributions of different visual regions to neural responses, which allows background noise to interfere with semantic alignment in key regions. In this paper, we propose SG-OT, a semantic-aware optimal transport framework that transforms OT from global uniform matching to selective matching over key regions. Specifically, we first estimate the semantic importance of image tokens to identify informative visual regions; then, we use their indices to constrain the participation of corresponding brain tokens and perform optimal transport only within the selected subset. This design preserves full visual context while effectively suppressing background interference. Extensive experiments on the NSD dataset show that SG-OT consistently outperforms the prior SOTAs across ten evaluation metrics, achieving an average improvement of 1.8% and improving caption quality (CIDEr from 72.90% to 77.92%). Our code is available at \urlhttps://github.com/CzhiL/SG-OT.

Abstract:
\beginabstract The advancement of mobile GUI agents has opened new opportunities for automating tasks on mobile devices. Training these agents requires large-scale high-quality data, which is prohibitively expensive when relying on human labor. Given the vast population of global mobile phone users, if automated data collection from them becomes feasible, the resulting data volume and the subsequently trained mobile agents could reach unprecedented levels. % Given the vast number of mobile phone users worldwide, automatic data collection from users can result in unprecedented data volume and the subsequently trained mobile agents. Nevertheless, two major challenges arise: % (1) extracting user instructions to build complete datasets without human intervention and (2) utilizing distributed data from diverse users while preserving privacy. (1) extracting user instructions without human intervention and (2) utilizing distributed user data while preserving privacy. To tackle these challenges, we propose MobileA\textsuperscript3gent, a collaborative framework that trains mobile GUI \underlineAgents using self-sourced data from diverse users. The framework comprises two components, each targeting a specific challenge: (1) \underlineAuto-Annotation, which enables the automatic collection of high-quality datasets during users' routine phone usage with minimal cost. (2) FedVLM-\underlineA, which enhances federated VLM training under non-IID distributions by incorporating adapted global aggregation based on both episode-level and step-level variability. Extensive experiments prove that MobileA\textsuperscript3gent achieves superior performance over traditional approaches at only 1% of the cost, highlighting its potential for real-world applications. % \footnoteOur code is publicly available at: \hrefhttps://anonymous.4open.science/r/MobileA3gent-Anonymoushttps://anonymous.4open.science/r/MobileA3gent-Anonymous. Our code is publicly available at: \hrefhttps://github.com/wwh0411/MobileA3genthttps://github.com/wwh0411/MobileA3gent. \endabstract

Abstract:
As generative image models achieve widespread adoption, watermarking has emerged as a critical defense for copyright protection and provenance. However, the lack of a standardized evaluation framework has led to fragmented and incomparable performance assessments of existing algorithms. To address this, we propose GenWM-Bench, a standardized evaluation framework tailored for diffusion-based watermarks. We systematically categorize 12 state-of-the-art watermarking schemes into distinct technical paradigms. To simulate complex real-world threats, our framework incorporates four critical attack surfaces: Image-level, Model-level, Detector-oriented, and Forgery. Our evaluation uncovers several critical yet unexplored vulnerabilities within current watermarking. For instance, weight-integrated schemes are susceptible to parameter averaging during model merging; content-agnostic schemes face the risk of template isolation via pattern extraction; and semantic-based methods exhibit significant failure rates during regeneration variants. Through a Pareto Front analysis across different schemes, we identify the operational limits of various watermarking methods under extreme threats and highlight the most challenging attacks. GenWM-Bench establishes a reference standard for the research and deployment of robust watermarks. Code is available at https://github.com/REPO-EXP/GenWM-Bench.git.

Abstract:
Video Anomaly Detection (VAD) aims to identify abnormal events in surveillance. To circumvent prohibitive frame-level annotation, Weakly-Supervised VAD (WSVAD) relies on coarse-grained video-level labels, yet is fundamentally bottlenecked by the inherent intra-video ambiguity. Existing Vision-Language methods attempt to introduce external semantic guidance, but often suffer from temporal granularity mismatch. Moreover, the pervasive yet overlooked issue of gradient saturation severely hampers the model's learning of challenging anomaly patterns. To address these challenges, we propose UPR-VAD, a unified framework based on Uncertainty-aware Purification and Regularization. Unlike static pipelines, UPR-VAD establishes a synergistic co-evolutionary loop: (1) The Uncertainty-aware Temporal Information Bottleneck (U-TIB) leverages epistemic uncertainty estimation to dynamically suppress high-response yet high-uncertainty ambiguous segments; (2) Intra-Video Negative Contrastive Learning (IVNC) shares uncertainty masks to treat background contexts as hard negatives for robust feature representation, creating a mutually reinforcing feedback loop with U-TIB; (3) Sparsity-Guided Confidence Regularization (SGCR) incorporates Sparsity Prior Alignment (SPA) to constrain predictions within the linear sensitive region of the activation function, forcing the model to mine and learn subtle yet critical anomaly details. Extensive experiments conducted on widely used benchmarks conclusively validate the effectiveness and superiority of our proposed UPR-VAD. The source code is publicly available at \hrefhttps://github.com/WeiliangHuang-UM-connect/UPR-VAD \textcolorblack\nolinkurlhttps://github.com/WeiliangHuang-UM-connect/UPR-VAD.

Abstract:
Referring Video Object Segmentation (RVOS) aims to segment referred objects at the pixel level in video sequences based on natural language descriptions. Existing methods typically introduce motion information within a unified cross-modal temporal modeling framework, where language cues are used for target localization and segmentation. However, the dependency of expressions on motion semantics is not explicitly modeled, making it difficult to adaptively adjust the use of motion information according to different semantic requirements. To address these issues, we propose an Expression-driven Motion Calibration (EMC) framework for RVOS that explicitly unlocks and leverages the motion semantics within expressions. The proposed method extracts interpretable motion control signals from expressions via a Motion Signal Processing (MSP) module, and employs a Motion Influence Calibration (MIC) module to adjust the contribution of motion cues during temporal decision making. In addition, a Semantic Temporal Stage Construction (STSC) module is introduced to build expression-relevant temporal stages, providing a compact temporal candidate space for motion calibration. Through extensive evaluation on six standard benchmarks, including Ref-YouTubeVOS, Ref-DAVIS17, MeViS (valid/validu), A2D-Sentences, and JHMDB-Sentences, the superiority of our method is validated. We will release the code on https://github.com/Jeven7/EMC.

Abstract:
Proteins are the primary executors of biological functions, and Functional Protein Design (FPD) holds immense promise. Despite the rapid advances of Large Language Models (LLMs), existing protein design methods have yet to fully harness their potential for targeted FPD, and still rely on simplified textual inputs such as template-style schemas or keyword annotations. As a result, they remain inadequate for FPD from free-form complex intents expressed in natural language. To this end, we introduce NL2Protein, a new task that aims to generate targeted functional proteins from free-form natural instructions. Since directly mapping such complex intents to full protein sequences presents an intractable search space, we employ functional motifs as an intermediate representation. Building on this formulation, we propose Motif-FPD, the first framework for NL2Protein, featuring a dual-core design centered on a specialized LLM, MFPD-R1. Specifically, MFPD-R1 translates free-form instructions into structured functional motifs through biochemical chain-of-thought reasoning, followed by the generation of complete protein sequence and 3D structure by biological experts. As high-quality functional motif generation constitutes the core bottleneck, we address it through a dual strategy of data curation and hierarchical reasoning alignment. First, we develop the MotifForge engine to construct MFPD-210K, a large-scale dataset for motif-grounded FPD. We then optimize MFPD-R1 through a two-stage pipeline: curriculum-based supervised alignment to adapt the base LLM into a specialized reasoner, followed by targeted preference optimization with a confusion-oriented data ranking mechanism and a unified reward group, to stabilize its motif-centric reasoning capability while further refining biological fidelity. Besides, we establish MFPD-Bench coupled with a hierarchical self-consistent assessment protocol, enabling evaluations at both the framework and LLM-core levels. Extensive quantitative and qualitative experiments indicate that our Motif-FPD outperforms expert models and leading LLMs, exhibiting SOTA performance on NL2Protein. Our data and related code are publicly available at https://github.com/shiyunyun1221/MFPD-R1.

Abstract:
Large language models (LLMs) have unified diverse linguistic tasks within a single framework, yet such unification remains unexplored in human motion generation. Existing methods are confined to isolated tasks, limiting flexibility for free-form and omni-objective generation. To address this, we propose OmniMoGen, a unified framework that enables versatile motion generation through interleaved text-motion instructions. Built upon an RVQ-VAE and a modality-specific MoT, OmniMoGen preserves strong language reasoning capabilities and decreases quantization error, supporting instruction-driven motion generation. We construct X2Mo, a large-scale dataset of over 137K interleaved text-motion instructions, and introduce AnyContext, a benchmark for evaluating true motion generation capabilities. Experiments show that OmniMoGen achieves state-of-the-art performance on text-to-motion, motion editing, and AnyContext, exhibiting emerging capabilities such as compositional editing, self-reflective generation, and knowledge-informed generation. These results mark a step toward the next intelligent motion generation. Project Page: \urlhttps://OmniMoGen.github.io.

Abstract:
Despite advances in Video Large Language Models (VLLMs) that have displayed promising outcomes in video understanding, the redundancy in the long-duration frames remains a hindrance to efficient reasoning. This paper introduces a training-free Persistence-Aware Compression and Aggregation (PCA) method designed to preserve high-fidelity raw visual information before the encoding stage. PCA can be built on arbitrary VLLMs and consists of two modules: 1) A Dynamic Downsampling (DD) module that adaptively removes redundant frames by analyzing frame-wise similarity. 2) A Persistence-Aware Motion Enhancement (PAME) module that enriches each selected keyframe by aggregating the temporal context of its neighbors, ensuring that essential information is preserved even after aggressive frame reduction. Our approach substantially reduces the computation of long-context modeling, while enhancing the performance of the baseline model. Extensive experiments demonstrate that PCA consistently outperforms existing state-of-the-art approaches in both efficiency and accuracy, achieving a speedup of 1.8× to 2.5× compared to the baseline VLLM. The code is open-sourced at \hrefhttps://github.com/Heisenberg10110/PCA\textttHeisenberg10110/PCA.

Abstract:
Graphical User Interface (GUI) agents powered by vision-language models hold promise for automating real-world mobile tasks. However, progress is limited by the lack of high-coverage, long-horizon interaction trajectories collected from element-rich and rapidly evolving apps. Existing pipelines often rely on costly human demonstrations or on-policy framework, which tends to over-sample common flows while missing rare transitions and complex multi-step procedures. To address this problem, we propose SEE, a two-stage data synthesis framework consisting of (i) an efficient exploration stage that builds an explicit UI transition graph over screens and elements, and (ii) a graph-based synthesis stage that composes diverse multi-step trajectories via planning and controlled sampling. This design yields reproducible and explainable data generation, while explicitly preventing spurious cycles and enabling long-horizon composition. Across multiple real-world apps, SEE produces trajectories with an average length of 14.8 steps while avoiding spurious loops, and agents fine-tuned on SEE achieve improved task success and generalization to unseen screens. Project page \hrefhttps://github.com/Zhh-Fan/SEE-Structure-aware-Exploring-Exploiting-for-Long-horizon-GUI-Agent-Trajectory-Synthesisis here.

Abstract:
Classical paintings preserve rich spatial, cultural, and historical content, making their reconstruction as explorable 3D scenes valuable for digital preservation, immersive exhibition, and cultural engagement. Yet, unlike photographs, they often depict scenes in a single-view, stylized manner, with weak perspective, lighting, and depth cues. Existing 3D reconstruction methods are largely built on natural-image priors, making it difficult to recover geometrically plausible and visually faithful 3D representations from such inputs. To address this challenge, we introduce Classical Painting-to-3D (CP3D), a new task that aims to recover a 3D representation from a single classical painting while jointly ensuring geometric plausibility, appearance fidelity to the source artwork, and plausible novel-view synthesis. We further propose GeAR, a training-free two-stage framework for Geometry Grounding and Appearance Restitution. GeAR first converts the input painting into a geometry-grounded representation with more coherent shading and illumination cues, improving the stability of 3D Gaussian reconstruction. It then restores artwork-faithful appearance across views under spatial constraints and multi-view consistency, recovering the painterly textures and details weakened during grounding. In addition, we construct HeriArch, a curated benchmark of 10,160 high-resolution classical artworks for systematic evaluation of CP3D. Extensive experiments and user studies show that GeAR consistently outperforms strong baselines in geometric plausibility, appearance fidelity, and human preference. Code is publicly available at \urlhttps://github.com/liuxindazz/GEAR.

Abstract:
Visual speech dubbing (VSD) aims to generate high-fidelity speech synchronized with silent video. While current flow-matching models achieve implicit alignment, they often suffer from over-smoothed acoustics and textual alignment inconsistency. Moreover, directly applying reinforcement learning (RL) post-training to VSD leads to reward collapse due to the inherent conflict between heterogeneous objectives: strict temporal synchronization and high-variance expressive acoustics. In this paper, we propose RL-Dubber, an online multi-reward decoupling RL framework tailored for multimodal dubbing. To ensure stable optimization across divergent reward scales, we introduce a multi-reward decoupling mechanism that balances exploration across dubbing objectives. The designed conditional KL penalty mechanism is further incorporated to dynamically anchor policy exploration, maintaining robust audio-visual alignment. Additionally, we employ an inference-aligned Diffusion Transformer (DiT) with locality-aware block-diagonal cross-attention to improve cross-modal alignment and pronunciation accuracy. Extensive experiments on public dubbing benchmarks demonstrate that RL-Dubber achieves state-of-the-art performance in intelligibility, synchronization accuracy, speaker similarity, and emotional expressiveness. Code and pretrained models are available at \urlhttps://github.com/wendavid687/RL-Dubber.

Abstract:
Controllable generation of high-quality human images is a fundamental yet challenging problem. Existing generative models often struggle to achieve precise spatial alignment and attribute consistency simultaneously, especially under complex descriptions. Moreover, existing human image datasets lack fine-grained attribute annotations and rich multi-modal supervision. This further hinders progress toward spatially and semantically consistent generation. To address these issues, we propose AttriHuman, a diffusion-based framework that jointly models spatial control and attribute-level supervision for human image generation. We propose a Time-Feature Aggregation Module (TFAM) that integrates timestep embeddings with UNet features. This enables fine spatial control and improves structural alignment with human parsing maps. We further propose an Attribute-Aware Reward Module (AARM) that provides fine-grained attribute supervision via cross-modal alignment. This improves semantic consistency between generated images and text. To train AttriHuman, we introduce T2Human-MultiModal, a high-quality dataset of full-body human images with comprehensive multi-modal annotations. Extensive experiments demonstrate that AttriHuman consistently outperforms state-of-the-art methods. It achieves over 10% reduction in FID and more than 15% improvement in attribute accuracy. It also achieves the highest preference in human evaluations. Code is available at \urlhttps://github.com/wenzhuofanfan/AttriHuman.

Abstract:
Maintaining global geometric consistency is a central challenge in long-sequence 3D reconstruction, with scale drift being the most critical failure mode. In chunk-based inference pipelines, the scale degree of freedom in sequential Sim(3) alignment is left unconstrained, causing estimation errors to compound multiplicatively and distort global trajectories and point cloud geometry. We present a scale-consistency enhancement framework built on a key insight: in structured environments such as driving scenes, geometric quantities arising from environmental regularity remain inherently invariant across temporal segments, and discrepancies in their per-chunk measurements directly expose inter-chunk scale drift. We propose Scene Geometric Invariant Anchoring (SGIA), which extracts dominant geometric invariants from each chunk's predicted point cloud via coarse-to-fine robust estimation and exploits their cross-chunk consistency to establish scale constraints independent of point cloud registration, explicitly degenerating 7-DoF Sim(3) alignment into 6-DoF rigid-body transformation and severing chain-wise scale error propagation at its source. We further introduce a lightweight test-time adaptation strategy that fine-tunes only normalization-layer parameters via multi-objective self-supervision, progressively improving intra-chunk predictions along the sequence. Both modules are plug-and-play and require no offline retraining. Experiments on multiple long-sequence benchmarks demonstrate state-of-the-art performance, reducing absolute trajectory error by up to 32% with significant gains in trajectory stability and reconstruction quality. Code:~\small\urlhttps://github.com/WZ-CS/VGGT-Align.

Abstract:
Electroencephalography (EEG) signals provide high temporal resolution, are non-invasive, and directly reflect neural activity, making EEG-based emotion recognition a widely adopted paradigm in multimedia induced affective computing. However, practical deployment remains challenging due to the high sensitivity of EEG signals to various physiological artifacts. Most existing approaches adopt a decoupled pipeline that performs artifact removal and emotion recognition in a cascaded manner, which may cause error propagation and fail to exploit the potential synergies between the two tasks. Moreover, many EEG-based emotion recognition frameworks implicitly assume clean or well-denoised signals and lack explicit designs for handling artifacts. To address these limitations, an Artifact-Robust Emotion Mamba network (AREMamba) is proposed to unify EEG artifact removal and emotion recognition within an end-to-end learning framework. The proposed framework is built upon three key components: (1) A confidence-aware collaborative optimization mechanism is introduced to dynamically rebalance gradients, enabling stable and synergistic optimization between the artifact removal and classification tasks. (2) A dual-domain Mamba-based encoder is applied to jointly model temporal and spectral representations, enabling effective artifact suppression while preserving emotion-related neural structures. (3) A hybrid temporal modeling strategy is developed to integrate state-space sequence modeling with multi-statistics aggregation, capturing both long-range dependencies and discriminative temporal cues. Experimental results on four public EEG emotion datasets (DEAP, DREAMER, AMIGOS, and SEED) show that AREMamba consistently outperforms eleven state-of-the-art methods in both artifact removal and emotion recognition. The code is available at https://github.com/PaperCode-ARE/AREMamba/tree/master.

Abstract:
Open-vocabulary semantic segmentation in the remote sensing (RS) field requires both language-aligned recognition and fine-grained spatial delineation. Although CLIP offers robust semantic generalization, its global-aligned visual representations inherently struggle to capture structural details. Recent methods attempt to compensate for this by introducing RS-pretrained DINO features. However, these methods treat CLIP representations as a monolithic semantic space and cannot localize where structural enhancement is required, failing to effectively delineate boundaries while potentially interfering with CLIPs semantic integrity. To address this limitation, we propose DR-Seg, a novel decouple-and-rectify framework. Our method is motivated by the key observation that CLIP feature channels exhibit distinct functional heterogeneity rather than forming a uniform semantic space. Building on this insight, DR-Seg decouples CLIP features into semantic-favorable and structure-favorable subspaces, enabling targeted structural enhancement by DINO without distorting language-aligned semantics. Subsequently, a prior-driven graph rectification module injects high-fidelity structural priors under DINO guidance to form a refined branch, while an uncertainty-guided adaptive fusion module dynamically integrates this refined branch with the original CLIP branch for final prediction. Comprehensive experiments across eight benchmarks demonstrate that DR-Seg establishes a new state-of-the-art. Code is available at \urlhttps://github.com/fzl1216-research/DR-Seg.

Abstract:
3D Gaussian Splatting (3DGS) delivers high-fidelity real-time rendering but suffers from geometric and photometric degradations under sparse-view constraints. Current generative restoration approaches are often limited by insufficient temporal coherence, a lack of explicit spatial constraints, and a lack of large-scale training data, resulting in multi-view inconsistencies, erroneous geometric hallucinations, and limited generalization to diverse real-world artifact distributions. In this paper, we present ArtifactWorld, a framework that resolves 3DGS artifact repair through systematic data expansion and a homogeneous dual-model paradigm. To address the data bottleneck, we establish a fine-grained phenomenological taxonomy of 3DGS artifacts and construct a comprehensive training set of 107.5K diverse paired video clips to enhance model robustness. Architecturally, we unify the restoration process within a video diffusion backbone, utilizing an isomorphic predictor to localize structural defects via an artifact heatmap. This heatmap then guides the restoration through an Artifact-Aware Triplet Fusion mechanism, enabling precise, intensity-guided spatio-temporal repair within native self-attention. Extensive experiments demonstrate that ArtifactWorld achieves state-of-the-art performance in sparse novel view synthesis and robust 3D reconstruction. Code available at: \hrefhttps://github.com/fyting/ArtifactWorldthis https URL.

Abstract:
Object detection in adverse weather remains challenging because severe degradations weaken visual quality and disrupt semantic feature representations across diverse scenes. Existing methods usually rely on condition-specific designs, which limits their ability to generalize within a unified detector. In this paper, we propose a Coarse-to-Fine Scene Expert Network (C2FXNet) that achieves unified detection through hierarchical scene guidance. Specifically, C2FXNet introduces a dual-level guidance mechanism consisting of a Multi-step Reasoning Router (MRR), which performs GRU-based recurrent scene reasoning over compressed multi-scale visual cues and frozen coarse scene prototypes, and a Fine Scene Refinement (FSR) module, which uses image-specific semantic cues to modulate high-level features for local variation handling. Furthermore, a Scene-aware Mixture-of-Experts (SMoE) dynamically combines scene-specific experts under the joint guidance of MRR and FSR. By coupling coarse scene reasoning with fine-grained semantic refinement, C2FXNet enables robust multi-scene detection without scene-specific training. Extensive experiments on RTTS, ExDark, and our newly constructed Adverse Weather Dataset (AWD) demonstrate that C2FXNet consistently outperforms state-of-the-art methods across foggy, dark, and clear conditions, reaching 63.70%, 71.14%, and 54.19% mAP on RTTS, ExDark, and AWD, respectively. The source code will be released at \urlhttps://github.com/PolarisFTL/C2FXNet.

Abstract:
In low-light image enhancement, methods based on the horizontal/vertical intensity (HVI) color space have been widely adopted due to their ability to preserve perceptual color consistency. However, existing approaches typically rely on fixed intensity modeling, which limits their adaptability to complex and diverse low-light conditions and often results in degraded structural details, particularly in severely underexposed regions. To address this limitation, we propose SIR-HVINet, a simulated infrared guided enhancement network built upon the HVI color space. By leveraging simulated infrared imaging as an auxiliary luminance prior, our method provides complementary brightness guidance under challenging illumination conditions. Specifically, we introduce a wavelet-based infrared fusion (WIF) module to inject simulated infrared-derived luminance cues into the Intensity component, thereby enhancing structural details in dark regions while maintaining perceptual color consistency. Furthermore, we design a Dual-Path Cross-Contextual Attention (DPCCA) mechanism to facilitate coordinated feature learning among illumination, color, and structure through multi-scale interactions. Extensive experiments demonstrate that SIR-HVINet consistently outperforms state-of-the-art methods across multiple benchmark datasets. The code is available at: \urlhttps://github.com/hukkee/SIR-HVINet.

Abstract:
Accurate food nutrition estimation (FNE) is critical for dietary assessment and intelligent health management. However, most FNE methods adopt a single modality or use a simple multi-modality fusion strategy for food representations, thereby ignoring prior and contextual information across different modalities. Textual information can provide explicit and complementary semantic knowledge but may also introduce ambiguity and noise. How to avoid their incorrect guidance as well as effectively align them with visual features remains difficult. In this paper, we propose a visual-text interacted multi-modal framework, i.e., MGIR-Net, for FNE that addresses these issues. First, we employ depth maps as the geometric priors to guide a more robust self-attention mechanism for visual feature extraction. Second, a Multi-Granularity Ingredient Reasoning scheme is proposed to exploit textual ingredient information at both coarse-grained and fine-grained levels. At the coarse level, a reliability-aware ingredient aggregation mechanism is introduced to assess each ingredient from the perspectives of textual coherence and visual grounding, enabling the model to emphasize reliable ingredients while suppressing noisy candidates. At the fine level, we categorize ingredients into visually grounded and contextually inferred groups under the guidance of textual context, facilitating better alignment between individual ingredients and visual representations. Finally, we formulate nutrition prediction as a probabilistic task with hierarchical uncertainty modeling, capturing both instance-level ambiguity and task-level optimization imbalance. Extensive experiments on the Nutrition5K benchmark demonstrate that our method achieves superior performance over other state-of-the-art methods across all five nutritional targets. Code is available at https://github.com/Hanym666/MGIR-Net.

Abstract:
Micro-expression recognition (MER) remains challenging due to the extremely subtle and spatially localized nature of facial micro-movements. Existing methods either rely on external optical flow that cannot be end-to-end corrected or model only single-directional dynamics while treating every spatial location as equally reliable. This paper presents a reliability-aware self-supervised dual-view displacement learning framework that explicitly answers "where a change occurs and how well it can be trusted." To this end, we propose Dual-View Apex Warping (DVAW), which simultaneously learns dense onset-to-apex and offset-to-apex displacements by reconstructing the apex frame under a shared coordinate system. From the reconstruction process, we derive two pixel-level reliability evidences: Dynamic Support and Round-Trip Consistency that naturally partition the spatial domain into trusted, uncertain, and silent regions. Building on these evidences, we introduce Reliability-guided Consistency (ReCo), a spatially adaptive training paradigm that applies symmetric transport alignment to trusted regions, channel-swap invariance to uncertain regions, and exempts silent regions from local consistency constraints, while a global temporal ranking regularizer preserves overall structural stability. The learned displacements are further converted into scale-normalized differential deformation descriptors (divergence and curl) to emphasize local shape change. The entire pipeline operates end-to-end using only the three canonical frames (onset, apex, offset) at inference time, eliminating any external optical-flow preprocessing. Extensive experiments on SMIC-HS, CASME II, SAMM, and their composite 3DB benchmark under the standard leave-one-subject-out protocol demonstrate that the proposed method consistently achieves new state-of-the-art UF1 and UAR scores across all datasets. Source codes: https://github.com/cszyzhang/DVAW

Abstract:
AIGC has shown remarkable success in CV and NLP, and has recently demonstrated promising potential in the wireless domain. However, significant data imbalance exists across RF modalities, with abundant WiFi data but scarce mmWave and RFID data due to high acquisition cost. This makes it difficult to train high-quality generative models for these data-scarce modalities. In this work, we propose RF-CMG, a diffusion-based cross-modal generative method that leverages data-rich WiFi signals to synthesize high-fidelity RF data for scarce modalities including mmWave and RFID. The key insight of RF-CMG is to decouple cross-modal generation into high-frequency guidance and low-frequency constraint, which respectively learn high-frequency distribution from limited target modality data and preserve the underlying physical structure via low-frequency constraints during generation. On this basis, we introduce a Modality-Guided Embedding (MGE) module to steer the reverse diffusion trajectory toward the target high-frequency distribution, and a Low-Frequency Modality Consistency (LFMC) module to progressively enforce low-frequency constraints to suppress the accumulation of source-modality structural biases during inference, enabling high-quality target-modality generation. Performance comparison with several prevalent generative models demonstrates that RF-CMG achieves superior performance in synthesizing RFID and mmWave signals. We further showcase the effectiveness of the data generated by RF-CMG in gesture recognition tasks, and analyze the impact of the proportion of synthetic data on downstream performance. The code is available at https://github.com/jl-jinglong/RF-CMG.

Abstract:
Change detection is a fundamental task in remote sensing, aiming to quantify the impacts of human activities and ecological dynamics on land-cover changes. Existing change detection methods are limited to predefined classes in training datasets, which constrains their scalability in real-world scenarios. In recent years, numerous advanced open-vocabulary semantic segmentation models have emerged for remote sensing imagery. However, there is still a lack of an effective framework for directly applying these models to open-vocabulary change detection (OVCD), a novel task that integrates vision and language to detect changes across arbitrary categories. To address these challenges, we first construct a category-agnostic change detection dataset, termed CA-CDD. Further, we design a category-agnostic change head to detect the transitions of arbitrary categories and index them to specific classes. Based on them, we propose Seg2Change, an adapter designed to adapt open-vocabulary semantic segmentation models to change detection task. Without bells and whistles, this simple yet effective framework achieves state-of-the-art OVCD performance (+9.52 IoU^c on WHU-CD and +5.50 mIoU^c on SECOND). Our code is released at \hrefhttps://github.com/yogurts-sy/Seg2Changehttps://github.com/yogurts-sy/Seg2Change.

Abstract:
Scene Graph Generation (SGG) produces structured scene representations, yet its training-time backdoor risk remains unexplored. Unlike classification, SGG exposes a pair-level attack surface that enables selective corruption of relations involving a designated object category. However, because SGG largely relies on localized evidence, globally placed triggers may face difficulty in manipulating such predictions. Recent SGG architectures increasingly incorporate global context to improve scene modeling, which may in turn facilitate such attacks. In this paper, we systematically study this backdoor risk and investigate the role of the global context pathway in backdoor behavior. Concretely, we first instantiate this attack with an effective method, and propose a three-axis evaluation protocol to characterize backdoor behavior. Experiments on Visual Genome with representative open-vocabulary and closed-set SGG models demonstrate the effectiveness of the attack. Mechanistic analyses reveal that the explicit global context pathway serves as a major carrier of the backdoor effect across the evaluated SGG models. In particular, replacing the poisoned global context representation with its clean counterpart during inference sharply suppresses the attack, yielding 100% and 97% reductions in attack success for the OvSGTR family and RelationFormer, respectively. Moreover, training-time interventions in the OvSGTR family show that suppressing trigger signals in this pathway either reduces attack success or substantially degrades utility. Together, these findings suggest that explicit global context designs may pose a backdoor risk in SGG architectures. Code is available at \urlhttps://github.com/rafa-cxg/SGGBackdoor.

Abstract:
Multimodal object detection in remote sensing faces challenges due to semantic heterogeneity and modality-specific noise interference. To this end, we propose SuppreSensing, which reformulates multimodal fusion as a selective collaboration process that jointly models shared information and modality-specific cues. SuppreSensing first designs an Expert-driven Multimodal Feature Recalibration (EMFR) module, which reformulates shared-consensus extraction as an input-adaptive multi-expert selection process to alleviate the symmetry trap in multimodal fusion. Complementing this, a modality-specific attribute augmentation strategy is employed to enhance specific modality features by modeling bidirectional discrepancy patterns, mitigating cross-modal heterogeneity. Furthermore, we propose an Expert-driven Customized Feature Purification (ECFP) module based on a "specialized inspection-comprehensive analysis-diagnostic update" physical examination paradigm to iteratively filter redundancies and reinforce task-relevant semantics. Extensive experiments on the DroneVehicle and VEDAI datasets demonstrate that SuppreSensing achieves state-of-the-art detection performance. Cross-domain evaluations on natural scene datasets (FLIR and LLVIP) further validate its superior robustness and generalization capability across diverse environmental conditions. The code will be available at \urlhttps://github.com/Victoria-xin1009/SuppreSensing for reproducibility.

Abstract:
In many open environment applications, data are continuously collected over time, leading to feature space evolution and label space expansion. For example, in forest pest monitoring, when some sensors are replaced with new types, old features may vanish and new features may emerge, thereby forming an evolving feature space. Meanwhile, the emergence of new pest species causes the label space to shift from closed-set to open-set. How to simultaneously achieve theoretically grounded knowledge transfer and unknown-class detection in this setting remains a critical challenge. In this paper, we formalize this important yet underexplored problem as Open-Set Heterogeneous Domain Adaptation with Evolving Features (OSHDAEF) and propose a unified theoretical and methodological framework. Theoretically, we characterize the key properties of OSHDAEF and derive a new upper bound on the target domain generalization error. Practically, we propose a new method based on the theoretical results, termed Latent Variable Inference and Prototype Alignment (LVIPA). Specifically, LVIPA first infers shared latent variables from the composite features of the source and target domains. Then, LVIPA aligns the distributions of cross-domain latent representations using surviving features as a bridge and achieves pseudo-label generation and open-set detection through a two-stage learning strategy. Finally, we conducted extensive OSHDAEF experiments on seven widely used datasets. The effectiveness of LVIPA is demonstrated by outperforming all comparison methods in almost all experiments. Code and supplementary material are available at https://github.com/SunqiangloveSuhairong/LVIPA-Code.

Abstract:
Intelligent fetal ultrasound (US) interpretation is crucial for prenatal diagnosis, but high annotation costs and operator-induced variance make unsupervised pre-training a highly promising paradigm. However, existing pre-training methods largely ignore US-specific characteristicssevere data redundancy, fan-shaped locality, and polar coordinate beamforminglimiting their effectiveness in downstream tasks. To address this, we propose PolarMAE, a novel and efficient pre-training framework tailored for US images. Specifically, to mitigate continuous scanning redundancy, we introduce Progressive Visual-Semantic Screening (PVSS) that adaptively extracts high-value samples, significantly boosting pre-training efficiency. Furthermore, we design Acoustic-Bounded Region Constraint (ABRC) to accommodate US locality, forcing the model to focus strictly on valid acoustic regions rather than invalid dark backgrounds. Finally, leveraging the beamforming prior and local details, we propose Polar-Texture Collaborative Masking (PTCM), enabling the model to capture underlying radial imaging patterns and critical tissue structures. Extensive experiments across diverse datasets and downstream interpretation tasks demonstrate that our method achieves state-of-the-art performance with strong pre-training scalability and efficiency. The code and models are available at https://github.com/lvmenggit/PolarMAE.

Abstract:
Diffusion Transformers (DiT) have emerged as a widely adopted backbone for high-fidelity image and video generation, yet their iterative denoising process incurs high computational costs. Existing training-free acceleration methods rely on feature caching and reuse under the assumption of temporal stability. However, reusing features for multiple steps may lead to latent drift and visual degradation. As recognized in recent work, model outputs evolve smoothly along much of the diffusion trajectory. We leverage this property to enable principled predictions rather than naive reuse. Based on this insight, we propose PrediT, a training-free acceleration framework that formulates feature prediction as a linear multistep problem. We employ classical linear multistep methods to forecast future model outputs from historical information, combined with a corrector that activates in high-dynamics regions to prevent error accumulation. A dynamic step modulation mechanism adaptively adjusts the prediction horizon by monitoring the feature change rate. Together, these components enable substantial acceleration while preserving generation fidelity. Extensive experiments validate that our method achieves up to 5.54x latency reduction across various DiT-based image and video generation models, while incurring negligible quality degradation. Our code is publicly available at https://github.com/hsc113/PrediT.

Abstract:
Octree-based anchor Gaussian Splatting has emerged as a scalable representation for city-scale novel view synthesis, where multi-level anchors adaptively capture scene content from coarse building structures to fine architectural details. However, we identify a fundamental limitation in existing methods: cross-level feature isolation, where each level's anchor features are optimized independently with no inter-level communication, causing color drift on building facades and over-smoothing in textured regions. We present HiCo-GS, a high-fidelity reconstruction framework with two complementary modules. Cross-Level Context Aggregation (CLCA) enables bidirectional hierarchical prior injection by leveraging the octree's spatial containment structure to aggregate per-level context vectors into parent-self-child triplets, fused via a lightweight MLP with residual connection. Coarse-level structural priors flow down to inform fine-level anchors, while fine-level detail statistics feed back to prevent over-smoothing, at negligible computational overhead. Depth-Normal Geometric Consistency (DNGC) regularization enforces agreement between rendered normals and depth-derived normals through an alpha-weighted consistency loss, complemented by edge-aware smoothness losses with progressive warmup that exploit the strong planar priors ubiquitous in urban geometry to suppress floating artifacts. We further introduce the China-Pagoda dataset comprising 8 ancient Chinese pagodas with over 1,200 images each, featuring dense ornamental carvings, curved multi-layer eaves, and repetitive fine-grained textures. Extensive experiments on Mill19, UrbanScene3D, MatrixCity, and China-Pagoda demonstrate that HiCo-GS achieves state-of-the-art rendering quality and substantially cleaner geometry across real-world and synthetic urban benchmarks. Code and dataset will be publicly available.

Abstract:
With the widespread deployment of visible-infrared multimodal perception systems in safety-critical domains such as autonomous driving, evaluating their cross-modal adversarial robustness has become increasingly vital. However, existing approaches exhibit significant limitations in approximating the intrinsic laws of imaging. Most studies either focus on a single modality, failing to bypass cross-modal verification, or simplify infrared modeling into heuristic pixel-intensity distributions, neglecting the impact of ambient temperature fluctuations on adversarial stability. To bridge this gap, this paper proposes MEC-Patch, a cross-modal adversarial attack framework driven by intrinsic physical laws. By leveraging the StefanBoltzmann Law, we establish a physics-grounded cross-spectral mapping that explicitly links material emissivity to thermal radiation. Building on this formulation, we reveal that, under a fixed emissivity distribution, ambient temperature variations induce consistent global scaling while preserving relative emissivity-induced contrast. We exploit this property to construct temperature-robust adversarial perturbations whose discriminative patterns remain stable in the infrared modality, thereby effectively reducing environmental sensitivity. Furthermore, we employ the physics-constrained NSGA-II algorithm to synergistically optimize the material-distribution-based patch parameters effective across both modalities, while enhancing generalization through a Dynamic Adversarial Resampling (DAR) strategy. Experimental results demonstrate that MEC-Patch effectively deceives state-of-the-art multimodal detectors and exhibits high robustness within high-fidelity, physically consistent, and multi-scene simulation environments. This research provides a physical-law-driven perspective for the security assessment of multimodal perception systems. The code is released at \urlhttps://github.com/star0521/MEC-Patch.

Abstract:
Multi-modal fusion has shown promise for crowd counting by leveraging complementary information across modalities. While recent works have explored adaptive fusion strategies, the role of density in modulating modality complementarity remains underexplored. Through systematic empirical analysis on RGB-Thermal and RGB-Depth benchmarks, we discover a critical insight: modality superiority exhibits hierarchical density-dependency, where optimal modality selection is determined by the interaction between global image context and local density patterns, rather than local density alone. The same dense region prefers different modalities depending on the overall image density, revealing a previously overlooked global-local interplay. Beyond this, we further observe that even within the same density level, different spatial regions exhibit heterogeneous modality preferences that a single global context vector cannot capture. Motivated by these findings, we propose a Semantically-Guided Region-Adaptive Mixture-of-Experts (Region-MoE) framework, where four experts with distinct semantic roles---RGB-dominant, Thermal-dominant, Balanced, and Context-aware---are routed by a gating network supervised with density-derived pseudo-labels. This design enables principled, interpretable region-level modality specialization without manual annotation. Extensive experiments demonstrate state-of-the-art performance on RGBT-CC and competitive results across RGB-Thermal and RGB-Depth benchmarks. Code is available at \urlhttps://github.com/EdvinCecilia/RegionMoE.

Abstract:
Cross-modal medical image registration is a fundamental but challenging task due to inherent discrepancies between imaging modalities. Existing methods typically tackle this challenge either by synthesizing target-modality images using generative models (GAN-like) or adopting modality-independent similarity metrics. However, both strategies overlook a crucial aspect of the registration process: How to explicitly distinguish and utilize task-relevant features in multimodal features. In this paper, we propose a novel cross-modal registration network: DECAF (Dual-path Evolving Cascaded Adversarial Filter), which explicitly integrates a Progressive Dual-path Decoupling Framework during the encoding stage. In this framework, we progressively decouple and filter modality-specific (task-irrelevant) components from modality-invariant (task-relevant) features via the Parallel Decoupling Splitter (PDS). Then, we designed an Adversarial Modal-decouple Driver (AMD) to ensure the distinctiveness of the decoupled features. Finally, we designed a Contrastive Optimization Paradigm (COP), which constrains the functional roles of decoupled features to the registration task at a top level. This work highlights the importance of exploiting the underexplored feature learning stage in cross-modal registration. Extensive experiments on three cross-modal medical image datasets demonstrated the superiority of our method in registration performance and robustness. (https://github.com/St-Antonio/DECAF)

Abstract:
Visual state-space models (SSMs) have shown strong potential for medical image segmentation, yet their effectiveness is often limited by two practical issues: axis-biased scan ordering weakens the modeling of oblique and curved structures, and naive multi-branch fusion tends to amplify redundant responses. We present TopoMamba, a topology-aware scan-and-fuse framework for segmenting heterogeneous medical visual media. The method combines a diagonal/anti-diagonal TopoA-Scan branch with the standard Cross-Scan branch to provide complementary structural priors, and introduces ScanCache, a device-aware caching mechanism that amortizes explicit scan-index construction across recurring resolutions. To fuse heterogeneous scan features efficiently, we further propose a lightweight HSIC Gate that regulates branch interaction using a dependence-aware scalar gating rule. We also instantiate a volumetric TopoMamba-3D for practical 3D clinical segmentation. Experiments on Synapse CT, ISIC 2017 dermoscopy, and CVC-ClinicDB endoscopy show that TopoMamba consistently improves segmentation quality over strong CNN, Transformer, and SSM baselines, with particularly clear gains on thin or curved targets such as the pancreas and gallbladder, while maintaining favorable deployment efficiency under dynamic input resolutions. These results suggest that topology-aware scan ordering and lightweight dependence-aware fusion form an effective and practical design for medical multimedia segmentation. The code is publicly available at \hrefhttps://github.com/lzeeorno/TopoMambalzeeorno/TopoMamba.

Abstract:
Person re-identification (ReID) has evolved toward unified multimodal retrieval, learning a shared embedding space for flexible retrieval under arbitrary modality combinations. However, existing unified multimodal ReID methods assume modality completeness during training. In practice, sensor failures, communication interruptions, or constrained acquisition conditions often lead to identity-level missing-modality in training data, where some identities fail to be captured in certain modalities, preventing complete cross-modal supervision. Such missing-modality training data introduce two key challenges: I) insufficient cross-modal correspondences weaken alignment supervision, leading to unstable semantic representations and multi-modal alignment drift; II) identity-level missing-modality makes discriminative identity relations across modality latent spaces limited and inconsistent, reducing per-modality discriminability and causing identity discriminability relation disorder. To address these issues, we propose RobReID, a robust unified multimodal person re-identification framework for training-time missing-modality scenarios. Specifically, we design a Modality-Aware Feature Imputation method that leverages the observed source modality to infer missing-modality features and restore cross-modal supervision. Furthermore, we introduce a Topology-Guided Consistency Alignment strategy that builds a stable topology and preserves identity neighborhood structures across modalities, enhancing relational consistency and discriminability. Extensive experiments show that RobReID outperforms state-of-the-art methods under missing-modality settings, demonstrating effectiveness and robustness. The code is available at \hrefhttps://github.com/Crayi/RobReidhttps://github.com/Crayi/RobReid.

Abstract:
In Multimodal Sequential Recommendation (MMSR), many recent methods treat modality-level features as the basic fusion units. However, we argue that the fusion units are coarse-grained and we empirically find that their performance does not always achieve complementary gains from different modalities under masked-modality fusion analysis. Motivated by this, we revisit MMSR from a new perspective, shifting the basic fusion units from coarse-grained modality-level features to fine-grained signal-patch-level ones, and propose PixelRec. In PixelRec, textual content is transformed into visualized text and integrated with the original image to form a unified visual input, which is then encoded into signal patches serving as the basic fusion units. To adaptively fuse these signal patches into item representations, we introduce a Vision-Recommendation Aggregator. Additionally, to address the high GPU cost associated with feature processing, we design a Re-Construction Compression module to reduce GPU consumption. Extensive experiments demonstrate that PixelRec outperforms state-of-the-art MMSR methods (e.g., PRISM, HM4SR, etc.), achieving improvements of 4.6%-15.9% in terms of recommendation accuracy. The code is available at \urlhttps://github.com/7xjerry1/ACM-MM-26-PixelRec.

Abstract:
Parameter-Efficient Tuning (PET) provides an efficient paradigm for adapting frozen large-scale visionlanguage models to Referring Image Segmentation (RIS). However, RIS heavily relies on fine-grained visionlanguage alignment. Existing PET methods typically improve transfer performance through complex cross-modal interaction designs or additional multi-scale convolutions, while paying relatively limited attention to alignment consistency between intermediate visual and language representations, which in turn restricts the modeling of fine-grained visionlanguage correspondence. Motivated by this observation, we propose BiRIS, a parameter-efficient bidirectional cross-modal alignment framework for RIS. Specifically, the proposed Bidirectional Cross Aligner (BCA) constructs a compact bidirectional interaction space across multiple paired visuallanguage backbone layers, and its core Cross-Alignment Attention (CAA) jointly models cross-modal correlations between intermediate visual and language representations, thereby consistently improving fine-grained bidirectional alignment across layers. In addition, we introduce a Semantic Adaptive Cosine Gate (SACG), which performs fine-grained adaptive gated scaling over visual features conditioned on sentence semantics, enabling more effective modeling of the dynamic modulation of visual responses by sentence-level semantics. Extensive experiments on the RefCOCO, RefCOCO+, and G-Ref benchmarks demonstrate that BiRIS achieves competitive performance while maintaining parameter efficiency. The code is available at https://github.com/KATe-AGI/BiRIS.

Abstract:
Multimodal Sentiment Analysis (MSA) aims to comprehend human sentiment by integrating information from different modalities. However, prior works often use discrete supervision signals to guide continuous sentiment representations, causing the loss of fine-grained emotional information. Furthermore, current dynamic fusion strategies typically assign weights based on feature strength, making models vulnerable to high-intensity modal noise. To address these issues, we propose the Soft Contrastive Enhanced Confidence-Adaptive Fusion Framework (SCAF). Specifically, we first design an intra-modal soft contrastive learning mechanism to establish a unified and continuous feature structure via soft-label guidance. Subsequently, we devise a hierarchical confidence-adaptive fusion module to quantify semantic uncertainty using information entropy, thereby dynamically suppressing unreliable features for robust representation. Extensive experiments on four benchmark datasets demonstrate that SCAF effectively preserves fine-grained sentiment information, enables more robust multimodal fusion, and achieves superior performance. The code will be made publicly available at https://github.com/MKMaS-GUET/SCAF.

Abstract:
With the increasing popularity of federated learning (FL), numerous optimized algorithms have emerged to enhance its performance from different perspectives. However, the presence of algorithm heterogeneity has given rise to isolated algorithm-level silos, posing a new challenge of cross-silo FL. To address this new issue more precisely, we define this problem as cross-algorithm federated learning (CAFL), distinguishing it from traditional cross-silo FL. In this paper, inspired by Kalman filtering in the multi-sensor system, we take a further step to explore the feasibility of aggregating models derived from different FL algorithms through Kalman filtering. Specifically, we propose a two-level fusion framework consisting of state correction and federated fusion, called CAFL-FKF. CAFL-FKF operates without interfering with local training or aggregation process within individual silos. Instead, it leverages uncertainty quantification from Bayesian inference to capture process noise in the CAFL system, which is subsequently mitigated using the federated Kalman filter (FKF). We demonstrate the efficacy of our framework through experiments on real-world datasets compared with nine carefully chosen traditional FL algorithms and six CAFL methods. The experimental results indicate that our method has better convergence performance and rate than the baseline method. Furthermore, the results highlight the potential of CAFL-FKF to address complex FL tasks that traditional cross-silo FL struggles to handle effectively. The code is available in: \urlhttps://github.com/inoryye/CAFL-FKF.

Abstract:
Building pixel-level correspondence between event and image data is a fundamental task for multi-sensor systems. However, existing cross-modal matching methods are largely restricted by their reliance on either matching labels or strictly aligned hardware, which limits them to unlabeled and unconstrained real-world scenarios where neither matching ground truth nor prior sensor relationships are available. To address this, we propose a novel two-stage training paradigm. First, we leverage large-scale data to perform label-agnostic distillation pretraining, upgrading optimization objectives with distribution-based and contrastive losses to learn highly generalizable representations. Second, to tackle unlabeled and unconstrained downstream data, we introduce an epipolar-guided self-distillation framework. By utilizing consistency verification to isolate robust matches and incorporating geometric confidence derived from an external epipolar prior, our model can effectively self-evolve directly on target domains without any supervision. Furthermore, we introduce a rigorous cross-modal evaluation benchmark based on TUM-VIE, featuring physically separated cameras with distinct intrinsic parameters and resolutions. Extensive experiments demonstrate that our proposed method achieves state-of-the-art performance on both MVSEC and TUM-VIE pose estimation tasks. The source code and benchmark will be made publicly available at https://github.com/ZhonghuaYi/nexus2-official.

Abstract:
Backdoor attacks undermine the reliability and trustworthiness of machine learning systems by injecting hidden behaviors that can be maliciously activated at inference time. While such threats have been extensively studied in unimodal settings, their impact on multimodal foundation models, particularly vision-language models (VLMs), remains largely underexplored. In this work, we introduce BackdoorVLM, the first comprehensive benchmark for systematically evaluating backdoor attacks and defenses on VLMs across a broad range of settings. It adopts a unified perspective that injects and analyzes backdoors across core vision-language tasks, including image captioning and visual question answering. BackdoorVLM organizes multimodal backdoor threats into 5 representative categories: targeted refusal, malicious injection, jailbreak, concept substitution, and perceptual hijack. Each category captures a distinct pathway through which an adversary can manipulate a model's behavior. We evaluate these threats using 12 representative attack methods spanning text, image, and bimodal triggers, tested on 2 open-source VLMs and 3 multimodal datasets along with 5 representative defense methods. Our analysis reveals that VLMs exhibit strong sensitivity to textual instructions, and in bimodal backdoors the text trigger typically overwhelms the image trigger when forming the backdoor mapping. Notably, backdoors involving the textual modality remain highly potent, with poisoning rates as low as 1% yielding over 90% success across most tasks. Moreover, existing defenses lack generalization across trigger types and backdoor scenarios, leading to unreliable protection against diverse triggers across modalities and forms. These findings highlight significant, previously underexplored vulnerabilities in current VLMs. We hope that BackdoorVLM can serve as a useful benchmark for analyzing and mitigating multimodal backdoor threats. Code is available at: https://github.com/bin015/BackdoorVLM.

Abstract:
Registration-based Few-shot medical image segmentation (RFMIS) aims to generate pseudo-labels for unlabeled images by warping a labeled image through registration. However, existing methods primarily perform pixel-level optimization and inference in Euclidean space, treating anatomical structures as flat and disjoint. This neglect of inherent hierarchies degrades pseudo-label quality and weakens the discrimination of ambiguous regions, limiting the segmentation performance. To overcome this challenge, we propose a Hyperbolic Hierarchy-aware Aggregative Learning framework for RFMIS, termed HAL, that enhances both deformation plausibility and anatomical discrimination for dual-task learning. Specifically, we introduce a Hyperbolic Hierarchy-aware Infusion (H2I) module, which leverages the hierarchical modeling capability of hyperbolic space to learn precise hierarchy-aware representations via transformation-guided supervised hyperbolic contrastive learning, and injects such hierarchical priors into Euclidean space through a gated infusion block while preserving semantic richness. Furthermore, we propose an end-to-end joint optimization algorithm by gradient aggregation, where the gradients from the registration and segmentation decoders, embedding semantic and hierarchical cues, are aggregated to update the shared encoder to promote collaborative learning across tasks. Extensive experiments on two anatomical regions, with five experimental settings, demonstrate the effectiveness and efficiency of our method in both registration and segmentation. The code is publicly available at https://github.com/JiamingCai469/H2AL.

Abstract:
Deep multi-view clustering (DMvC) has achieved superior performance due to its powerful capability of representation learning and cross-view information integration. However, most DMvC methods ignore cross-view heterogeneity and struggle to semantic alignment among various views. This paper proposes a novel Global-and-Local Collaborative Mixture-of-Experts (GLCMoE) framework for DMvC, which mainly consists of Local MoE (LMoE), Global MoE (GMoE), and Dual Contrastive Learning (DCL) modules. The LMoE filters cross-view heterogeneity via parameters shared expert bank and preserves complementarity rich in fine-grained view-specific information by collaborative experts that are adaptively activated and reorganized via gating routing mechanism. The GMoE with attention-based routing focuses on extracting cross-view consistency and capturing high-order cross-view interaction from unified multi-view feature space. The profound semantic alignment across views is achieved by DCL module from various levels, including coarse-grained granularity among local view-specific representations and fine-grained granularity between local fusion and global representation. Further, the gating routing-balance regularizer is designed to alleviate the expert collapse and improve the availability of experts. The GLCMoE method adheres to a split-then-aggregate, hierarchical contrastive learning strategy, achieving discriminative cross-view representation fusion and semantic alignment. The experimental results on six benchmark datasets verify the effectiveness of the proposed GLCMoE method. The source code is publicly available at https://github.com/Qingguchat/GLCMoE.

Abstract:
Video aesthetic assessment (VAA) aims to predict how aesthetically pleasing a video is, yet remains far less explored than other visual assessment tasks. Its progress is hindered not only by the scarcity of large-scale benchmarks, but also by the intrinsic subjectivity of aesthetic judgment, which is shaped by human perception. In this paper, we revisit VAA from a psychological perspective and propose Peak-End-Net, a lightweight and interpretable framework inspired by the peak-end rule, which suggests that people tend to judge a temporal experience mainly according to its salient moments and the ending. Building on this intuition, we first transfer knowledge from image aesthetic assessment (IAA) to VAA by introducing a pretrained IAA head to produce frame-wise aesthetic priors, which serve as surrogate signals for identifying aesthetically salient moments and guiding peak-end rule-based temporal aggregation. To further capture how a video evolves aesthetically over time, we design an aesthetic rhythm encoder that models temporal progression beyond isolated moments. Additionally, we refine the overall assessment through a dynamic gated fusion mechanism to improve robustness under distribution shift. Our method is built on a frozen vision transformer (ViT) and requires only a small number of trainable parameters, making it scalable and parameter-efficient. Extensive experiments on two existing VAA benchmarks, including in-domain evaluation on VADB and cross-domain testing on DIVIDE-3K, demonstrate that our approach achieves state-of-the-art performance, affirming the value of psychologically grounded modeling for VAA. Our code and models are available at https://github.com/AMAP-ML/Peak-End-Net.

Abstract:
Visual Retrieval-Augmented Generation (VRAG) empowers Vision-Language Models to retrieve and reason over visually rich documents. To tackle complex queries requiring multi-step reasoning, agentic VRAG systems interleave reasoning with iterative retrieval. However, existing agentic VRAG faces two critical bottlenecks. (1) Visual Evidence Sparsity: key evidence is scattered across pages yet processed in isolation, hindering cross-page reasoning; moreover, fine-grained intra-image evidence often requires precise visual actions, whose misuse degrades retrieval quality; (2) Search Drift in Long Horizons: the accumulation of visual tokens across retrieved pages dilutes context and causes cognitive overload, leading agents to deviate from their search objective. To address these challenges, we propose VISOR (Visual Retrieval-Augmented Generation via Iterative Search and Over-horizon Reasoning), a unified single-agent framework. VISOR features a structured Evidence Space for progressive cross-page reasoning, coupled with a Visual Action Evaluation and Correction mechanism to manage visual actions. Additionally, we introduce a Dynamic Trajectory with Sliding Window and Intent Injection to mitigate search drift. They anchor the evidence space while discarding earlier raw interactions, preventing context from being overwhelmed by visual tokens. We train VISOR using a Group Relative Policy Optimization-based Reinforcement Learning (GRPO-based RL) pipeline with state masking and credit assignment tailored for dynamic context reconstruction. Extensive experiments on ViDoSeek, SlideVQA, and MMLongBench demonstrate that VISOR achieves state-of-the-art performance while maintaining reasonable and controllable computational costs for long-horizon visual reasoning tasks. The source code is available at \urlhttps://github.com/syc1336/VISOR.

Abstract:
Graph clustering aims to partition nodes in a graph into distinct groups without ground truth labels. Typically, structural and attribute information on graphs characterize semantics at different levels. This difference may induce semantic discrepancy, leading to differentiated clustering structures. However, existing methods generally neglect this semantic discrepancy. They often perform indiscriminate fusion of structures and attributes during representation learning, which ultimately results in suboptimal clustering performance. To address this issue, we propose a novel Dual-view Uncertainty Driven Entrusted Teaching framework for contrastive graph clustering, named DUET. DUET first employs a dual-level contrastive learning module to globally align embedding and cluster semantics across structure and attribute views. Nevertheless, global alignment inevitably leaves residual nodes with inconsistent cross-view predictions, which we define as hard samples. To prevent them from corrupting cluster topologies and fully exploit cross-view information, we design an innovative criterion, named Gamma, to identify trustworthy hard samples. Accordingly, an entrusted teaching mechanism mutually calibrates the structural and attribute views, using confident predictions from one as supervision for the other. Finally, a cross-attention module adaptively integrates these refined representations, generating highly discriminative embeddings for clustering. Extensive experiments on five benchmark datasets demonstrate our proposed DUET's state-of-the-art performance, and our code is available at \urlhttps://github.com/guguolibai/DUET.

Abstract:
Visual Retrieval-Augmented Generation (VRAG) has emerged as a promising paradigm for equipping Vision-Language Models (VLMs) with external visual evidence, enabling them to go beyond parametric knowledge when answering visually grounded questions. However, in such multi-image settings, VLMs still often suffer from visual hallucinations and struggle to accurately identify the question-relevant evidence needed for reliable reasoning. Existing methods usually lack an explicit cross-image evidence collection process, and also provide limited credit assignment when jointly optimizing perception and reasoning. To address this issue, we propose EVisRAG, an evidence-guided visual retrieval-augmented framework for multi-image reasoning. EVisRAG first observes the retrieved images, records question-relevant visual evidence from each image, and then performs reasoning and answer generation based on the aggregated evidence. We further introduce RS-GRPO, which aligns reward signals with token spans from different stages, improving training stability and strengthening the joint optimization of evidence localization and reasoning. Experiments on multiple visual question answering benchmarks show that EVisRAG consistently outperforms the backbone VLM by an average of about 19%, while substantially reducing visual hallucinations. These results demonstrate that explicit evidence collection and scoped reward design are effective for improving visual grounding and reasoning reliability in multi-image settings. Codes and data are available at \urlhttps://github.com/OpenBMB/VisRAG.

Abstract:
Graph Neural Networks (GNNs) have emerged as a cornerstone for representing complex relational dependencies in diverse multimedia tasks, particularly in cross-platform user interest modeling and cross-modal semantic alignment. In the real world, a practical defense against graph adversarial perturbations is needed. However, we observe that the prevailing adversarial purification methods are essentially domain-restricted defenses, which leads to the following shortcomings: (1) single-domain data provides insufficient structural and semantic diversity for learning robust purification criteria; (2) training of domain-specific defense strategies from scratch consumes substantial computational cost. To address the above limitations, we propose a transferable graph purification scheme, named ProGAP, to bridge adversarial defense knowledge via vulnerability-aware graph prompt learning. Firstly, to capture universal adversarial patterns, a perturbation-capture edge detector is pretrained on data-rich graphs by jointly modeling topological and semantic information. Subsequently, to achieve more knowledge transfer w.r.t. robustness, vulnerability-aware prompts are designed that inject targeted purification guidance into biased nodes, during which the pretrained detector adapts to distribution shifts in downstream graphs without parameter-laborious updates. Experimental results demonstrate that compared with state-of-the-art baselines, our ProGAP achieves 1%-9% improvement, and reduces the time consumption by up to 2.2x. The code for ProGAP is available at https://github.com/Lieyoufffff/ProGAP.

Abstract:
Multimodal embodied agents are increasingly required to solve long-horizon tasks by integrating visual observations, textual goals, and interaction history into closed-loop decision making. However, state-of-the-art large-model-based planners often rely on a single dominant planning style during execution. Once this execution mode becomes ineffective, the agent may remain stalled for many steps, repeatedly interacting with the environment without making meaningful progress. We address this limitation by proposing a Quality-Diversity (QD) framework for discovering diverse planning policies for multimodal embodied agents. The proposed method treats planning-policy templates as evolvable individuals and organizes them into a behavior-indexed archive rather than collapsing search to a single prompt style. In the offline stage, rollout trajectories are summarized into structured success and failure experiences, which guide policy variation through recombination and experience-guided mutation. The resulting policies are mapped into a behavior space defined by interaction intensity and goal-directedness, and the highest-quality policy in each niche is retained in the archive. In the online stage, the agent executes one policy at a time while monitoring task progress. When persistent stall is detected, the system rolls back to the latest checkpoint and switches to a behaviorally distinct archive policy to resume execution. Experiments on the ThreeDWorld transport benchmark show that the proposed framework improves both task success and interaction efficiency over representative baseline planners. These results suggest that discovering diverse policy repertoires is an effective way to support adaptive multimodal planning and online failure recovery. Code is available at https://github.com/EvoNexusX/2026XuQD-Plan.git.

Abstract:
Existing weakly supervised referring expression comprehension (WREC) and segmentation (WRES) methods perform anchor-text matching using flat sentence-level representations, thereby discarding the multi-granularity hierarchical semantics inherent in both visual and linguistic modalities. Because referring expressions naturally form hierarchical clauses of increasing specificity and flat Euclidean space distorts such tree-like structures, explicit hierarchical modeling is necessary yet overlooked. We propose HCNet (Hyperbolic hierarchical Clause Network), which first decomposes each expression into four hierarchical clauses via dependency parsing, then refines sentence representations through a hierarchical text enhancer that attends to coarser clause levels. Subsequently, both visual and textual features are projected into a dual Euclidean-Lorentzian hyperbolic space, where an entailment cone loss enforces hierarchical nesting on the text side and a structure alignment loss transfers this hierarchy to the visual modality. Experiments on RefCOCO, RefCOCO+, and RefCOCOg demonstrate that HCNet achieves state-of-the-art on both WREC and WRES, with improvements of up to +3.54% over the previous best. The code is available at \urlhttps://github.com/VILAN-Lab/HCNet.

Abstract:
Existing Video Object Segmentation (VOS) solutions primarily rely on short-range frame-to-frame propagation and monolithic spatial representations. % Consequently, they remain vulnerable in extended sequences where objects undergo severe occlusions or complex non-rigid deformations. % To alleviate these limitations, we propose Joint Temporal Conduction and Structural Experts Decomposition (ConStruct), a unified method that balances long-range temporal consistency and precise structural semantics. % Instead of fragile discrete matching, our temporal conduction mixer models temporal evolution as a continuous physical conduction process. This explicitly drives global temporal diffusion, naturally bridging information gaps caused by prolonged occlusions. % To accommodate complex structural diversity across dynamic scenes, a dynamic structural experts decomposer routes temporally-mixed features into specialized pathways via a heterogeneous mixture-of-experts. % Furthermore, a frequency-aware decoder leverages high-frequency cues to recover crisp boundaries during multi-scale feature fusion. % Extensive evaluations across five SVOS benchmarks (DAVIS-2017, YouTube-VOS 2018/2019, MOSE, LVOS) and three UVOS benchmarks (DAVIS-2016, FBMS, YouTube-Objects) demonstrate ConStruct's universal effectiveness. Our method achieves highly competitive performance against state-of-the-art approaches, highlighting the immense potential of a unified temporal-structural inference paradigm. % The code is available on \hrefhttps://github.com/Bruce1eezz/ConStruct-VOS\textcolormagentahttps://github.com/Bruce1eezz/ConStruct-VOS.

Abstract:
LiDAR data generation holds considerable potential for scalable 3D scene perception in autonomous systems. Existing works typically adapt image-domain generative architectures and train from scratch, but the lack of comparable-scale LiDAR data bottlenecks their accuracy in modeling 3D geometric details. The notable semantic consistency between LiDAR and image modalities provides an opportunity to utilize the image diffusion model pretrained on massive image data for LiDAR generation. However, the substantial modality gap poses a significant challenge for straightforward adaptation. To address this question, we propose IMPEL, a novel framework for LiDAR point cloud generation by exploiting the prior knowledge contained in the pretrained image diffusion. Specifically, we first utilize a Hough voting-based method to project 3D LiDAR point clouds into 2D range images accurately. We then propose a dual spatial consistency optimization strategy to guide the fine-tuning of the pretrained image diffusion model to perceive structural semantics and spatial relationships from both 2D and 3D aspects. Additionally, we design a LiDAR-specific ControlNet that integrates the LiDAR scanning characteristics into the control mechanism for camera-to-LiDAR conditional generation, which holds significant practical value. Experiments conducted on the 64-beam and 32-beam scenarios demonstrate the effectiveness of our framework, substantiating that the rich prior knowledge embedded in the pretrained image diffusion model can be effectively transferred to LiDAR point cloud generation. Code is available at https://github.com/SINCOSLab/IMPEL.

Abstract:
Stereo matching and surface normal estimation are fundamental tasks in 3D vision. However, existing feed-forward stereo methods still struggle to produce reliable predictions in challenging regions, mainly due to the lack of strong geometric priors. In this paper, we propose GeoStereo, a unified stereo geometry estimation framework that leverages powerful diffusion priors to jointly predict disparity and surface normals. Specifically, GeoStereo couples a feed-forward stereo matching pipeline with a diffusion-based normal estimation branch. To enable effective interaction between the two tasks, we introduce a disparity to normal initialization strategy and construct a warp to left-view condition for the diffusion process. This coupled design allows the diffusion branch to provide strong structural priors that enhance disparity estimation in ill-posed regions, while the feed-forward branch offers reliable geometric guidance for accurate normal prediction. Extensive experiments show that GeoStereo performs reliably in challenging scenarios, including low-light environments, highly reflective surfaces, and transparent objects. Under zero-shot settings, it achieves Rank-1 disparity estimation on multiple benchmarks, including KITTI and NYUv2, and delivers the best normal estimation accuracy on many real indoor benchmarks, such as iBims-1 and ScanNet. Project page: \urlhttps://qz-wei.github.io/GeoStereo.github.io/

Abstract:
Autonomous UAV systems increasingly rely on multimodal large language models (MLLMs) to operate in complex real-world environments. Such embodied scenarios require not only understanding the surrounding space but also maintaining a coherent representation of the agent itself. However, existing UAV-oriented approaches and benchmarks remain largely environment-centric, primarily focusing on spatial understanding tasks, with the agent's self-awareness remaining implicit. To address this gap, we introduce SIS-Bench, a benchmark for evaluating embodied spatial intelligence in UAV scenarios under a unified self-in-space formulation. SIS-Bench organizes evaluation along two complementary dimensions, \emphspace and \emphself, and a three-level hierarchy of perception, memory, and reasoning. It contains 4,856 question--answer pairs across 13 tasks derived from 1,646 real-world UAV videos through a task-conditioned construction pipeline with expert verification. Extensive evaluations reveal that current MLLMs exhibit fundamental limitations in modeling dynamic and agent-centered processes. In particular, we observe a clear imbalance between spatial cognition and self-awareness, as well as a progressive performance degradation across cognitive levels. Motivated by these findings, we further explore a motion-aware representation that incorporates self-related dynamics through optical flow and visual feature fusion. Experimental results show that modeling agent motion consistently improves perception and memory performance, not only in spatial cognition but also in self-awareness, and generalizes to downstream UAV decision-making tasks. Our results highlight the importance of self-awareness for advancing embodied spatial intelligence, and provide both a new benchmark and empirical evidence for motion-aware self-in-space modeling.

Abstract:
RGB-based 3D hand pose estimation (HPE) in dynamic real-world environments is a foundational technology for human-computer interaction. Although current methods achieve remarkable performance under controlled conditions, they remain sensitive to extremely low-light conditions. Specifically, it involves low signal-to-noise ratios and missing visual cues, making it difficult to acquire precise 3D motion annotations from RGB. As a compromise, existing low-light datasets capture static poses, failing to reflect real-world complex dynamic motions. To this end, we develop a multi-modal, multi-view capture system with adjustable illumination and a high-fidelity illumination-invariance automated annotation pipeline, decoupling 3D annotation from adverse lighting, preserving photometric realism and illumination variations. Based on this pipeline, we propose DimHand, the first large-scale 3D hand pose dataset capturing continuous motion under dynamic low-light conditions, comprising 745,280 frames from 15 subjects performing 15 distinct sequential actions across 5 viewpoints with illumination ranging from 0.1 to 10 lux. However, the coupling between severe visual degradation and the intrinsic complexity of hand structure induces spatial and semantic ambiguity. To address these challenges, we propose DimNet, an end-to-end framework that leverages 3D structural priors via a Kinematic Prior Codebook (KPC) for structural regularization and a Spatial-Semantic Projection (SSP) for local semantic refinement. Extensive experiments demonstrate that DimNet outperforms state-of-the-art methods, particularly in generalizing to unseen dynamic hand motions, thereby establishing a strong low-light baseline for 3D HPE. Our dataset and supplementary materials are available on the project website: https://hanlingzhan.github.io/DimHand.

Abstract:
Electroencephalography (EEG) connectivity graphs provide an effective representation of neural interactions, but most existing methods model them as static or single-view graphs, limiting their ability to capture temporally evolving and complementary brain-network dynamics. To address this, we propose HiGMR-Net (Hierarchical Graph Memory Refinement Network), a hierarchical graph sequence learning framework for classifying Frontotemporal Dementia (FTD) and Healthy Controls (HC) from EEG. The proposed model forms a sequence of Event-Related Potential (ERP) peak-wise EEG connectivity graphs that capture temporally varying neural interactions. It further introduces a hierarchical memory-guided multi-view fusion mechanism that jointly models channel-level and region-level graphs, where graph-level memory preserves persistent region-wise connectivity patterns and edge-level refinement updates channel-level interactions using higher-level graph context. The resulting representations are encoded by a shared graph neural network and aggregated for classification. Using subject-wise stratified 5-fold cross-validation, HiGMR-Net achieves a mean accuracy of 95.34% \pm 4.29% for FTD versus HC classification, with a 9.00% accuracy gain over the strongest multi-view baseline model. These findings highlight the significance of hierarchical multi-view fusion for robust EEG-based FTD versus HC classification. The source code is available at: \urlhttps://github.com/DDasChakladar/HiGMR-Net.

Abstract:
Existing image manipulation localization (IML) methods rely heavily on 2D forensic cues, such as low-level artifacts, noise traces, and semantic inconsistencies in the manipulated image. While effective in many cases, these cues become much less discriminative when manipulated regions are well blended with their surrounding context in appearance. In such cases, a manipulated region may remain locally appearance-consistent, but still violate the geometric structure of the surrounding scene. This limitation motivates us to go beyond purely 2D evidence and introduce geometric reasoning into IML. To this end, we leverage monocular reconstruction to obtain auxiliary geometric cues, including depth and surface normals. However, a key challenge lies in the fact that reconstructed geometry on manipulated images is inherently noisy and cannot be used naively. Rather than treating depth and normals as direct evidence, we estimate their reliability and exploit them selectively for localization. Based on this principle, we design a geometry-aware framework (GFrame) that fuses reliable geometric cues with RGB features and propagates them across scales to improve fine-grained localization. Extensive experiments show that the proposed method achieves excellent performance under limited budget constraints. These results indicate that reliable 3D geometry provides complementary forensic evidence beyond traditional 2D cues for IML. The code is available at https://github.com/vpsg-research/GFrame.

Abstract:
While the privacy risks of multimodal large language models (MLLMs) have drawn significant attention, the unique vulnerabilities of domain-specific MLLMs remain largely underexplored. Focusing on document understanding MLLMs for identity document processing, this paper investigates the privacy issues inherent in Key Information Extraction (KIE) tasks. We reveal that when input images lack sufficient visual evidence, these models often rely on memorized field relations from training data to infer missing content, thereby leaking multiple correlated fields containing sensitive personal information. To mitigate this risk, we make three key contributions. First, we propose the Dynamic Relational Unlearning Framework (DRUF) which comprises a Relational Decoupling Unlearning (RDU) module and a dynamic set update mechanism. It suppresses the leakage of high-risk field pairs while preserving KIE performance. Second, we introduce DocPrivacyBench, a novel benchmark to systematically evaluate a model's susceptibility to privacy leakage under conditions of absent or minimal visual evidence. Third, we evaluate three MLLMs and six unlearning methods using this benchmark, assessing both post-unlearning leakage suppression and utility preservation. Our results demonstrate that existing MLLMs consistently exhibit privacy leakage when visual evidence is scarce, particularly on noisier datasets. In contrast, DRUF outperforms the strongest baseline by improving leakage suppression by 4.8 percentage points, effectively mitigating privacy risks while maintaining robust document information extraction performance. The code and benchmark are available at \urlhttps://github.com/xubeining/Beyond-Visual-Evidence.

Authors: Di Wen (Karlsruhe Institute of Technology), Zeyun Zhong (Karlsruhe Institute of Technology), David Schneider (Karlsruhe Institute of Technology), Manuel Zaremski (Karlsruhe Institute of Technology), Linus Kunzmann (Karlsruhe Institute of Technology), Yitian Shi (Karlsruhe Institute of Technology), Ruiping Liu (Karlsruhe Institute of Technology), Yufan Chen (Karlsruhe Institute of Technology), Junwei Zheng (ETH Zurich), Jiahang Li (Karlsruhe Institute of Technology), Jonas Hemmerich (Karlsruhe Institute of Technology), Qiyi Tong (Istituto Italiano di Tecnologia), Patric Grauberger (Karlsruhe Institute of Technology), Arash Ajoudani (Istituto Italiano di Tecnologia), Danda Pani Paudel (INSAIT, Sofia University 'St. Kliment Ohridski'), Sven Matthiesen (Karlsruhe Institute of Technology), Barbara Deml (Karlsruhe Institute of Technology), Jrgen Beyerer (Karlsruhe Institute of Technology), Luc Van Gool (INSAIT, Sofia University 'St. Kliment Ohridski'), Rainer Stiefelhagen (Karlsruhe Institute of Technology), Kunyu Peng (Karlsruhe Institute of Technology)

Abstract:
We introduce IMPACT, a synchronized five-view RGB-D dataset for deployment-oriented industrial procedural understanding, built around real assembly and disassembly of a commercial angle grinder with professional-grade tools. To our knowledge, IMPACT is the first real industrial assembly benchmark that jointly provides synchronized ego--exo RGB-D capture, decoupled bimanual annotation, compliance-aware state tracking, and explicit anomaly--recovery supervision within a single real industrial workflow. It comprises 112 trials from 13 participants totaling 39.5 hours, with multi-route execution governed by a partial-order prerequisite graph, a six-category anomaly taxonomy, and operator cognitive load measured via NASA-TLX. The annotation hierarchy links hand-specific atomic actions to coarse procedural steps, component assembly states, and per-hand compliance phases, with synchronized null spans across views to decouple perceptual limitations from algorithmic failure. Systematic baselines reveal fundamental limitations that remain invisible to single-task benchmarks, particularly under realistic deployment conditions that involve incomplete observations, flexible execution paths, and corrective behavior. The full dataset, annotations, and evaluation code are available at \urlhttps://github.com/Kratos-Wen/IMPACT.

Abstract:
Scene Graphs (SGs) provide structured representations of visual scenes by modeling objects and their pairwise relationships. Despite recent progress, existing datasets primarily focus on generic natural contexts, leaving domain-specific and function-oriented scenes largely underexplored. This limitation restricts the evaluation of relational reasoning in scientific experimental scenes, thereby hindering the development of intelligent monitoring, analysis, and related applications in such scenes. To address this gap, we introduce PhysScene, the first SG dataset tailored to physics experiments. PhysScene encompasses specialized instruments, structured experimental setups, and functional relations intrinsic to experimental environments, enabling reasoning that extends beyond spatial co-occurrence to logical dependencies. Rather than pursuing large data scale, PhysScene focuses on strong semantic constraints and high relation density in experimental scenes, posing new challenges for existing scene parsing algorithms while offering opportunities for further improvements. Extensive analyses and experiments show that PhysScene complements existing benchmarks and establishes a valuable testbed for advancing scientific visual reasoning. The dataset is publicly available at \hrefhttps://github.com/ZMH-SDUST/PhysScenehttps://github.com/ZMH-SDUST/PhysScene.

Abstract:
The scholarly exegesis of ancient Chinese characters demands integrating visual observation, linguistic analysis, and historical context. However, existing computational approaches focus narrowly on subtasks such as character recognition and retrieval, lacking the structured datasets and benchmarks required for comprehensive scholarly analysis. To address this limitation, we introduce Ancient Chinese Character Exegesis (ACCE), a vision-language question answering (VQA) task that models the scholarly exegesis process. ACCE is organized into four progressive levels: basic character identification, glyph-form analysis, meaning exegesis, and diachronic evolution analysis. To support this task, we construct two complementary resources. JieZi-Dataset is the first large-scale, expert-audited VQA training dataset for ACCE, comprising over 500K QA pairs. It is constructed via a pipeline that reduces factual errors by constraining generation with expert-designed templates and source-text references. Human verification is further applied at each key stage to ensure scholarly accuracy. JieZi-Bench is an evaluation benchmark aligned with the exegesis process, constructed and verified by human experts to ensure evaluation reliability. It consists of four levels with reference answers curated from authoritative lexicographic works held separate from the training data. Experiments on multimodal large language models show that current models perform well on basic identification but struggle with glyph analysis, semantic reasoning, and diachronic understanding. Fine-tuning on JieZi-Dataset substantially improves performance across all four levels. Code and dataset are available at https://github.com/Ran00w/JieZi.

Abstract:
Long video question answering (Long-Video QA) increasingly relies on agentic tool use to retrieve evidence from long videos. In realistic settings, this process often requires multi-hop retrieval, where agents must iteratively gather multiple discontinuous evidence clips. However, existing long-video benchmarks are largely static: they rarely enforce strict multi-hop retrieval and typically lack a standardized evidence-access interface, making it difficult to separate failures in retrieval planning from those in answer generation. To address this gap, we introduce LongVidSearch, a benchmark for evaluating agentic multi-hop evidence retrieval planning in long videos under standardized access constraints. LongVidSearch enforces retrieval necessity: a Hop-k question requires exactly k necessary evidence clips, and removing any single clip renders the question unsolvable. The benchmark contains 3,000 questions over 447 long videos (average length 26 minutes), covering four reasoning categoriesState Mutation, Causal Inference, Global Summary, and Visual Trackingwith 2-, 3-, and 4-hop evidence requirements. To ensure fair and controlled evaluation, all agents interact with LongVidSearch through a unified tool interface, which fixes the retrieval backend and isolates the agents ability to formulate queries and plan iterative retrieval. In addition to answer accuracy, we measure tool-call cost to analyze the accuracyefficiency trade-off under identical access conditions. We evaluate VideoAgent-style QA agents with multiple backbone LLMs using three-judge majority voting. GPT-5 achieves the highest accuracy (42.43), outperforming Gemini 3 Pro (30.97) and GPT-4o (19.20), yet remaining below 50%, highlighting the difficulty of multi-hop retrieval planning. With gold evidence clips, performance becomes near-perfect, confirming retrieval planning as the primary bottleneck. Our code and data are available at \urlhttps://longvidsearch.github.io/.

Abstract:
The growing integration of vision and language models is driving a fundamental shift in video understanding and processing. This evolution calls for datasets that jointly capture visual content and its semantic representations at scale. To address this need, we introduce \dataset, a large-scale curated multimodal dataset comprising \num10000 high-fidelity 4K video sequences at 60\,fps with rich semantic and perceptual annotations. We developed an automated acquisition pipeline to curate videos from the Pexels repository, using targeted search queries and strict filtering criteria to capture a wide range of real-world scenes. Beyond the video sequences, \dataset is enriched with comprehensive multimodal annotations that integrate low-level visual features with high-level semantic information. These include LLM-generated semantic descriptors, no-reference quality metrics, spatial-temporal complexity metrics, and visual diversity attributes. By combining structured annotations with high-quality video data, \dataset provides a versatile resource for a wide range of applications, including video enhancement, content-aware compression and streaming, neural video coding, multimodal learning, generative video modeling, and perceptual quality modeling. Dataset URL: \hrefhttps://cd-athena.github.io/LMM-10Khttps://cd-athena.github.io/LMM-10K.

Abstract:
Existing instructional video benchmarks mainly focus on observable actions or coarse procedural steps, providing limited support for pedagogical events whose interpretation depends on lesson context and instructional function. We introduce PEA (Pedagogical Event Analysis), a hierarchical benchmark for function-aware event detection in real classroom videos, where visually similar behaviors may play different pedagogical roles across lesson contexts. PEA contains 100 classroom videos with 10,343 annotated instances from Mathematics and Information Technology classrooms. It defines a two-level taxonomy of 32 classes, including 5 lesson stages and 27 fine-grained verbal and non-verbal events. To support reliable evaluation, PEA adopts annotator training, double annotation on 10% of the data, senior adjudication, and an official fixed train/validation/test split, achieving a Fleiss' Kappa of 0.78. We provide representative baselines and release annotations, taxonomy files, split files, evaluation tools, and baseline code, while raw videos are available under controlled access due to privacy constraints. Project page: https://github.com/mark1001-ljj/dataset_event.git.

Abstract:
The rapid development of Multimodal Large Language Models (MLLMs) has led to growing interest in egocentric video understanding, specifically the ability for MLLMs to recognize fine-grained hand-object interactions, track object state changes over time, and reason about manipulative processes in dynamic environments from a first-person perspective. However, existing egocentric video benchmarks suffer from limited grounded rationale evaluation, offering limited support for fine-grained operation-centric reasoning and rarely examining whether model rationales are grounded in explicit spatio-temporal evidence. To address this gap, we introduce EgoCoT-Bench, a fine-grained egocentric benchmark for grounded and verifiable operation-centric reasoning with explicit step-by-step rationale annotations. Overall, EgoCoT-Bench comprises 3,172 verifiable QA pairs over 351 egocentric videos separated into four task groups for a total of 12 subtasks, encompassing perception and retrospection, anticipation, and high-level reasoning. The benchmark is constructed through a spatio-temporal scene graph (STSG)-guided generation framework and is further refined by human annotators to ensure correctness, egocentric relevance and fine-grained quality. Experimental results show continuing difficulties with egocentric fine-grained reasoning and further reveal that many multimodal models produce explanations that are answer-correct, but have evidence that is inconsistent with the answer. We hope EgoCoT-Bench can serve as a useful testbed for grounded and verifiable reasoning in egocentric video understanding. Project page and supplementary materials are available at: \urlhttps://dstardust.github.io/EgoCoT/.

Abstract:
Recent advances in multimodal large language models (MLLMs) have significantly improved the performance of multimodal emotion recognition (MER) and enabled interpretable description generation by jointly modeling video, audio, and language, etc. However, these performance improvements are often accompanied by an increase in model parameter size (e.g, at least 7B), which simultaneously incurs high computational costs and reduces inference efficiency, thereby hindering real-time deployment on resource-constrained platforms such as robots and mobile devices. This raises a fundamental question: do we really need the multimodal MER model larger than 1B parameters for high-quality MER? In this paper, we challenge the assumption that larger models are inherently necessary and proposes a lightweight MER framework (called Light-MER), which achieves better and faster multimodal sentiment understanding and recognition through knowledge distillation. It can transfer knowledge from a strong, large-scale teacher model to a lightweight sub-billion-parameter student model, aiming to preserve rich multimodal emotion reasoning and recognition while substantially improving deployment efficiency. Specifically, we introduce two new optimization strategies to enhance knowledge transfer: (1) a new optimal transport loss that combines Sliced Wasserstein Distance with hidden-state alignment, and (2) a new multi-reward optimization strategy based on GRPO that balances MER performance and efficiency, aimed at further enhancing the learning capabilities of student models. Extensive experiments on nine benchmark datasets demonstrate that Light-MER achieves state-of-the-art performance while significantly improving inference efficiency. This highlights the strong potential of small multimodal emotion language models for future research. Code is available at https://github.com/GAIR-Lab/Light-MER.

Abstract:
Existing remote-sensing change captioning benchmarks mainly reward surface-level descriptions of visual differences, but provide limited support for evaluating whether models can perform rule-constrained semantic reasoning over observed changes, filter them under application-specific constraints, and align their outputs with geographic expert judgment. This limitation exposes a critical gap between current benchmark design and the demands of real-world Earth observation, where practical workflows require not only semantic understanding, but also higher-level reasoning and filtering. To address this gap, we introduce Geo-BC, a large-scale, preference-aligned multimodal benchmark that reframes change analysis as structured change interpretation rather than caption generation. The benchmark contains 109224 spatiotemporal samples, supplemented by 327672 reasoning clues and 382284 review comments, enabling process-level supervision beyond final-answer evaluation, as well as a curated Pro subset for challenging evaluation. In addition to benchmark instances, Geo-BC provides language-grounded reasoning trajectories, hard negative supervision for pseudo-change rejection, uncertainty-aware confidence annotations, and chosen/rejected preference pairs that support supervised fine-tuning, reward modeling, preference optimization, and fine-grained error diagnosis. We further establish baseline results across representative multimodal large models, highlighting the persistent difficulty of structured geo-world understanding. By focusing on the under-explored problem of semantically grounded change reasoning, Geo-BC provides critical data infrastructure to support the development of geospatial semantic spatiotemporal understanding. The dataset, generation code, and baseline models are publicly available at: https://github.com/kaynqi/Beyond-Captioning.

Abstract:
Vision-Language Models (VLMs) often lose visual grounding during multi-step reasoning: as reasoning chains grow longer, later inference steps rely increasingly on language priors rather than image evidence. We identify a consistent benchmark-level signature associated with this degradation: across 2,510 re-examined samples from four benchmarks, attention entropy over image tokens typically decreases during Round 1 and rises again after image re-injection. However, we find that effective visual re-examination requires two complementary ingredients: image re-injection and targeted self-diagnosis. Without targeted diagnosis, re-examination can even hurt performance, whereas accurate self-diagnosis yields substantial gainsa swing of several points on key benchmarks, indicating that diagnostic quality is a key factor in whether re-examination helps or hurts in our setting. We present ReGround, a two-stage framework that teaches VLMs to self-diagnose grounding failures and selectively re-examine visual evidence, without architectural modifications or external tools. Through capability bootstrapping, a stronger variant from the same model family provides diagnostic scaffolding only during data construction, while the policy model learns to diagnose autonomously at inference time and retains most of the assisted gains. Experiments on eight benchmarks across two VLM backbones demonstrate consistent gains, especially on visually intensive multi-step reasoning tasks, while incurring only modest inference overhead relative to tool-augmented baselines. Project page: https://sespoir.github.io/reground-page/. Code: https://github.com/sespoir/ReGround.

Abstract:
Large Language Models (LLMs) have recently shown exceptional potential in time series forecasting (TSF), leveraging their inherent sequential reasoning capabilities to model complex temporal dynamics. Existing approaches typically employ an autoregressive generation strategy to adapt LLMs for TSF. However, we identify a theoretical flaw in this paradigm: during inference, the model operates in an open-loop manner, recursively consuming its own generated outputs. This leads to error accumulation, where minor early deviations cascade into significant rollout drift over long horizons. In this paper, we reformulate autoregressive forecasting through the lens of control theory, proposing Feedback-driven LLM (F-LLM), a novel closed-loop framework. Unlike standard methods that passively propagate errors, F-LLM actively stabilizes the trajectory via a learnable residual estimator functioning as a system observer. Furthermore, we provide a mathematical proof that, under explicit contraction assumptions, this closed-loop mechanism guarantees a uniformly bounded step-wise error sequence within the local surrogate dynamics. Extensive experiments demonstrate that F-LLM significantly mitigates error propagation, achieving good performance on time series benchmarks. Our code is publicly available at \urlhttps://github.com/Zh-XY22/F-LLM.

Abstract:
Biomedical image analysis spans diverse modalities and tasks, yet real-world deployment is hindered by severe distribution shifts across scanners, protocols, and patient populations. High-performing models consequently require repeated domain-specific fine-tuning, which is a costly cycle that becomes impractical when labels are scarce or privacy constraints limit data sharing. We propose OPERA (Offline Policy-guided Expert Routing and Adaptation), a multi-agent ensemble framework that addresses this deployment bottleneck by treating expert weight assignment as an offline policy learning problem: a routing policy is learned from a small validation set without gradient updates to any expert agent, then deployed with test-time adaptation to handle distribution shift. OPERA coordinates heterogeneous specialist agents through complementary mechanisms. The expert profiling module learns selection policies offline, enabling informed allocation of expertise. Each agent undergoes confidence calibration through temperature adjustment, ensuring more reliable probabilistic outputs. OPERA also incorporates distribution aware adaptation, where class weights are dynamically adjusted at the batch level using statistics derived from unlabeled test data. Instance level routing assigns each sample to the most suitable expert by leveraging inter model agreement and predictive entropy. We evaluate OPERA on 9 datasets covering fundus photography, chest X-ray, CT, MRI, and multimodal diagnostic benchmarks, comparing against 30+ baselines across classification, segmentation, and multimodal settings. OPERA consistently improves performance and calibration quality, demonstrating that offline policy-guided expert agents coordination is a practical path to deployable biomedical AI without retraining. Code is on GitHub\footnote\hrefhttps://github.com/HUANGLIZI/OPERAhttps://github.com/HUANGLIZI/OPERA.\blfootnoteCorresponding author: Zihan Li and Qingqi Hong. \\Li, Z and Liu, F have the equal contribution.

Abstract:
Benchmark accuracy is commonly used as a proxy for visual understanding in video large language models (LLMs). We test this assumption across twenty models spanning 2--78B parameters and ten architecture families. On MVBench, paired McNemar tests identify a model pair that differs in original-video accuracy (p_\textorig=0.0003) but not on a black-screen baseline (p_\textblack=0.53), showing that accuracy and visual dependency need not vary together. We define the Visual Dependency Gap (\VDG) as the per-question difference between correctness with the original video and with a black screen. The task-type \VDG ranking recurs across all twenty models, although its magnitude is not monotonic across model generations: Qwen3-VL has lower \VDG than Qwen2.5-VL despite stronger language priors. A four-condition ladder (black screen \to single frame \to shuffled frames \to original video) separates spatial, frame-diversity, and temporal contributions. Restoring temporal order adds at most +0.078 accuracy across sixteen open-weight models; frame diversity contributes more. An ablation from 0.5 to 24\,FPS does not support frame sparsity as the cause of low Temporal Reasoning \VDG. Under H.264 compression, flat aggregate accuracy masks bidirectional question flips, and CRF-sensitive questions are concentrated among visually dependent items. Four API-accessed models range from near-zero \VDG (0.025) to the second-highest value overall (0.315), indicating that the accuracy-dependency dissociation is not confined to one model family. Code and annotations are available at \urlhttps://github.com/JaeLee18/accuracy-without-grounding.

Abstract:
At its core, robotic manipulation is a problem of vision-to-geometry mapping (f(v) \rightarrow G). Physical actionssuch as reaching, grasping, and orientingare fundamentally defined by geometric properties like 3D positions, rotations, and spatial relationships. Consequently, we argue that the foundation for generalizable robotic control should be a vision-geometry backbone, rather than the widely adopted vision-language or video models. Conventional Vision-Language-Action (VLA) and video-predictive models rely on backbones pretrained on large-scale 2D image-text or temporal pixel data. While effective, their representations are largely shaped by semantic concepts or 2D priors, which do not intrinsically align with the precise 3D geometric nature required for physical manipulation. Driven by this insight, we propose the Vision-Geometry-Action (VGA) model, which directly conditions action generation on pretrained native 3D representations. Specifically, VGA replaces conventional language or video backbones with a pretrained 3D world model, establishing a seamless vision-to-geometry mapping that translates visual inputs directly into physical actions. To further enhance geometric consistency, we introduce Progressive Volumetric Modulation and jointly train action and 3D property prediction to preserve geometric representations. Extensive experiments validate the effectiveness of our approach. Across simulation benchmarks, VGA outperforms leading VLA, 3D-VLA, and WAM baselines, including \pi_0.5, OpenVLA-OFT, GeoVLA, and Motus, demonstrating precise spatial manipulation and robust generalization. In real-world deployments, VGA surpasses \pi_0.5 under unseen viewpoints and accurately follows language instructions for target grasping. These results highlight that operating on native 3D representationsrather than relying primarily on language or video priorsoffers a promising direction toward generalizable physical intelligence. Project page: \urlhttps://hcplab-sysu.github.io/VisionGeometryActionModel/.

Abstract:
Existing audio forgery detection and localization (AFDL) methods often overfit dataset-specific low-level artifacts, limiting their generalization to subtle, localized, and unseen manipulations. Recent audio large language model (ALLM)-based approaches cast AFDL as question answering but still model forensic evidence implicitly, without linking manipulation cues to predictions. To bridge this gap, we propose ThinkOmni, a reasoning-driven omni-modal large language model that jointly performs explicit forensic reasoning, spoofing detection, and temporal manipulation localization. To enable explicit reasoning supervision, we construct Forensic-Aware Chain-of-Thought (FACoT), a 100K-sample dataset with structured forensic evidence and reasoning annotations. Leveraging FACoT, we introduce Forensic-Aware Modality-Incremental Learning (FMIL), which progressively aligns semantic, acoustic, and spectral-visual representations with the LLM backbone to capture complementary forensic cues. We further propose Forensic-Consistent Multi-task Loss (FCML), which combines weighted cross-entropy with an adaptive localization loss to coordinate reasoning generation, spoofing detection, and temporal localization. Extensive experiments show that ThinkOmni achieves strong cross-dataset generalization in both detection and localization. Code, models, data, and inference examples are available at \urlhttps://beyond0814.github.io/ThinkOmni/.

Abstract:
Recently, large pretrained multimodal embedding models such as Qwen3-VL Embedding have shown strong promise for sequential recommendation, as they provide reusable semantic item representations across modalities and domains. However, directly using these embeddings often leads to suboptimal performance because of domain misalignment. Efficient side adaptation is therefore an attractive solution. Although adapting all backbone layers should help, existing side adapters often degrade with depth, prompting layer dropping despite the loss of useful hidden states. This is due to two major challenges: (1) the lack of modeling in selecting fused representations during residual addition, and (2) the insufficient preservation of earlier representations during progressive sigmoid fusion. This paper therefore asks a practical question: How can we design a side adaptation approach that effectively unlocks the potential of large pre-trained multimodal embedding models? To address this question, we propose \emphStresa, a stream-aware side-adaptation framework for frozen large pre-trained multimodal embedding models in sequential recommendation. Stresa introduces Stream-aware Hidden-Adapter Fusion (SHAF) to preserve historical side memory during fusion and Residual Stream Adapter (ReSA) to produce selective residual updates across layers. Empirically, Stresa consistently outperforms standard side adapters and state-of-the-art baselines on public datasets across multiple backbone embedding models. These results highlight the promise of adapting large embedding models for sequential recommendation. Our code is publicly available at~\urlhttps://github.com/GAIR-Lab/Stresa.

Abstract:
Multimodal Recommender Systems (MRSs) typically rely on a flawed ``modality harmony'' assumption, presuming that multimodal features are inherently beneficial and strictly aligned with users' collaborative interaction patterns. However, modality-topology conflicts are ubiquitous in real-world scenarios due to deceptive visual clickbaits and mismatched semantics. Blindly integrating these noisy modalities inevitably pollutes the pristine collaborative space, causing severe representation distortion. To address this, we propose Orthogonal purification and topology-guided MoE for conflict-aware multimodal Recommendation (OrthoRec). At its core, OrthoRec introduces Collaborative-Guided Orthogonal Purification (CGOP), which geometrically decouples multimodal features into directions parallel and orthogonal to a pure collaborative anchor. By adaptively truncating the orthogonal noise with an energy-preserving normalization, CGOP rectifies deceptive semantic directions while preserving the modality's intrinsic representation capacity. Furthermore, we design a Topology-Aware Routing Mixture-of-Experts (TAR-MoE). Guided by the collaborative topology, TAR-MoE employs decoupled sigmoid gating to break the zero-sum bottleneck of traditional softmax attention, autonomously determining the injection scale for each purified modality. Finally, a safe-SSL objective is introduced to dynamically penalize the forced contrastive alignment of contradictory pairs. Experiments on three real-world Amazon datasets show that OrthoRec consistently outperforms competitive recent baselines and exhibits improved robustness under modality noise and item sparsity. Our code is available at \urlhttps://github.com/Camilla-jl/Orthorec.

Abstract:
\newcommand\method\textscMarsTSC As a rich fusion of multiple types of media, time series encompass correlated and complementary representations across numeric, textual, and visual formats, and are pervasive across diverse real-world application domains. Harnessing recent advances in large vision-language models (VLMs) offers a promising direction for exploiting the rich intrinsic temporal characteristics embedded across these complementary media for few-shot time series classification. However, existing approaches rely on static context that fail to automatically evolve with the few-shot training samples, and lack tailored agentic reasoning mechanisms to maximally squeeze out discriminative knowledge from the limited labeled data. In this paper, we propose the first VL\underlineM \underlineagentic \underlinereasoning framework for few-\underlineshot multimodal \underlineTime \underlineSeries \underlineClassification (\method), which introduces a self-evolving knowledge bank as a dynamic context iteratively refined via reflective agentic reasoning. The framework comprises three collaborative roles: i) Generator conducts reliable classification via reasoning; ii) Reflector diagnoses the root causes of reasoning errors to yield discriminative insights targeting the temporal features overlooked by Generator; iii) Modifier applies verified updates to the knowledge bank to prevent context collapse. We further introduce a test-time update strategy to enable cautious, continuous knowledge bank refinement to mitigate few-shot bias and distribution shift. Extensive experiments across 12 mainstream time series benchmark datasets demonstrate that \method delivers substantial and consistent performance gains across 5 VLM backbones, outperforming both classical and foundation model-based time series baselines under few-shot conditions, while producing interpretable rationales that ground each classification decision in human-readable feature evidence. Code is available at \repourlhttps://github.com/HuangJW0821/MarsTSC.

Abstract:
Urban socio-economic sensing plays a vital role in advancing global sustainable development goals. With the advent of Large Vision-Language Models (LVLMs), new opportunities have emerged to address this challenge by framing it as a multi-modal perception and reasoning task. However, recent studies show that LVLMs still struggle to make accurate and interpretable socio-economic predictions from visual data. To overcome these limitations and fully exploit the potential of LVLMs, we propose CityRiSE, a novel framework for Reasoning urban Socio-Economic status in LVLMs via reinforcement learning (RL). With carefully curated multi-modal dataset and verifiable reward design, our approach guides the LVLM to focus on semantically meaningful visual cues, enabling structured and goal-oriented reasoning for generalist socio-economic status prediction. Experiments demonstrate that CityRiSE, equipped with emergent reasoning, significantly outperforms existing baselines, improving both prediction accuracy and generalization across diverse urban contexts, especially on unseen cities and unseen indicators. This work highlights the promise of combining RL and LVLMs for interpretable and generalist urban socio-economic sensing.

Abstract:
Medical vision-language models (VLMs) hold significant promise for clinical image understanding, but scaling them at training time is limited by data privacy constraints and the cost and scarcity of expert annotations. Test-time scaling (TTS) is therefore a practical alternative, as it can improve performance without retraining. However, existing TTS methods are designed mainly for LLMs, where extra inference compute is spent on language-space reasoning. This paradigm does not transfer directly to medical visual reasoning, where generating more text does not yield new visual findings. Instead, progress requires iteratively seeking and integrating fine-grained, localized visual evidence from different image regions. We introduce MedVES, a training-free framework for visual evidence scaling at test time for frozen VLMs in medical VQA. MedVES casts test-time scaling as iterative visual evidence seeking rather than extended language-space reasoning. Its core component is a Visual Overlay State (VOS), a bounded visual state representation that externalizes accumulated evidence on the original image while preserving global context. Built on VOS, MedVES performs iterative visual evidence acquisition with adaptive stopping, enabling test-time compute to scale with question complexity. Across five medical VQA benchmarks covering diverse imaging modalities, a 4B MedVES model achieves strong average performance at its scale and is competitive with, or better than, several models with 8× more parameters, without any additional training. \hrefhttps://github.com/wxzwxzwxz/MedVESCode is available.

Abstract:
Multimodal recipe generation suffers from severe hallucinations. When visual evidence is ambiguous, Multimodal Large Language Models (MLLMs) often fall back on deep-rooted dataset biases to fabricate unverified ingredients. Even conventional Retrieval-Augmented Generation (RAG) struggles, as models blindly copy retrieved noise. To bridge this visual-culinary gap, we first introduce C2MR (Chinese Cross-Modal Recipes) dataset, a high-fidelity dataset of 32,072 expert-vetted image-recipe pairs. Building upon this, we propose the \textscRecipe (REtrieval-constrained Culinary Inference via Probabilistic Entropy) framework, shifting generation from probabilistic guessing to verifiable logical reasoning. Instead of passively accepting retrieved priors, \textscRecipe actively quantifies its own epistemic uncertainty (entropy) via stochastic sampling. By fusing this internal hesitation with statistical frequency, we establish a non-linear decision boundary that effectively excises bias-driven ``stubborn hallucinations''. Finally, a dynamic logits mask enforces these verified ingredients as hard constraints during synthesis. Extensive experiments demonstrate \textscRecipe achieves state-of-the-art performance on Recipe1M and C2MR. It slashes hallucination rates to 4.8% with >93.5% Precision. Crucially, our framework acts as a universal catalyst, enabling 2B-scale edge models to rival 13B baselines and pushing 72B flagship models toward stronger ingredient-hallucination control. Our code is available at \hrefhttps://github.com/Agentyzu/Recipehttps://github.com/Agentyzu/Recipe.

Abstract:
The difficulty of predicting highly articulated and long-horizon human motion stems not only from limited forecasting capacity, but more fundamentally from how motion representations are learned. Optimizing future prediction through a single direct regression objective biases the model toward over-smoothed continuations, rather than learning representations that capture the joint dynamics and temporal dependencies governing future motion. We present MoReFun, a two-stage framework that first learns future-aware motion representations and then adapts them for downstream forecasting. During pretraining, the model reconstructs masked future joints under full past-motion guidance, while jointly reconstructing masked past motion. We further introduce velocity-aware masking, which preferentially masks highly dynamic joints so that the model focuses on the most informative and hardest-to-predict motion regions rather than low-motion structure. During fine-tuning, the pretrained model predicts the entire future sequence, now treated as fully masked. Experiments on Human3.6M, AMASS, and 3DPW show that MoReFun reduces MPJPE by an average of 8.8% over strong baselines, with especially clear gains on complex motions. Moreover, on the FineMotion dataset, we further equip the model with a lightweight text head to enable future motion understanding alongside prediction. Both prediction and understanding benefit from the learned representation and further complement each other by aligning low-level motion dynamics with high-level action semantics during fine-tuning. Code is publicly available at \urlhttps://github.com/SorenShi/MoReFun.

Abstract:
Dyadic face-to-face interaction is a fundamental form of human communication, where facial motions are continuously shaped by mutual engagement between two interlocutors rather than being produced independently. In many real-world applications, however, direct visual access to the conversation partner is often restricted or unavailable due to strict privacy requirements, making privacy-preserving audio-only generation especially desirable. Existing methods confront three main limitations: (1) single-speaker 3D talking head approaches generate isolated facial animations, failing to model mutual interactive dependencies; (2) current dyadic interaction frameworks predominantly rely on the partner's visual input as conditioning, rendering them difficult to apply in privacy-sensitive scenarios; and (3) asymmetric speaker-listener paradigms fall short of handling the dynamic turn-taking behavior in genuine bidirectional conversations. To tackle these challenges, we formally define the pure audio-driven dyadic 3D talking head generation task and propose DyadTalk, a unified framework requiring no visual cues during inference. The core of DyadTalk is the Hybrid Interaction Module (HIM) featuring a Mix-then-Decouple design. It effectively mixes dual speech streams to capture shared conversational dynamics while decoupling the representations to preserve speaker-specific motion patterns. Extensive experiments validate that DyadTalk consistently outperforms the audio-adapted baseline, DualTalk. Specifically, it achieves a 43% reduction in expression Frchet Distance (FD-EXP) and a 35% decrease in Mean Squared Error (MSE-EXP) on the in-domain test split. Furthermore, DyadTalk demonstrates strong out-of-domain robustness, reducing generalization errors by over 23% across key perceptual metrics. By successfully recovering realistic and coherent bidirectional motions solely from synchronized audio, our work provides a viable solution for privacy-friendly conversational avatar generation. See the supplementary video for additional results: https://slnrm.github.io/DyadTalk/

Abstract:
Few-shot segmentation aims to segment novel object categories given only a handful of annotated examples. Existing approaches either rely on meta-trained specialist models whose learned matching functions generalize poorly across domains or adopt training-free pipelines based on point-wise feature similarity that treat each patch embedding independently and fail to capture the global statistical structure of the target concept. In this paper, we present SegmentGDA, a fully training-free framework that recasts few-shot segmentation as probabilistic inference in the representation space of frozen vision foundation models. SegmentGDA models foreground and background support features as class-conditional Gaussian distributions and applies Gaussian Discriminant Analysis to classify each query patch via Bayesian decision rules, producing a dense posterior probability map that aggregates global class-level statistics across all support samples and suppresses noisy correspondences more effectively than deterministic local matching. The posterior map is further converted into spatially diverse positive and negative point prompts for the Segment Anything Model through an iterative refinement loop, yielding high-resolution segmentation masks without any task-specific training or fine-tuning. Extensive experiments across four standard benchmarks demonstrate the effectiveness of the proposed framework, underscoring its potential for real-world applications. These results highlight distribution-based modeling as a promising approach for adapting vision foundation models to few-shot dense prediction. Code is available at \urlhttps://github.com/wavachao/SegmentGDA.

Abstract:
Prior solutions and benchmarks for medical hallucination evaluation remain insufficient for systematically modeling diverse fine-grained clinical hallucination types and fail to capture the length-dependent evolution of hallucinations in extended clinical reports. To bridge these gaps, we introduce MGHalluDet\footnoteThe data source is available at: \textcolorbluehttps://github.com/halhrha/MGHalluDet, a multi-grained benchmark specifically designed to evaluate hallucination behaviors across different narrative scales and diverse clinical hallucination types. It comprises short-, medium-, and long-context subsets, while systematically covering fine-grained hallucination categories including object, attribute, relation, and imaging errors. Concurrently, we propose TRACE (Temporal Routing of Autoregressive Context and Entropy), a novel dynamic token-level hallucination detection framework tailored to autoregressive medical text generation. Unlike conventional approaches that rely on static similarity matching, TRACE explicitly tracks the evolution of internal representation states throughout the decoding process, enabling more effective detection of subtle hallucination onset and transition patterns. Specifically, our method leverages Spatial Attention Entropy (SAE) to quantify the visual dispersion associated with unbounded lesions, while incorporating temporal context to detect abrupt shifts in autoregressive generation dynamics that are indicative of hallucinations. Extensive experiments across different context lengths and hallucination settings consistently demonstrate the effectiveness of TRACE for medical LVLM hallucination detection.

Abstract:
The social interactions among crowds via Danmaku (a.k.a., bullet comments) on modern multimedia platforms can facilitate both viewpoint conflicts and consensus, providing fine-grained discriminative social signals that can benefit fake news detection. However, the inherent accumulation latency of Danmaku in real-world scenarios violates the real-time necessity of fake news detection, making the studies of Danmaku-related fake news detection underexplored. To break this violation, we simulate this temporal-aware user interactive process by proposing a novel temporal Generative danmaku framework, called Genda, which consists of: (1) a Danmaku Trigger for predicting the timing and intensity of user reactions; and (2) a Danmaku Generator for synthesizing corresponding semantic and emotional expressions, thereby mutually constructing a temporally aligned and human-like pseudo Danmaku streams. To make the generated Danmaku useful for identifying fake news videos, we further design a Danmaku-guided Temporal Multimodal fake news detection model - DM-FEND, which enables fine-grained multimodal interactions among video, audio, text, and Danmaku, enhancing dynamic modalities alignment and semantic noise inhibition. The experimental results demonstrate that \emphDM-FEND consistently outperforms state-of-the-art baselines across both Chinese (FakeSV) and English (FakeTT) benchmarks. Further ablations validate the crucial role of temporal Danmaku modeling in enhancing robustness and discriminative capability. Finally, this study offers a bright and robust solution for multimodal fake news detection in modern social interactive fashions by bridging the temporal inconsistency between news and user behaviors. To ensure reproducibility, the code and data used in this study are released at: \urlhttps://github.com/126541/Let-the-Bullets-Fly.

Abstract:
The emergence of medical deepfakes, i.e., medical images manipulated by deep generative models, poses a significant threat to clinical workflows. However, existing detectors suffer from two critical limitations: poor generalization to unseen generative architectures for manipulation detection and lack of interpretability. In this context, we present HexMIL (Hierarchical EXplainable Multiple Instance Learning), a mask-free medical deepfake detector that simultaneously addresses both limitations using only binary volume-level supervision. HexMIL decomposes each CT volume into a two-level hierarchy of patches and slices, aggregated via independent Gated Attention modules whose weights are directly combined into a full-resolution 3D attention volume that localizes the manipulated sub-region without any pixel-level annotation. Unlike post-hoc methods such as Grad-CAM, HexMIL's attention weights constitute the exact forward computation driving the classification decision, providing ante-hoc and structurally faithful spatial attribution. We evaluate HexMIL on M3DSynth and CT-GAN datasets under a rigorous cross-generator generalization protocol, training on a single generative architecture and testing on unseen ones. HexMIL outperforms all baselines by +9.1 AUC and +9.4 F1 in out-of-domain classification, and achieves the best average IoU and Pointing Game score in localization. Project page: \hrefhttps://opontorno.github.io/hexmil/opontorno.github.io/hexmil.

Abstract:
\beginabstract Although Multimodal Large Language Models (MLLMs) have advanced rapidly in video understanding, their ability to produce video aesthetic captions remains poorly evaluated: existing metrics (e.g., BLEU, ROUGE) penalize linguistic diversity and cannot verify whether captions conform to established aesthetic norms. To bridge this gap, we present VAC-Bench, the first benchmark for systematic video aesthetic captioning evaluation, comprising 1,800 expert-annotated videos across diverse visual domains with fine-grained aesthetic attribute tags, quality ratings, and hierarchical captioning tasks. We further propose VANE (Video Aesthetic Norm Evaluation), a protocol that shifts assessment from reference text matching to norm conformance verification, introducing a three-tier classification that disentangles factual hallucinations from normative deviations and yielding three metrics: Norm Conformance, Coverage, and Completeness. Evaluation of 11 state-of-the-art MLLMs reveals a striking perception-expression gap: high dimensional coverage but substantially lower norm conformance, indicating that current models perceive which aesthetic dimensions to address yet struggle to articulate norm-conformant judgments. The code are available at \urlhttps://github.com/flyingbird1993/VAC-Bench. \endabstract

Abstract:
The scarcity and privacy concerns of anomalous videos in real-world scenarios jointly lead to severe shortages in training data, forming a critical bottleneck for the advancement of weakly-supervised video anomaly detection (WSVAD). Recent studies utilize video generation techniques to synthesize anomalous videos, thereby mitigating data scarcity. However, these generated videos often have significant discrepancies from the real videos, creating a pressing need to enhance the quality and controllability of the generated videos. To address these limitations, we propose a novel Customized Anomalous Video Generation Engine (CAVGE). CAVGE synthesizes realistic anomalous videos through a multi-stage pipeline: (1) LLM-driven anomaly script generation decomposes abnormal behaviors into structured scenes; (2) Text-guided keyframe synthesis as a critical visual prompt; (3) Generating anomalous videos guided by scripts and keyframes. This method not only supports the incorporation of multi-stage human guidance, but also simultaneously outputs corresponding multi-granularity annotations and detailed text descriptions for the generated videos. Furthermore, we attempt to leverage the generated videos to endow the model with the capability of effectively identifying novel anomalies, thereby breaking through the generalization limitations of previous methods. Thus, we propose a Vision-Language Video Anomaly Detection framework with Incremental Learning (VLVAD-IL). This framework integrates multimodal vision-language representations with incremental learning mechanisms, enabling continuous assimilation of new anomalous knowledge, while mitigating catastrophic forgetting of previously acquired knowledge. Extensive experiments conducted on widely used benchmarks conclusively validate the effectiveness of our proposed generation engine and framework. The source code is publicly available at \hrefhttps://github.com/WeiliangHuang-UM-connect/Customized-Anomalous-Video-Generation-for-Incremental-Learning-in-WSVAD \textcolorblack\nolinkurlhttps://github.com/WeiliangHuang-UM-connect/Customized-Anomalous-Video-Generation-for-Incremental-Learning-in-WSVAD.

Abstract:
Recent advances in video diffusion models have substantially enhanced character animation techniques. However, existing methods primarily depend on structural conditions, such as DWPose or SMPL-X, to animate character images, which limits their effectiveness in open-domain scenarios involving dynamic backgrounds or complex character-scene interactions. This study presents AniCrafter, a diffusion-based human-centric animation model designed to seamlessly integrate and animate a given character within open-domain dynamic backgrounds while adhering to specified human motion sequences. Built upon advanced Image-to-Video (I2V) diffusion architectures, the model introduces an innovative ``avatar-background'' conditioning mechanism that reformulates open-domain human-centric animation as a restoration problem, thereby achieving versatile, occlusion-aware animation results. Experimental evaluations demonstrate that the proposed approach outperforms current state-of-the-art methods and exhibits an exceptional capability in handling challenging scenarios. Homepage: \urlhttps://myniuuu.github.io/AniCrafter

Abstract:
Teacher--Student (T--S) distillation is promising for 3D point cloud anomaly detection, but existing frameworks suffer from a \emphpassive discrepancy bottleneck. Task-agnostic teachers provide limited separation between normal and anomalous features, allowing students to over-generalize to defects and weakening the anomaly signal. We propose the Semantic-Geometric Distillation Network (SGD-Net), an active, dual-sided discrepancy amplification framework. On the teacher side, language-driven cross-modal alignment uses textual descriptions as semantic anchors to learn anomaly-aware representations. On the student side, a Geometric Perception Branch with curvature-guided anisotropic attention captures fine-grained structural irregularities. At inference, their semantic and geometric scores are fused using the Cauchy combination test. Experiments on Real3D-AD and Anomaly-ShapeNet demonstrate state-of-the-art object-level detection and point-level localization. Code will be available at \urlhttps://github.com/hanlinhao156-ops/SGD-Net.

Abstract:
While traditional Image Aesthetics Assessment (IAA) relies heavily on 2D images, 3D Object Aesthetics Assessment (3D-OAA) remains underexplored due to challenges in modeling spatial geometry and reconciling contradictory multi-view aesthetic perceptions. To bridge this gap, we formulate a multi-view, 3D-aware assessment paradigm that is compatible with existing 2D IAA frameworks and comprises multi-view information embedding and adaptive view-weight allocation. We instantiate and evaluate this paradigm in Hairstyle Aesthetics Assessment (HAA), a texture-sensitive domain in which appearance varies across viewpoints, by introducing the Hairstyle Aesthetics Assessment Network (HAANet). On the one hand, HAANet evaluates hairstyle aesthetics by embedding multi-view portrait features and dynamically weighting their contributions. On the other hand, HAANet employs a two-stage multi-task learning strategy to extract specialized features and utilizes symmetric attention to capture aesthetic compatibility. Furthermore, we construct HAA10K, a comprehensive dataset of over 10,000 expert-annotated portrait images. On HAA10K, HAANet outperforms nine evaluated methods across all metrics, providing an initial benchmark for multi-view HAA and evidence for its use as an aesthetic signal in the evaluated downstream tasks. All resources will be available at https://github.com/woshidandan/ACE-Hairstyle-Aesthetics-Assessment.

Abstract:
Despite numerous Graphical User Interface (GUI) agents claiming to automate user interaction tasks, to date, few achieve satisfactory interaction capability with human users in real-world scenarios. Through empirical analysis, this paper identifies the root cause of the limited interaction capability as the rigid swipe execution. In particular, unlike humans, who perform swipes with fine-grained control over trajectory, speed, and timing, existing agents can only conduct simplistic, deterministic swipe behaviors, leading to frequent failures on complicated user-like interaction tasks. Due to the lack of open-source human-like swipe training data, we propose SwipeGen, the first tool for synthesizing diverse and human-like swipe interactions, and SwipeBench, the first benchmark for evaluating agents' swipe interaction quality. Extensive experiments show that SwipeGen can improve the swipe execution success rate of existing agents by up to 2.46x. Our code, dataset, and model are available at https://github.com/TSKGHS17/SwipeGen.

Abstract:
The rapid advancement of generative AI has enabled highly realistic image manipulations, posing challenges to trustworthy media analysis. Existing multimodal forgery detection methods often rely on single-stage fusion and one-shot generation, failing to establish reliable cross-modal grounding and leading to inconsistent explanations. We propose TraceSleuth, a coarse-to-fine, evidence-driven framework that explicitly models the formation and verification of forensic evidence. Instead of direct prediction, it first constructs a distribution-aware hypothesis via prototype-based aggregation, and then enforces consistency-aware reasoning through cross-modal alignment for faithful explanation generation. For localization, a graph-based propagation module captures non-local trace consistency and is fused with semantic cues for precise mask prediction. Counterfactual and consistency regularization further ensure that both predictions and explanations are grounded in genuine trace regions. Experiments demonstrate improved detection and localization performance, along with more faithful and interpretable explanations. The code is available at \hrefhttps://github.com/GCLion/TraceSleuthhttps://github.com/GCLion/TraceSleuth.

Abstract:
Maritime perception is a critical component of trade security, with multi-modal semantic segmentation serving as a foundational technology for fine-grained scene understanding. However, achieving accurate and robust segmentation performance remains severely constrained by heterogeneous composite degradation and foreground-background imbalance. Existing cascaded pipelines attempt to restore images before the downstream task, but their emphasis on visual fidelity is misaligned with machine perception and can amplify background interference. To address these challenges, we propose Bio-PCNet, a biologically-inspired framework for multi-modal maritime segmentation under composite degradation. It introduces a feedback-inspired joint training framework for task-oriented representation learning. Bio-PCNet establishes a training-time closed loop comprising of three core components: 1) Synaptic Gain-Control Fusion (SGCF) module, which employs a denoise-then-attend strategy to suppress interference before global context integration; 2) Semantic-Guided Predictive Feedback Learning (S-PFL) that injects top-down semantic priors during training to selectively recover target details while inhibiting background distractions; and 3) Self-Supervised Magno-Parvo Frequency Loss (MP-Loss), leveraging frequency-spatial decoupled supervision as a task-specific inductive bias for auxiliary fusion learning. Extensive experiments on IVMSD and additional public benchmarks show consistently stronger performance and improved cross-dataset generalization. The source code is publicly available at https://github.com/GPIOX/Bio-PCNet.

Abstract:
Audio-visual generalized zero-shot learning leverages the synergy between auditory and visual modalities to recognize instances from seen and unseen categories. However, existing methods usually rely on a rigid alignment paradigm, mapping complex multi-modal features to singular class category. This leads to semantic information loss and misalignment. To break this bottleneck, we propose a novel Textual Cloud strategy. It elaborates singular category into a multi-faceted semantic landscape, including visual, auditory, and contextual cues. These facets represent physical appearance, acoustic signatures, and abstract definitions, respectively, forming a set of orthogonal and distinct semantic subspaces. Furthermore, to orchestrate these disentangled semantic cues, we introduce a Gated Global Attention (GGA) module. GGA dynamically recalibrates the importance of each aspect, preventing over-reliance on any single facet. It simultaneously allows each textual cloud aspect to be enriched by the others, fusing comprehensive semantics to guide the zero-shot recognition. Extensive experiments on VGGSound, UCF-101, and ActivityNet benchmarks demonstrate that our method establishes new state-of-the-art performance, achieving improvements in Harmonic Mean metric by margins of 4.17% and 5.22% on VGGSound and ActivityNet, respectively. Our source code and supplementary material are available at: \urlhttps://github.com/Asia-Leee/Textual-Cloud.

Authors: Shibo Ma (Shandong Artificial Intelligence Institute, Qilu University of Technology (Shandong Academy of Sciences)), Chunjie Ma (Shandong Artificial Intelligence Institute, Qilu University of Technology (Shandong Academy of Sciences)), Shengxun Wei (School of Computer Science and Engineering, Tianjin University of Technology), Nian Hu (School of Computer Science and Engineering, Tianjin University of Technology), Xinhui Li (School of Computer Science and Engineering, Tianjin University of Technology), Yibo Zhao (School of Computer Science and Engineering, Tianjin University of Technology)

Abstract:
Prohibited object detection in X-ray images suffers from limited annotated data, which hinders the training of robust detectors. Moreover, the cumulative attenuation and severe object overlap in X-ray images make data generation challenging in terms of realism, diversity, and controllability. In this paper, we propose PhysDiff, a physics-inspired diffusion framework that explicitly incorporates X-ray image priors into controllable generation through semantic enrichment, realistic generation, and physics-inspired control. First, LLM-guided dense captioning alleviates the lack of rich text annotations in prohibited object detection datasets by augmenting sparse labels with material, pose, and scene-context descriptions. Second, Log-Domain Mixup is designed to improve the distributional fidelity and coverage of generated samples via Beer-Lambert-law-inspired interpolation in approximate optical-density space. Finally, a physics-inspired dual-condition ControlNet is designed to adaptively fuse density masks and edge cues to achieve more accurate and stable controllable generation by jointly modeling attenuation patterns and structural boundaries. Extensive experiments demonstrate that PhysDiff improves generation fidelity and controllability over representative baselines, while the resulting synthetic data further benefits downstream prohibited object detection, especially for small objects. The source code is publicly available at https://github.com/VPMS-Open-Source/PhysDiff.

Abstract:
Most existing video compression algorithms follow a paradigm of transformation and quantization, optimizing the trade-off between distortion and bitrate. However, extremely low-bitrate compression remains an underexplored frontier where perceptual quality optimization under severely constrained coding resources has not been adequately addressed. In this paper, we propose a unified generative framework that leverages pre-trained Diffusion Transformer (DiT) priors to achieve high perceptual quality at extremely low bitrates. We first introduce a flexible Group-of-Latents (GoL) strategy within the latent space of a causal tokenizer, explicitly partitioning the latent stream into intra I-latents and inter P-latents. The Deep Compression Module (I-DCM) then encodes key I-latents to preserve perceptual anchors with minimal overhead. Building upon these anchors, the DiT-based Unified Latent Denoising Module (U-LDM) refines intra-frame textures and synthesizes P-latents from noise, reconstructing temporal dynamics at zero additional bitrate cost. Extensive experiments demonstrate that our method uniquely operates in the extreme-low bitrate regime (e.g., < 0.005 bpp), achieving state-of-the-art perceptual fidelity with rich spatial details and robust temporal consistency. Code will be available at https://github.com/sk-pku/Group-of-Latents.

Abstract:
This paper investigates how to leverage the powerful capabilities of the Segment Anything Model 2 (SAM2) to tackle few-shot segmentation (FSS) tasks. One major challenge lies in the substantial appearance differences between the support and query images, which may diminish the effectiveness of the memory-matching process in SAM2. Through systematic analysis, we find that the memory generated based on the categorical semantic information is robust to this appearance divergence and is more compatible with SAM2. Based on these observations, we propose a new Memory-Enhanced SAM2 (MESAM) method, which first introduces a Semantic-Visual Memory Generation (SVMG) module to separately construct a query semantic memory, a query foreground visual memory, and a support adaptive visual memory. Then, in our designed Interactive Semantic Memory Refinement (ISMR) module, we further enrich the obtained semantic memory with fine-grained visual details, and suppress its misatcivated background responses. Extensive experiments on the PASCAL-5^i, COCO-20^i, and LVIS-92^i datasets demonstrate that MESAM consistently outperforms existing FSS methods, achieving new state-of-the-art segmentation performance. Our source code is available at: \hrefhttps://github.com/tianyu-zou/MESAMhttps://github.com/tianyu-zou/MESAM.

Abstract:
Vision-language-action (VLA) models have shown strong performance in robotic manipulation, yet their robustness to physically realizable adversarial attacks remains underexplored. Existing studies reveal vulnerabilities through language perturbations and 2D visual attacks, but these attack surfaces are either less representative of real deployment or limited in physical realism. In contrast, adversarial 3D textures pose a more physically plausible and damaging threat, as they are naturally attached to manipulated objects and are easier to deploy in physical environments. Bringing adversarial 3D textures to VLA systems is nevertheless nontrivial. A central obstacle is that standard 3D simulators do not provide a differentiable optimization path from the VLA objective function back to object appearance, making it difficult to optimize through an end-to-end manner. To address this, we introduce Foreground-Background Decoupling (FBD), which enables differentiable texture optimization through dual-renderer alignment while preserving the original simulation environment. To further ensure that the attack remains effective across long-horizon and diverse viewpoints in the physical world, we propose Trajectory-Aware Adversarial Optimization (TAAO), which prioritizes behaviorally critical frames and stabilizes optimization with a vertex-based parameterization. Built on these designs, we present Tex3D, the first framework for end-to-end optimization of 3D adversarial textures directly within the VLA simulation environment. Experiments in both simulation and real-robot settings show that Tex3D significantly degrades VLA performance across multiple manipulation tasks, achieving task failure rates of up to 96.7%. Our empirical results expose critical vulnerabilities of VLA systems to physically grounded 3D adversarial attacks and highlight the need for robustness-aware training. Code is available at https://vla-attack.github.io/tex3d.

Abstract:
Low-light imaging often introduces color bias caused by the low signal-to-noise ratio and the image formation process. Although recent low-light image enhancement methods have achieved strong brightness recovery, faithful color restoration remains challenging, manifesting as overall color bias together with local under- and over-saturation. To address this issue, we propose CAGE, a cylindrical color correction framework with adaptive color debiasing and gamut-harmonized saturation rectification for color-faithful low-light image enhancement. We first introduce AdaLAB, a cylindrical adaptive LAB color space that provides a decoupled and image-specific basis for uniform color correction. Building on this color space, we further develop AdaCCT, an adaptive cylindrical color transform with forward and inverse transforms for the conversion between RGB and AdaLAB color space, as well as necessary color debiasing and saturation rectification. The forward transform suppresses embedded color bias before backbone enhancement by reorganizing the chromatic distribution through chromatic-plane shifting and scaling, while the inverse transform achieves faithful saturation rectification through out-of-gamut lightness compensation. Extensive experiments on multiple benchmarks show that CAGE achieves more faithful color restoration, specifically reduces color bias and saturation abnormality, and delivers better overall visual quality across different low-light enhancement backbones. The code is available at \urlhttps://yangzhichen763.github.io/CAGE/.

Abstract:
Blind RAW video demosaicing and denoising remains challenging because RAW imaging is camera dependent, and collecting large-scale paired RAW video datasets is prohibitively expensive. In this paper, we present a scalable synthetic-only framework for blind RAW video demosaicing and denoising. Instead of relying on paired RAW video supervision, our framework learns from physics-based synthetic RAW videos whose inverse imaging process and degradations preserve both camera realism and temporal coherence. Specifically, we develop a RAW video synthesis pipeline that converts clean sRGB videos into degraded Bayer RAW sequences using real camera parameters, along with temporally shared sensor noise. On the model side, we propose STFNet, a flow-free causal demosaicing and denoising framework that combines Stable Temporal Propagation and Historical Temporal Aggregation to improve temporal reliability under online inference. Extensive experiments suggest that, under current benchmarks, synthetic-only training transfers competitively to real RAW videos while achieving a favorable balance between restoration quality and computational efficiency. In addition, we introduce Ultra-RAW, a synthetic benchmark with exact paired ground truth for full-reference evaluation, while using real benchmarks as transfer testbeds for cross-camera assessment. Code is available at \hrefhttps://github.com/nsqg/STFNet https://github.com/nsqg/STFNet.

Abstract:
Cross-modal translation between satellite imagery observations and radar reflectivity is a fundamental yet challenging problem in weather nowcasting. Existing approaches typically learn direct pixel-to-pixel mappings, which are prone to instability, over-smoothing, and poor preservation of rare extreme events due to long-tail intensity distributions and spatial misalignment across modalities. In this work, we reformulate cross-modal translation as two structured subproblems: (i) modality-specific sparse representation learning, and (ii) representation-to-representation mapping across modalities through sparse coding. We pretrain two modality-specific branches on IR-Lightning and Radar using SSC-VAE, respectively, to learn sparse dictionaries and latent representations capable of stably reconstructing observations via sparse compositions. With reconstruction capability fixed, we then train a decoupled translation module that maps satellite sparse codes to Radar sparse codes. Translation is further guided by continuous intensity-aware reconstruction, multi-level structural and semantic alignment constraints. By restricting cross-modal learning to a structured intermediate sparse space, the proposed framework achieves both accurate modality reconstruction and more robust satellite-to-radar translation. Extensive experiments on the SEVIR and Kendari datasets demonstrate consistent improvements over existing methods, including GAN-based approaches, particularly under challenging conditions of satelliteradar misalignment. These results highlight the effectiveness of structured sparse-space translation for cross-modal weather modeling. Our code is available at https://github.com/m5379mm/DDCT-S2R.

Abstract:
Trajectory prediction has shifted toward structured formulations with explicit social modeling. However, existing methods inadequately distinguish the functional roles of social influence in trajectory planning. Observing that agents typically form motion plans by anticipating others' future behaviors before making local reactive adjustments, we identify social interactions as playing staged roles, namely planning precedes reaction. We propose INTraJ, a unified framework that decomposes social influence into two stages: a planning stage constructs reference trajectories using future social information, and a reaction stage recovers local adjustments from the residual between full-context prediction and the reference. INTraJ supports both multi-target and single-target paradigms. Extensive experiments on four standard benchmarks, including Argoverse 2, Argoverse 2-ped, ETH/UCY, and SDD, demonstrate consistent improvements, particularly in FDE and long-horizon consistency, with state-of-the-art performance achieved in several settings. INTraJ reframes trajectory prediction as a planning-driven two-stage process, validating that staged social modeling is critical for stable predictions. The code is publicly available at https://github.com/11isnotavailable/INTraJ.

Abstract:
The automated generation of high-fidelity, city-scale 3D environments remains a formidable challenge with profound academic and industrial implications. However, existing methods struggle to achieve the necessary quality, fidelity, and scalability. To address this, we propose UrbanWorld2.0, a reality-aligned intelligent multimodal synthesis engine that creates detailed, city-scale 3D worlds of high fidelity. We introduce an agentic framework that leverages diverse multimodal foundation tools to acquire real-world knowledge, maintain robust intermediate representations, and construct complex 3D scenes. This agentic design, featuring dynamic data processing, iterative self-reflection and refinement, and the invocation of advanced multimodal tools, minimizes cumulative errors and enhances overall performance. Extensive quantitative experiments and qualitative analyzes validate the superior performance of UrbanWorld2.0 in real-world alignment, shape precision, texture fidelity, and aesthetics level, achieving a win rate of over 86% against existing baselines for overall perceptual quality. This combination of 3D quality, reality alignment, scalability, and seamless compatibility with computer graphics pipelines makes UrbanWorld2.0 a promising foundation for applications in immersive media, embodied intelligence, and world models.

Abstract:
Infrared and visible image fusion (IVIF) plays a critical role in enhancing scene perception by integrating information from different imaging modalities. Existing knowledge distillation-based IVIF methods generally follow a teacher-imitation paradigm, which inherently limits the students performance to that of the teacher, not to mention whether the student can fully inherit the teachers capability. In this paper, inspired by the fact that human students can possibly surpass their teachers during learning, whose key lies in enabling students simultaneously to engage in self-learning and exploration in more challenging environments under teacher guidance, we for the first time introduce the idea of student self-learning into the IVIF knowledge distillation framework. To address the challenge of coordinating teacher guidance with student self-learning, we propose a reinforcement learning (RL) agent-based approach that jointly regulates the intensity of teacher supervision and the difficulty of self-learning. Specifically, training samples with varying levels of difficulty are automatically generated to enable adaptive self-learning, and an RL agent is designed to dynamically adjust the learning strategies based on the students current learning status, which is characterized by the students performance and student-teacher performance gap at each time. Through the reinforcement learning with well-defined state space and reward functions, the agent can automatically uncover the subtle learning dynamics that enables the student to well inherit the fusion knowledge while maximally surpassing the teacher. Extensive experiments demonstrate that the proposed approach is able to achieve state-of-the-art fusion performance with a quite lightweight model, yielding significant improvements over existing methods in both fusion quality and inference speed. The code is available at https://github.com/wyhlaowang/AgentFuse.

Abstract:
Text-guided image fusion aims to provide flexible and interactive control that accommodates diverse user preferences and task-specific requirements. However, existing approaches still suffer from insufficient instance-level controllability, leading to inaccurate target enhancement or unintended spread to irrelevant regions, thereby reducing the precision and effectiveness of text-guided fusion. To address this limitation, we propose a precision-oriented parallel guidance framework for infrared and visible image fusion, termed CRIF. Firstly, a parallel collaborative architecture composed of an instance localization branch (ILB) and a bi-modal integration branch (BIB) is established to achieve fine-grained spatial alignment between textual intent and image content. The ILB employs the Semantic Gating Module to resolve referential ambiguity between textual descriptions and source images, yielding text-aligned modulation masks for precise target localization. Meanwhile, the BIB constructs a global fusion representation through the Cross-Modal Aggregation Module. Subsequently, the global fusion representation is modulated by the region-aware decomposer, enabling locally controllable fusion in target regions while preserving overall visual coherence. Extensive experiments demonstrate that CRIF achieves state-of-the-art performance on both base image fusion and infrared-visible referring image segmentation, while also enabling precise instance-level control of designated targets. The code is available at https://github.com/lxq-jnu/CRIF.

Abstract:
Reward models guide text-to-image (T2I) systems toward outputs aligned with human preferences. However, typical reward models such as HPSv3 are trained on pre-annotated data from earlier T2I models, without accounting for quality discriminative shifts arising from evolving model capabilities and reinforcement learning (RL) iterations, limiting their broader applicability. In this work, we propose HPSv3++, a reward model framework that elevates the HPSv3 model for varying T2I model capabilities and their RL iteration changes across the full capability-iteration spectrum. Specifically, we first introduce HPDv3++, a 212K dual-dimension preference dataset annotated for text fidelity and aesthetic quality using a recent high-capability (Qwen-Image) model with human supervision. We then propose a two-stage training framework. Stage 1 employs data-aware orthogonal gradient projection to incorporate diverse aesthetic perception from HPDv3++ while preserving the original effective human preference knowledge in HPSv3. Stage 2 further leverages unlabeled data from T2I models spanning different capability levels and RL iterations, and introduces a joint capability-iterations conditioned signal for the reward model together with a standard deviation-driven unsupervised guidance mechanism, strengthening reward model across the capability-iteration spectrum. HPSv3++ achieves state-of-the-art preference prediction, outperforming HPSv3 9.8% on HPDv3, 5.5% on GenAI-Bench, while achieving 79.1%/88.1% on our proposed HPDv3++. When used for T2I RL training, it consistently improves GenEval scores across diverse T2I models, demonstrating its wide-range capabilities. The code is available at \urlhttps://github.com/PlantPotatoOnMoon/HPSv3-PlusPlus

Abstract:
3D scene graph generation aims to construct structured scene representations composed of object nodes and spatial relation edges, forming a core component for scene understanding. Although 3D Gaussian Splatting (3DGS) has emerged as a dominant paradigm for 3D scene representation, existing 3DGS-based methods primarily focus on isolated object retrieval, leaving a gap in relation modeling and compatibility with Gaussian primitives. In this paper, we propose G2G in bridging Gaussian representations and Graph-structured scene understanding, which operates in two stages. In the hierarchical open-vocabulary clustering stage, we mitigate multi-view semantic collapse by introducing adaptive neighborhood consistency, which reduces boundary feature contamination. To further structure the representation, Gaussians are organized using a two-level Gaussian codebook encoding. Then, a view bank module is adopted to preserve the full multi-view CLIP semantic distributions. In the relation graph generation stage, we introduce the relation modeling mechanism that predicts embeddings for object cluster pairs, while capturing their spatial relationships. To model relation directionality, the mechanism employs a gated module with dual-path cross-attention, forming a discriminative embedding space that enables end-to-end inference from Gaussian-based objects to structured scene graphs. Extensive experiments demonstrate that G2G surpasses state-of-the-art methods, improving the mIoU of open-vocabulary 3D semantic segmentation by 21.2% and relationship prediction by 17.6%. \hrefhttps://github.com/Tomswjy/G2G\textcolorblueProject page.

Abstract:
Perceiving multimodal cues and forecasting fine-grained actions from an egocentric (Ego) perspective is vital for applications like robot manipulation. However, previous studies either rely mainly on under-informed visual inputs to predict coarse human motions or follow the VRM/VLA paradigm, which suffers from insufficient robot data and the gap between human and robot embodiments. We observe that 3D hand pose naturally serves as a unified representation to bridge human-robot actions. Hence, we investigate an under-explored Vision-Language guided Egocentric 3D Hand Pose Forecasting (VL-EHPF) task, which aims to predict future Ego 3D hand poses from visual observations, a language instruction, and pose states. To overcome the limited field-of-view and highly dynamic motions in the Ego view, we propose a framework dubbed Exo2EgoPose, which innovatively leverages holistic and stable exocentric (Exo) demonstrations as guidance to compensate for partial and dynamic Ego-view cues. Specifically, we introduce a Dual-level Exocentric Reconstruction Module (DERM), which incorporates the paired Exo videos as supervision to reconstruct their video-level and chunked frame-level representations, thereby modeling spatial contexts and temporal dynamics. Then, the Global-to-Local Modulation Module (GLMM) utilizes the reconstructed hierarchical Exo representations for progressive feature refinement via attention mechanisms and adaptive modulation, enabling comprehensive Exo guidance for accurate Ego hand pose forecasting. Extensive experiments on AssemblyHands, Ego-Exo4D, and our newly constructed EgoMe-pose benchmarks show the superiority of our method, which outperforms state-of-the-art methods by a large margin. Moreover, it demonstrates an effective human-to-robot transfer capability and yields improvements on the CALVIN dataset. Code is available at \hrefhttps://github.com/ZhaofengSHI/Exo2EgoPosehttps://github.com/ZhaofengSHI/Exo2EgoPose.

Abstract:
Ultra-High-Resolution (UHR) remote sensing image understanding requires Vision-Language Models (VLMs) to capture both the global scene layout and sparse yet task-critical local details under limited computational budgets. Existing methods mainly follow two paradigms. One is passive perception, which relies on resolution expansion or token compression and may therefore discard fine-grained details. The other is active perception, which depends on multi-round zooming and search, but suffers from high latency, contextual fragmentation, and error accumulation. We argue that a more effective path toward UHR understanding lies not in accessing more, but in organizing better. To this end, we propose WeaveEarth, a training-free framework that reformulates UHR understanding as a problem of structured evidence construction and reasoning under global context constraints. Specifically, WeaveEarth first employs Global-Aware Evidence Construction to select a compact, low-redundancy, and spatially complementary Minimal Support Evidence Set. It then introduces Structured Evidence Reasoning, which weaves local evidence, spatial metadata, and relative topology into a unified reasoning interface, thereby enhancing the VLM's ability to perform global-local joint reasoning. Extensive experiments show that WeaveEarth consistently outperforms strong baselines and existing UHR methods across multiple UHR remote sensing benchmarks and multiple frozen VLM backbones. Code is available at https://github.com/XianZhi-Ma/WeaveEarth.

Authors: Qiang Lyu (University of Chinese Academy of Sciences), Tao Yang (University of Chinese Academy of Sciences), Kunze Wu (Chongqing University), Yuheng Li (Sichuan University), Yuanjun Liao (Sichuan University), Zeyu Qiu (University of Chinese Academy of Sciences), Haolin Shi (Youku), Yuning Gong (Sichuan University), Jianlou Si (Alibaba Cloud), Fangda Chen (The Hong Kong University of Science and Technology (Guangzhou)), Bing Zhang (University of Chinese Academy of Sciences), Junyuan Xiao (Tsinghua University), Yanchao Zhao (University of Chinese Academy of Sciences), Zicong Chen (Beihang University), Yichen Liu (University of Chinese Academy of Sciences), Chongxiao Wang (Youku), Jun Huang (Alibaba Cloud), Fei Ding (Youku), Zhou Jiang (South China University of Technology), Zhihao Li (Harbin Institute of Technology), Chun Pong Lau (City University of Hong Kong), Weiqiang Wang (University of Chinese Academy of Sciences)

Abstract:
Cross-domain image composition aims to insert a photorealistic foreground into a background of a different visual domain (\emphe.g., sketch, painting, cartoon) while preserving object identity and achieving stylistic coherence. However, the lack of natural supervision where the same object appears in both photorealistic and artistic domains has confined prior work to training-free pipelines that hinge on accurate image inversion, or methods that rely on rough supervision constructed by retrieving photorealistic objects visually and semantically similar to those in artistic paintings. Our key insight is to repurpose style transfer as a scalable data synthesis pipeline. Leveraging content images from OpenImages and style images from Style30k, we construct OpenImages-Style-600k, a large-scale dataset spanning 100k content images and 500 diverse styles, where each content image is paired with 5 styles. The proposed pipeline naturally ensures that the foreground image shares the same appearance and identity as the corresponding object in the target composite. Built upon OpenImages-Style-600k, we propose StyleCom, the first training-based framework for cross-domain image composition \emphwith precisely paired supervision. Based on FLUX.1-Dev, StyleCom conditions on a \emphcollage prior (i.e., the foreground cutout stitched onto the background at the target location) together with a text prompt describing the target scene. The collage prior provides the model with an initial layout of the desired composition and effectively prevents foreground vanishing during inference. StyleCom eliminates the need for image inversion required by training-free methods, while achieving state-of-the-art results on the widely-adopted TF-ICON benchmark. Our code and dataset will be available at \urlhttps://github.com/fikry102/StyleCom.

Authors: Zilong Ling (Anhui Provincial International Joint Research Center for Advanced Technology in Medical Imaging, School of Computer Science and Technology, Anhui University), Huabin Wang (Anhui Provincial International Joint Research Center for Advanced Technology in Medical Imaging, School of Computer Science and Technology, Anhui University), Liang Du (School of Computer and Information Technology, Shanxi University), Xuejun Li (Anhui Provincial International Joint Research Center for Advanced Technology in Medical Imaging, School of Computer Science and Technology, Anhui University), Peng Zhou (Anhui Provincial International Joint Research Center for Advanced Technology in Medical Imaging, School of Computer Science and Technology, Anhui University)

Abstract:
Active image classification is a fundamental task in image processing. One of the most popular active image classification methods is the uncertainty-based method, which selects the uncertain images for querying annotations. Conventional uncertainty-based active learning methods often regard the class boundary samples as the uncertain samples. However, in this paper, we observe a new kind of uncertain samples in active learning, which the model often misclassifies with high confidence, and thus we call them misleading samples. Since the model often has high confidence in these samples, without the annotations, the model does not even know they are uncertain or difficult samples, not to mention that the model knows itself makes mistakes on them. Therefore, misleading samples are much more difficult to identify, and conventional active learning methods often ignore them. To tackle this problem, we theoretically analyze how they were generated, and inspired by multi-modal learning, we discover the inconsistency property of them in the original spatial modal and frequency modal. Based on this analysis, we design a simple yet effective bi-modal method to identify the misleading samples. Then, we integrate it into an uncertainty-based framework to discover both kinds of uncertain samples. The extensive experiments show that our method discovers uncertain samples more comprehensively and performs better in the image classification task. Code is available at: https://github.com/Luming580/BUAL.

Abstract:
Audio-Visual Instance Segmentation (AVIS) aims to simultaneously classify, segment, and track sounding objects within video sequences. Unlike Audio-Visual Semantic Segmentation (AVS), AVIS involves instance-level modeling across longer video sequences, introducing two key challenges: (1) complex modality-state changes disrupt long-range modeling, and (2) substantial structural and distributional discrepancies between modalities hinder precise instance-level association. Existing methods rely on fixed-step Transformers and recursive Mamba models, lacking adaptability to modality-state changes. In addition, methods performing implicit matching ignore the inherent distributional inconsistencies. To address these issues, we propose a framework with Adaptive Dynamic Step Modulation (ADSM) and Optimal Transport-based Matching Modulation (OT-MM). ADSM adaptively modulates Mamba step sizes using temporal variation, cross-modal discrepancy, and historical context, balancing rapid response to modality-state changes with stable long-range modeling. OT-MM explicitly formulates instance-level cross-modal matching as an entropy-regularized optimal transport problem solved via log-domain Sinkhorn iterations, and further enforces distribution-level coherence with an MMD regularizer. Extensive experiments demonstrate state-of-the-art performance on the AVIS benchmark (+3.76 FSLA, +2.75 HOTA, +2.58 mAP), verified through comprehensive qualitative visualizations. The code and model are available at \urlhttps://github.com/happylife-pk/SAMOT.

Abstract:
While visual token pruning is essential for efficient Multimodal Large Language Models (MLLMs), existing training-free methods suffer from a critical limitation: they rely on static, instantaneous heuristics to perform irreversible filtering. This approach ignores the hierarchical nature of MLLMs, where token importance often evolves dynamically rather than remaining fixed across layers. Consequently, tokens essential for deep-layer reasoning are often prematurely discarded by shallow-layer estimates. To address this, we propose Trend-aware Pruning, a novel framework that elevates pruning from a local snapshot decision to a temporal trajectory modeling problem. Instead of relying on isolated scores, our method captures the momentum of attention flow. This enables a dynamic rectification mechanism that selectively reactivates ``late-blooming'' tokens---those initially undervalued but exhibiting rising semantic importance---thereby preventing the loss of critical visual cues. Extensive experiments demonstrate that our approach achieves a superior efficiency-performance trade-off across diverse multimodal tasks. Notably, it reduces visual tokens by over 77.8%, retaining only approximately 23 tokens in the final layer while maintaining competitive performance, offering a robust and reversible solution for high-efficiency multimodal inference. The project page is available at https://github.com/JieMaMagic/Trend-aware-Pruning

Abstract:
Video customization based on Text-to-Video (T2V) models aims to learn specific features from reference data to generate controllable videos. While significant strides have been made in image stylization and video motion customization, simultaneously controlling multiple concepts, such as content, style, and motion, remains a major challenge. In this work, we systematically define the task of multi-concept video customization, which requires the joint control of content, style, and motion. To facilitate research in this area, we construct a comprehensive benchmark and propose Disco-LoRA, a unified framework designed to tackle this problem by disentangling and flexibly recombining different concepts in two stages: (1) We decompose the objective into two sub-tasks: Content-Style and Content-Motion. Each sub-task is addressed using our Iterative Dual-LoRA Disentanglement Framework, which effectively disentangles distinct concepts within the data. (2) We identify layer-wise weight trends as crucial for LoRA identity, while weight magnitudes dictate composability. To harmonize these scales, we propose a Z-score-based statistical regularization that aligns weight distributions, preserving layer-wise trends while minimizing interference between different LoRAs. Extensive experiments show that Disco-LoRA excels in multi-concept video customization, effectively preserving appearance, style, and motion for controllable text-to-video generation. Our project page is available at https://xuxuancheng0208.github.io/discolora_page/.

Abstract:
Human identity-preserving text-to-video generation remains challenging under large changes in viewpoint, facial expression, illumination, and motion. Existing methods condition the generator on a single reference portrait, but a static image cannot capture how identity-bearing cues evolve across views and expressions, leading to face deformation, pose locking, identity drift, or over-smoothed faces. We observe that a short reference clip naturally provides richer temporal and multi-view identity cues than any single image, motivating a video-referential formulation. This richer signal, however, introduces a new challenge: identity evidence is distributed across many frames and must be distilled into a compact, stable representation under a limited token budget. To this end, we propose Slot-ID, a lightweight identity-conditioning framework built on a frozen text-to-video backbone. Slot-ID employs a slot-based temporal identity encoder with Sinkhorn-routed iterative reading to distill a compact, stable set of identity tokens from the reference clip, complemented by an image-anchor stream for dual-source conditioning. Extensive experiments demonstrate that Slot-ID outperforms state-of-the-art methods in identity preservation and visual naturalness while remaining competitive in prompt following, with particularly large gains under challenging pose, expression, and motion variations. The project page is available at \urlhttps://gapszju.github.io/SlotID.

Abstract:
Anchor graph clustering is an effective approach for handling large-scale clustering. Its good performance relies on the assumption that the anchors distribution is approximately consistent with the samples distribution. However, this assumption is difficult to satisfy in real-world applications. When there is a significant discrepancy between the two distributions, the clustering performance degrades markedly. To address this issue, we treat the anchor graph as an effective representation of the anchors. By studying the balance of anchors across each cluster, we approximately ensure the consistency between the anchor distribution and the sample distribution, and we further derive the theory of balanced anchor distribution across clusters. After that, leveraging the probabilistic properties of the anchor graph itself and the smoothness of labels, we propose an anchor graph clustering model based on tensor regression. In our model, low-rankness and smoothness are unified in the gradient domain to simultaneously capture inter-view high-order correlations and intra-view local continuity. Extensive experiments on benchmark datasets demonstrate that the proposed method outperforms competing methods on most datasets. The source code and appendix are available at \urlhttps://github.com/XDZSX/Supplementary_Material.git.

Abstract:
3D MLLMs recognize objects via pattern matching, which fails when geometrically similar objects require functional reasoning to disambiguate (e.g., dresser vs.\ wardrobe). Our per-sample \emphGeometric Confusion Score (GCS) exposes this failure: baseline accuracy is 33.7% on ModelNet40's most ambiguous 20% versus 84.0% on its least ambiguous 20%. We identify a \emphfunction-as-annotation pattern, in which functional language decorates a known identity instead of supporting inference, and hypothesize that reversing this information flow should concentrate improvements on the most ambiguous samples. PFIR(Part-level Functional Inference for Recognition) reverses that flow. Built on PartNeXt part segmentation, its data pipeline produces 507K QA pairs with functional CoTs (geometry \to function \to identity), and its progressive curriculum culminates in dual-objective GRPO: strict functional reasoning and descriptive grounding for Objaverse's diverse distribution. The two RL objectives yield nearly orthogonal task vectors (\cos = 0.056), enabling composition via Task Arithmetic. Under a strict dual OOD protocol (Objaverse \to ModelNet40; XYZRGB \to XYZ-only), PFIR raises zero-shot accuracy from 60.5% to 66.7%, with gains concentrated 2.6× more on geometrically ambiguous samples---supporting part-level functional reasoning as an effective route to more generalizable 3D recognition. Code: \urlhttps://github.com/this233/pfir-3d.

Abstract:
3D Gaussian Splatting (3DGS) supports high-fidelity talking head synthesis, but fine-grained regional expression control remains difficult when speech-driven motion and explicit Action Unit (AU) controls affect overlapping facial regions. Frame-level AU trajectories may also contain high-frequency noise that produces unstable deformation. We present EmoZone-Talker, an AU-conditioned framework that treats these issues as spatial responsibility allocation and temporal conditioning. Synergy Zones with Prioritized Attention Bias (SZ-PAB) introduces soft region-aware priors that bias audio toward articulation regions and AUs toward upper-face regions while retaining a learnable transition zone. The Channel-Independent Temporal AU Encoder (CIT-AE) applies learnable local temporal filtering before multimodal fusion. Integrated with 3D Gaussian deformation, these components improve measured AU controllability and temporal stability while maintaining rendering quality and lip synchronization. The method uses editable AU trajectories as external controls rather than inferring emotion from audio alone. Code is available on \urlhttps://github.com/HMQH/Emozone-Talker.

Abstract:
As an important subfield of cross-modal generation, synthesizing static visual content in the form of images from audio, namely audio-to-image (A2I) generation, has attracted increasing research attention in recent years. Nevertheless, despite the remarkable visual quality of modern text-to-image (T2I) models, the performance of A2I remains fundamentally limited by traditional datasets, which often lack both high-fidelity images and precise cross-modal alignment. As a result, existing methods still struggle to achieve high-quality audio-to-image generation through finetuning strong T2I models, thereby constraining practical applications in this area. Motivated by this gap, we introduce A2I-Set, a unified, high-quality tri-modal dataset consisting of 323K paired audio, images, and detailed text captions, specifically designed for audio-visual research, including audio-conditioned image generation. Besides, we developed a new mixed-source test set for the A2I task through human supervision. We further propose an A2I model, AudioCanvas, fine-tuned on our A2I-Set. Experiments show that AudioCanvas achieves more visually expressive as well as cross-modal alignment results that generally outperforming existing approaches. Our dataset and source code are available at \urlhttps://github.com/gdx012/A2I-Generation.

Abstract:
Referring Multi-Object Tracking in satellite videos (RMOT-SV) aims to track multiple targets in remote sensing imagery using natural language descriptions. Existing benchmarks remain limited in scale and annotation diversity, with short and simple expressions that provide insufficient support for complex scenarios involving multiple tiny objects. To address these limitations, we present SatRMOT, the largest and most diverse benchmark for referring multi-object tracking in satellite videos. SatRMOT is characterized by three key properties: (i) larger scale, with 523 videos, 138k frames, and 19k referring expressions, substantially exceeding existing RMOT-SV benchmarks; (ii) richer expressions, where each video is annotated with diverse natural language descriptions involving motion patterns, spatial relations, and contextual cues; and (iii) multi-object referring for tiny targets, where expressions frequently describe multiple small objects, particularly cars, in dense and complex satellite scenes. A multi-agent collaborative framework is developed to enable scalable and high-quality annotation with reduced human labor and time costs. Building upon this dataset, we propose Dual Refinement Track (DRTrack), a strong baseline for RMOT-SV that performs joint visual-linguistic refinement. Extensive experiments demonstrate the challenging nature of our dataset and the effectiveness of our method. The dataset will be released at \urlhttps://satrmot.github.io/.

Abstract:
With the adoption of 3D Gaussian Splatting (3DGS) in real-time rendering, VR/AR, NeRF acceleration, and 3D content creation, the need for standardized evaluation on large-scale real-world scenes is becoming increasingly urgent. However, existing datasets and benchmarks are mainly designed for object-centric or small-scale settings, and thus fail to capture challenges in large scenes, such as severe scale variation, complex occlusion, repetitive textures, fine-grained structures, and reflective interference. As a result, methods still lack a unified and reliable benchmark for large-scale 3DGS reconstruction and compression-oriented evaluation. To address this gap, we introduce PKU-GS, a large-scale real-world dataset and benchmark for Gaussian Splat Compression (GSC). Built from high-resolution UAV imagery captured under controlled conditions, PKU-GS covers large-scale scene categories, including architectural structures and open spaces, with diverse challenging samples involving complex textures, glass reflections, ultra-large buildings, and severe tree occlusions. The dataset contains 188 scenes and 8.64 million images, totaling more than 20 TB of data. Based on this dataset, we establish a unified processing and evaluation protocol covering the full pipeline from video sampling and camera pose estimation to 3DGS training and novel-view evaluation, and provide baseline results and ablation studies for systematic comparison of reconstruction quality, perceptual fidelity, and model compactness. The dataset is available at \urlhttps://ykh6666.github.io/PKUGS/.

Abstract:
Recent advances in AIGC have exacerbated the misuse of malicious deepfake content, making the development of reliable deepfake detection methods an essential means to address this challenge. Although existing deepfake detection models demonstrate outstanding performance in detection metrics, most methods only provide simple binary classification results, lacking interpretability. Recent studies have attempted to enhance the interpretability of classification results by providing spatial manipulation masks or temporal forgery segments. However, due to the limitations of forgery datasets, the practical effectiveness of these methods remains suboptimal. The primary reason lies in the fact that most existing deepfake datasets contain only binary labels, with limited variety in forgery scenarios, insufficient diversity in deepfake types, and relatively small data scales, making them inadequate for complex real-world scenarios. To address this predicament, we construct a novel large-scale deepfake detection and localization (DDL) dataset containing 1.4M+ forged samples and encompassing 80 distinct deepfake methods. The DDL design incorporates four key innovations: (1) Comprehensive Deepfake Methods (covering 7 different generation architectures and a total of 80 methods), (2) Varied Manipulation Modes (incorporating 7 classic and 3 novel forgery modes), (3) Diverse Forgery Scenarios and Modalities (including 3 scenarios and 3 modalities), and (4) Fine-grained Forgery Annotations (providing 1.18M+ precise spatial masks and 0.23M+ precise temporal segments). Through these improvements, our DDL not only provides a more challenging benchmark for complex real-world forgeries but also offers crucial support for building next-generation deepfake detection, localization, and interpretability methods. The dataset and supplemental materials are publicly available at https://github.com/inclusionConf/DDL.

Abstract:
Vision-Language Models (VLMs) now achieve over 99% zero-shot accuracy on Chinese-character CAPTCHAs, rendering traditional hardening strategies (complex backgrounds, stylized fonts, character distortion) ineffective. Adversarial perturbation offers an emerging defense by exploiting model vulnerabilities at the decision-boundary level, yet existing methods originate from disparate communities and have never been compared under identical conditions; moreover, the role of \emphperturbation input modality has not been systematically characterized across heterogeneous VLM architectures. To bridge these gaps, we introduce CaptchaBench, a large-scale benchmark comprising 840,000 base Chinese-character CAPTCHA images (\rev3,500 characters sampled from GB2312 Level-1 × 120 backgrounds × 2 generative pipelines), with a stratified evaluation subset of \rev16,000 adversarially perturbed images. \revEight perturbation methods are organized into three input-modality groups (Image-only: ASPL, Glaze, AMP, XTransfer, \revAnyAttack; Text-only: Nightshade; Image+Text: MMCoA, \revCoA) and assessed under a unified three-question framework: \revQ1 Direct-Recognition Accuracy (DRA) and Q1 Failure, Text Visibility Rate (TVR), and \revQ3 Attack Success Rate (ASR), \revcapturing complementary black-box response behaviors. Baseline evaluation across five commercial VLMs (Qwen-VL, Kimi~2.5, GPT-5.2, Gemini-3.0, GLM-4V) yields three key observations: (1)~perturbation input modality is associated with distinct black-box response patterns: image-only methods achieve high visual confusion \revbut variable VLM-reported visibility, while methods differ in cross-VLM variation; (2)~Gemini-3.0 exhibits a pronounced, model-specific TVR \revresponse pattern (62.4--97.3% vs.\ \leq20.8% for all other VLMs; \geq3.0× vs.\ the highest non-Gemini TVR across all eight methods), \revthat persists across prompt variants; and (3)~stroke-bin Q3 ASR varies across methods, with the 16--20-stroke bin high in this evaluation; the sparsely sampled 21+ bin is reported descriptively. This motivates character structure as a CAPTCHA design dimension. \revA 29-participant four-choice study confirms high human readability of perturbed CAPTCHAs (98.9% accuracy). The dataset and evaluation code are released under CC~BY~4.0. \revExtended tables, figures, and visual case studies are available as online supplementary material on the CaptchaBench dataset website at \urlhttps://github.com/Linzy19/CaptchaBench.

Abstract:
Fine-grained visual understanding and high-level reasoning in real-world open-water environments remain under-explored due to the lack of dedicated benchmarks. We introduce MARINER, a comprehensive benchmark built under the novel Entity-Environment-Event (3E) paradigm. MARINER contains 16,629 multi-source maritime images with 63 fine-grained vessel categories, diverse adverse environments, and 5 typical dynamic maritime incidents, covering fine-grained classification, object detection, and visual question answering tasks. We conduct extensive evaluations on mainstream Multimodal Large language models (MLLMs) and establish baselines, revealing that even advanced models struggle with fine-grained discrimination and causal reasoning in complex marine scenes. As a dedicated maritime benchmark, MARINER fills the gap of realistic and cognitive-level evaluation for maritime multimodal understanding, and promotes future research on robust vision-language models for open-water applications. Appendix and supplementary materials are available at https://lxixim.github.io/MARINER.

Abstract:
Efficiently locating clinically relevant moments in long surgical videos is essential for surgical education, yet the medical domain still lacks a dedicated benchmark for surgical video moment retrieval (VMR). A practical benchmark should include both a clinically grounded dataset and a strong baseline, yet building such a benchmark is challenging in two aspects. Dataset construction requires reliable expert temporal annotations and medically grounded prior knowledge. Baseline construction must handle redundancy, sparse targets, limited input budgets, and domain-specific prior knowledge. To address these issues, we introduce SurgScope, a benchmark for surgical VMR built from long real surgical videos across five procedure types. SurgScope provides expert-annotated temporal boundaries and surgeon-verified prior knowledge, and supports both phase localization and instrument localization. We further establish a strong baseline with two key components: Domain-specific Prior-knowledge Prompting and Redundancy-guided Dynamic Sampling (RDS). The former adds surgeon-verified prior knowledge into prompt, while the latter suppresses redundant frames. Experiments show that SurgScope is challenging for current MLLM-based VMR methods, while fine-tuning on SurgScope substantially improves performance. Domain-specific prior knowledge benefits phase localization, and RDS consistently outperforms default sampling strategies. SurgScope is publicly available at \urlhttps://github.com/yhz2003/SurgScope.

Abstract:
Multimodal Emotion Recognition (MER) is a fundamental task in multimedia understanding, where state-of-the-art methods have achieved prominent success relying on the assumption of complete, well-aligned multimodal inputs. However, real-world unconstrained scenarios often suffer from unpredictable modality degradation and information loss, leading to catastrophic performance collapse of existing MER models that heavily depend on facial cues as the core inference anchor. Worse still, there remains a critical research gap: few systematic works exist to evaluate model robustness under diverse, fine-grained incomplete information settings, as prior works are limited to coarse-grained full-modality removal and neglect interpretable analysis of emotional reasoning processes. To bridge this gap, we formally define the Incomplete Multimodal Emotion Recognition (IMER) task, which benchmarks model generalization and robustness under multi-grained modality and emotional information loss. We further construct MIME, a dedicated IMER benchmark with 2,000 video segments covering realistic fine-grained facial-detail degradation, severe facial degradation, and extreme full-modality loss, surpassing the limited coarse-grained settings of prior works. We also propose the Chain of Emotion (CoE) analysis paradigm with tailored evaluation metrics to dissect the fine-grained emotional cue perception and reasoning process of Multimodal Large Language Models (MLLMs). Extensive experiments on seven mainstream MLLMs reveal that body language and scene context act as critical compensatory mechanisms under facial degradation, while a pronounced gap exists between models' basic visual perception ability and high-level cognitive reasoning. Our work provides a focused evaluation platform and actionable insights for advancing robust multimodal emotion recognition. Our dataset and code are public at https://yuxinokk.github.io/MIME.

Abstract:
Generative AI can now produce highly realistic images, yet current models still exhibit subtle but critical defects that undermine their reliability. While existing AI-generated image (AGI) evaluation benchmarks have made notable progress, comprehensive AGI defect diagnosis remains underexplored. To bridge this gap, we introduce AGIDefect-4K, a richly annotated dataset of 4,000 images from 15 state-of-the-art generative models spanning both open-source and closed-source systems. AGIDefect-4K features hierarchical defect annotations: (1) detection labels identifying whether defects exist, (2) pixel-level segmentation masks localizing defective regions, and (3) detailed textual explanations characterizing defect types and their perceptual impact. Each image is further annotated with an overall quality score. Building on this, we present AGIDA (AGI Defect Assistant), a baseline framework leveraging Multimodal Large Language Models (MLLMs) for joint defect detection, localization, explanation, and quality prediction. Comprehensive benchmarking on AGIDefect-4K reveals that AGI defect understanding remains challenging, underscoring the value of this dataset. The dataset is publicly available at https://github.com/sxfly99/AGIDefect-4K.

Abstract:
Scene text spotting requires high-precision alignment between textual recognition and spatial localization. While visual-token grounding has emerged as a promising formulation for Multimodal Large Language Models (MLLMs), the previous multi-patch paradigm often introduces redundant noise and localization ambiguity, particularly for dense or small text instances. To address this, we propose Single-Patch Text Spotting (SPaTS), a vision-centric framework that routes each text instance through a single anchor visual token and then recovers geometry via full-image refinement. To accurately identify this anchor without oracle labels, we introduce Single-Patch Selective Optimization (SPaSO), a reinforcement learning framework that optimizes discrete visual-token selection using patch-level rewards. To further improve representation robustness and localization precision, we introduce Directional Embedding Alignment (DEA) to suppress unstable norm bias by decoupling feature magnitude and direction, and Patch-Enhanced Decoding (PED) to fuse the routed anchor with language semantics and cross-attend over the full-image feature map for geometry-aware boundary regression beyond coordinate-space surrogates. Extensive experiments demonstrate that SPaTS consistently and significantly outperforms both frontier closed-source MLLMs and OCR MLLMs. Code is available at https://github.com/eeNickTang/SPaTS.

Abstract:
Recent studies have shown that multimodal large language models (MLLMs) are highly vulnerable to hidden-pattern visual illusions, where the hidden content is imperceptible to models but obvious to humans. This deficiency highlights a perceptual misalignment between current MLLMs and humans, and also introduces potential safety concerns. To systematically investigate this failure, we introduce IlluChar, a comprehensive and challenging illusion dataset, and uncover a key underlying mechanism for the models' failure: high-frequency attention bias, where the models are easily distracted by high-frequency background textures in illusion images, causing them to overlook hidden patterns. To address this issue, we propose the Strategy of Multi-Scale Perception (SMSP), a plug-and-play framework that aligns with human visual perceptual strategies. By suppressing distracting high-frequency background signals, SMSP generates images closer to human perception. Our experiments demonstrate that SMSP significantly improves the performance of all evaluated MLLMs on illusion images, for instance, increasing the accuracy of Qwen3-VL-8B-Instruct from 13.0% to 84.0%. Our work provides novel insights into MLLMs' visual perception, and offers a practical and robust solution to enhance it.

Abstract:
Chest radiographs often reach clinicians in low-resource telemedicine settings as display-recaptured smartphone images rather than native digital studies. These images introduce structured acquisition artifacts that create clinically meaningful distribution shift for radiology vision- anguage models (VLMs). Standard Conformal Risk Control (CRC) enforces average-risk guarantees, but it can still violate worst-group safety on harder corruption subgroups, a failure that remains invisible under aggregate evaluation. We trace this behavior to \emphcross-modal decoupling: when visual evidence is degraded, the language decoder remains fluent and confident, producing disagreement across stochastic samples without correspondingly high token entropy. Motivated by this effect, we introduce \emphFindings-Disagreement-Variance (FDV), an uncertainty score that combines entropy and cross-sample disagreement, together with \emphWorst-Group Conformal Risk Control (WG-CRC), a post-hoc calibration framework that targets subgroup-level safety without retraining. On TeleRadShift, a structured 15-group benchmark, WG-CRC achieves 0/15 violations at 23.9% coverage, corresponding to an 8.9× improvement over Bonferroni at matched safety. The method transfers without recalibration to 1,800 real smartphone captures (0/9 violations, worst-group risk 0.492) and remains stable across 30 random splits (worst-group risk 0.506\pm0.022). Project Page: \urlhttps://gen-ai-research.github.io/teleradshift.

Abstract:
Clustering faces by identity in long, continuous video streams is a challenging task that serves as a building block for human-centric video analysis. The majority of existing face clustering models are offline, i.e., they assume the entire data to be available altogether. In many practical scenarios, however, the entire data is either unavailable or poses computational challenges to be processed in its entirety. A handful of online methods that exist rely on geometric visual features and simple thresholding techniques, which fail when the faces exhibit significant appearance variations in long, unconstrained videos. To this end, we present VISTA, the first vision language framework for online face clustering in long videos. VISTA leverages semantic reasoning in a controlled, uncertainty aware fashion to maintain both high cluster purity and long range identity continuity. This makes it robust to the variations in illumination, pose, and costume changes that the faces and characters exhibit in long real world videos. Extensive experiments on the VPCD benchmark dataset (with 15 full episodes from The Big Bang Theory [2007-2019], Buffy the Vampire Slayer [1997-2003], and Sherlock [2010-2017]) demonstrate that VISTA achieves a robust precision and recall balance, consistently improving over the strongest matched online unsupervised baseline (COFC) under identical detector, embeddings, and evaluation protocol. Code and supplementary material: https://github.com/SkyJiashu/VISTA.

Abstract:
The integration of physical laws and machine learning has emerged as a compelling paradigm for modeling complex real-world systems. Compared to traditional numerical solvers, this approach significantly improves computational efficiency while preserving physical consistency. However, the source--sink dynamics of global carbon fluxes are driven by a complex coupling of diverse physical and biogeochemical mechanisms, resulting in pronounced spatiotemporal heterogeneity. Existing physics-informed methods typically embed governing equations directly into model architectures or loss functions. Inherently, this imposes strong assumptions of regime uniformity and mechanistic homogeneity, neglecting the intricate interactions across different regimes and physical processes. Consequently, while these methods excel on idealized partial differential equations (PDEs), they often exhibit limited generalizability when applied to complex real-world systems. To bridge this gap, we propose a hierarchical Mixture-of-Experts (MoE) framework, termed Regime and Mechanism Decoupled Mixture-of-Experts (REM-MoE), that explicitly encodes regimes and evolutionary mechanisms as two independent inductive biases during prediction. The core innovation lies in structurally decoupling complex systems: first-level experts model regime-specific heterogeneity, while second-level experts capture the diversity of underlying mechanisms. By dynamically combining these two levels, REM-MoE adaptively represents evolutionary dynamics under varying ``regime--mechanism'' conditions, thereby modeling state-dependent changes in regime--mechanism interactions. We validate our approach on a global carbon flux prediction task, where observational signals are jointly driven by multiple mechanisms across diverse regimes. Experimental results show that the proposed method consistently outperforms strong baselines in prediction accuracy while learning regime-dependent mechanism representations that are aligned with physical reference patterns. Furthermore, REM-MoE maintains strong performance over longer forecasting horizons and yields physically interpretable routing patterns. The code is publicly available at: \hrefhttps://github.com/Son4ta/REM-MoEhttps://github.com/Son4ta/REM-MoE.

Abstract:
Diffusion Transformers (DiTs) have demonstrated exceptional performance in high-fidelity image and video generation. To alleviate their massive computational overhead, temporal feature caching has been proposed to bypass redundant computations. However, existing cache-then-forecast methods driven by derivative-based polynomials often cause severe quality degradation under high acceleration due to unstable long-step predictions. To address this bottleneck, we propose Barycentric Rational Forecasting with Chebyshev Enhancement (BRACE). Motivated by the observation that DiT feature trajectories are globally smooth yet frequently exhibit sharp irregularities and local non-smoothness, BRACE shifts the paradigm from derivative-driven polynomial extrapolation to feature-driven rational forecasting. Specifically, it maintains a local sliding window to cache sparse historical features and leverages adapted Chebyshev weights to formulate a barycentric rational function, directly aggregating these raw features to ensure numerical stability. Extensive experiments demonstrate that BRACE achieves state-of-the-art qualityefficiency trade-offs across various DiT architectures with negligible computational overhead. Further details are available on our \hrefhttps://youngkinlon.github.io/BRACE-Taming-Sharp-Irregularities-via-Barycentric-Rational-Forecasting-for-Fast-DiT-Inference/project page.

Abstract:
Recent advances in generative models have made synthetic images increasingly photorealistic, posing growing challenges to reliable and generalizable detection of AI-generated images. Existing methods typically rely on holistic image-level classification, which may overlook local artifacts, especially in high-resolution images. To address this limitation, we formulate this task as a falsification-driven multiple instance learning problem, where a single fake patch can falsify the whole image, while a real image requires all patches to be authentic. Based on this formulation, we propose Learning to Falsify (LTF), a hierarchical multiple instance learning framework that models an image as multiple bags of instances across different spatial levels and aggregates their predictions under the falsification-inspired principle. We further introduce Cross-Shuffle during training to encourage the model to learn local artifact evidence rather than image-level shortcuts. Furthermore, we propose the Falsification-driven Test-time Aggregation Strategy (FATAS), a plug-and-play module that extends the same falsification-inspired principle to existing detectors by aggregating patch-level predictions. Extensive experiments demonstrate that LTF achieves strong generalization and robustness across multiple benchmarks, while FATAS substantially improves representative detectors in high-resolution scenarios without retraining. Code is available at \urlhttps://github.com/HuiGuanLab/LTF.git

Abstract:
Video Corpus Moment Retrieval (VCMR) is hindered by high computational complexity and a scarcity of large-scale, specialized datasets, as existing benchmarks rely on labor-intensive manual annotations that limit scale, scene diversity, and training utility. To bridge these gaps, we present ACE (Automated Calibrated Enrichment), a fully automated pipeline for constructing large-scale VCMR datasets. ACE yields 108K diverse videos and employs a fine-to-coarse generation strategy that progresses from frame-level cues to event-level semantics. A boundary refinement model further optimizes timestamps, while multi-model validation with hallucination checks ensures annotation quality. Building on ACE, we propose SegRe, a unified framework that reformulates temporal localization as segment-level retrieval. Central to SegRe is a segment-based training strategy that synergizes multi-objective losses with intra-video negative mining, enabling fine-grained discrimination among temporally similar candidates within the same video. Experiments show that SegRe achieves state-of-the-art performance across multiple benchmarks in both standard and zero-shot evaluation settings. The code is available at \urlhttps://github.com/camellia517/ACE-Dataset/.

Abstract:
Multimodal Remote Sensing Image Fusion (MRSIF) faces a critical challenge of fusion collapse in few shot or noisy scenarios, such as those obscured by fog and clouds, where models over-rely on the dominant modality or suffer from severe semantic shifts, leading to the failure of the multimodal fusion mechanism. We have observed that the intrinsic geometric representation of the feature space can correspondingly characterize semantic information, thereby providing an effective geometric-driven solution for avoiding and correcting fusion collapse challenges. To be specific, we propose Residual Manifold Geometric Constrainer (ReMaCo), a novel plug-and-play fusion enhancement module that treats MRSIF task as a constrained progressive evolution process on non-Euclidean manifolds. Firstly, dual manifold constraints are designed: Lorentz hyperbolic space is utilized to model the intrinsic confidence among multimodal features, while unit hyperspherical space is employed to stabilize the directional consistency of feature semantics against noise-induced drift. Furthermore, Residual Fusion Feature Calibration (RFFC) module is designed to guide the model to focus on uncertain regions in multimodal inputs through hyperbolic energy maps and implement residual feature correction, which prevents the suboptimal state propagation. By transforming the optimization from purely data-driven search to geometric-guided refinement, ReMaCo provides a robust fusion paradigm for sophisticated MRSIF tasks. Extensive experiments demonstrate that our method effectively mitigates fusion collapse and achieves superior segmentation performance in challenging, interference-prone remote sensing scenarios. The code is available at https://github.com/PhilipHan-HITSZ/ReMaCo.

Abstract:
Visual AutoRegressive Modeling (VAR) has excelled in natural image generation via next-scale prediction, but its use on topology-structured data like human skeletons is still unexplored. VARPose is proposed to adaptively densify 2D sparse poses, thereby enriching the anatomical information available for 3D lifting models. Our core contributions are twofold. First, we introduce a Granularity-agnostic Pose Tokenizer (GPT), which employs a single hybrid codebook and a residual quantization strategy to encode poses of varying densities into a unified, multi-scale discrete representation. Our results demonstrate the strong generalizability of this representation. By decoupling the representation from the projection, we can successfully decode novel pose granularities using a frozen codebook with a retrained decoder. Second, we propose UniSkelar, a unified autoregressive model that treats ``joint density'' as ``scale''. UniSkelar learns to predict the token sequence for the next density level in a coarse-to-fine manner, conditioned on the sparsest pose. VARPose not only outperforms state-of-the-art methods and generalizes to unseen granularities, but also confers tangible performance gains on downstream tasks, such as 3D Pose Estimation and Human Mesh Recovery, through 2D pose densification. Our code and model are available at \textcolorpurple\urlhttps://github.com/BRL-SYSU/VARPose.git.

Abstract:
\beginabstract Survival analysis is critical for cancer prognosis and treatment planning, yet existing methods lack the transparency essential for clinical adoption. While recent pathology agents have demonstrated explainability in diagnostic tasks, they face three limitations for survival prediction: inability to integrate multimodal data, ineffective region-of-interest (ROI) exploration, and failure to leverage experiential learning from historical cases. We introduce \ours, the first hierarchical chain-of-thought (CoT)-enhanced multi-agent system for multimodal survival prediction. \ours~consists of two stages: (1) WSI-Gene CoT-Enhanced Case Bank Construction employs hierarchical analysis through Low-Magnification Screening, Cross-Modal Similarity-Aware Patch Mining, and Confidence-Aware Patch Mining for pathology images to effectively explore WSI ROI, while Gene-Stratified analysis processes six functional gene categories. Both branches generate structured reports with CoT reasoning, storing complete analytical processes for experiential learning. (2) Dichotomy-Based Multi-Expert Agent Inference retrieves similar cases via RAG and integrates multimodal reports with expert predictions through progressive interval refinement. Extensive experiments on five TCGA cohorts demonstrate \ours's superiority over conventional methods, proprietary MLLMs, and medical agents, establishing a new paradigm for explainable AI-driven survival prediction in precision oncology. Source code is released at \urlhttps://github.com/hotdogisme/SurvAgent. \endabstract

Abstract:
Conversational Speech Synthesis (CSS) is a fundamental component of human-computer interaction, aiming to generate contextually appropriate, expressive, and empathetic speech. However, facial expressions encode subtle and rich affective cues that are crucial for empathetic speech interaction, whereas existing approaches often overlook this important modality. In addition, the lack of large-scale natural conversational datasets with both speech and visual modalities also limits the development of visual affect understanding in conversational settings. To address these limitations, we propose FacialTalker, a facial-expression-aware CSS framework built upon a large language model backbone. To efficiently encode facial expressions, we propose AUTokenizer, a single-codebook visual tokenizer that discretizes each frame-level facial expression into a compact token, trained with supervision from combinations of facial Action Units. We further introduce a dual direct preference optimization (DualDPO) strategy, which extends the DPO by jointly imposing preference constraints on both visual and speech token sequences, to enhance the model's understanding of facial expressions and speech semantics in multimodal conversational contexts. Moreover, we construct VSDD-1K, a large-scale multimodal dialogue dataset collected through a fully automated pipeline from real-world Internet conversations, comprising over 1,033 hours of synchronized speaker videos and speech, with more than 85% of frames containing valid faces. Extensive objective and subjective experiments demonstrate that FacialTalker consistently outperforms strong baselines in facial-expression perception and speech synthesis quality, generating speech that is more natural, expressive, and better aligned with the conversational context. The results also validate the effectiveness of our training strategy and dataset construction pipeline. The project page: \texthttps://github.com/walker-hyf/FacialTalker.

Abstract:
Multi-view pedestrian detection (MVD) mitigates occlusion and missed detections in single-view settings by aggregating multi-camera observations into a unified bird's-eye-view (BEV) representation. Recent methods lift 2D image features into 3D space but typically rely on fixed-size voxel grids as the scene representation, whose spatially discrete and rigid modeling limits fine-grained localization and scale-adaptive representation of pedestrians. In this paper, we propose GaussianMVD, the first end-to-end trainable implicit 3D Gaussian framework for multi-view pedestrian detection. It directly learns a detection-oriented implicit 3D Gaussian representation from multi-view images using only BEV supervision, without external models or multi-stage optimization, enabling fine-grained and scale-adaptive pedestrian modeling through learnable 3D Gaussians. To eliminate alignment errors from directly mapping 2D pixels to 3D Gaussians, a voxel-aligned 3D Gaussian predictor raises feature association from 2D images to a unified 3D space. Furthermore, rasterization renders the BEV as a continuous probability distribution, yet traditional MVD decoders sample only discrete peak pixels and ignore the surrounding gradient information; we therefore propose posterior probability distribution decoding, which computes the expectation of the pedestrian distribution for more precise localization. Experiments on WildTrack and MultiviewX demonstrate that GaussianMVD achieves state-of-the-art performance among methods without extra supervision. Code is available at \urlhttps://github.com/sheldoncooper5/GaussianMVD.

Abstract:
Multimodal Entity Linking (MEL) aims to link ambiguous mentions in complex multimodal context to the correct entities in a knowledge base. Recent MEL methods have achieved promising performance by combining multimodal representation learning with auxiliary LLM-based reasoning. However, they still face two key challenges. First, fine-grained crossmodal interactions remain vulnerable to redundancy and noise, while existing methods still fall short in modeling the asymmetric roles and dependencies of global and local features. Second, current LLM-enhanced methods typically invoke LLMs indiscriminately for all samples, leading to substantial computational overhead and potential instability caused by hallucinations. To address these issues, we propose AVM-EDR, a novel framework for MEL. Specifically, we develop an Asymmetric VIB-MoE Crossmodal framework (AVM), which introduces Variational Information Bottleneck (VIB) to suppress noise in fine-grained crossmodal interactions and incorporate global semantics to guide the expert routing. We further design an Entropy-Driven LLM Refinement strategy (EDR), which estimates sample by the entropy-based uncertainty estimation and performs LLM-based scoring only for hard cases. In this way, AVM-EDR improves performance while reducing unnecessary reasoning cost and hallucination interference. Extensive experiments demonstrate the effectiveness and efficiency of the proposed framework. The code is available at \hrefhttps://github.com/zhouyuanyu/AVM-EDRhttps://github.com/zhouyuanyu/AVM-EDR.

Abstract:
Diffusion model based fusion of low-resolution multispectral (LRMS) and panchromatic (PAN) images has attracted increasing attention in recent years. However, most existing methods jointly treat PAN and multispectral images as conditions to the denoising network without explicitly distinguishing their distinct roles in the fusion process. As a result, the model is unable to impose differentiated constraints and guidance, leading to unsatisfactory spectral information preservation and spatial detail enhancement. Recent advancements have demonstrated that diffusion models tend to generate overall structures in the initial denoising stages and then progressively recover fine details in latter stages. Motivated by this observation, we propose an Asynchronous Diffusion (AsyDiff) model that progressively performs image fusion from structure reconstruction to detail refinement. Specifically, AsyDiff employs two dedicated data streams: a spatial stream for detail enhancement and a spectral stream for spectral preservation. Correspondingly, the diffusion process is divided into three stages. The structure reconstruction stage primarily uses the PAN stream to reconstruct the initial spatial structure, while the spectral refinement stage uses the multispectral stream to refine spectral information and improve detail consistency. An asynchronous transition stage is introduced between them to alleviate conflicts between the two streams. Extensive experiments on multiple satellite datasets demonstrate that the proposed method achieves state-of-the-art performance in both visual quality and quantitative evaluation metrics. The source code is available at https://github.com/yangyingkun/AsyDiff.

Abstract:
While existing CLIP-based Image Quality Assessment (IQA) models have demonstrated enhanced capabilities in perceiving visual quality, due to the small scale of IQA datasets, they still struggle to evaluate out-of-distribution distorted images, which deviate from the pre-training datasets. Although some mixed-dataset training methods have been proposed to enable IQA models to be trained on more diverse datasets, the limited scale of IQA data still restricts these methods' generalization ability. To address these issues, we propose a novel CLIP-based blind IQA method named Retrieval-enhanced local meta training (Relocat)-IQA, which aims to augment both of the image and text embeddings by online retrieval for better generalization. Specifically, we first propose an effective pre-training framework to align and encode the fine-grained multi-modal quality knowledge as an external embedding memory. Then the sample-specific quality knowledge is retrieved from the memory to enhance the image and text embeddings during the training of \method, which can greatly improve the zero-shot predictions at inference time. In addition, we propose a local meta-training strategy to utilize mixed datasets efficiently. Remarkably, we demonstrate the benefits of the retrieval-enhanced module for generalization not only through experiments but also from a theoretical perspective. Extensive experiments validate the effectiveness and generalization of our method. Supp. Mat. and details can be found at \urlhttps://github.com/zhyan567/Relocat-IQA.

Abstract:
While adversarial prompt tuning can enhance robustness of vision-language models efficiently, we find that existing methods aggravate robust generalization overfitting on seen classes, leading to a rapid degradation in performance against adversarial examples of unseen classes as training progresses. We empirically identify that this degradation stems from the tendency of the model to learn pseudo-robust features (i.e., non-generalizable shortcuts). To mitigate this, we propose ADAPT (Adversarial DisentAngled Prompt Tuning), a robust prompt tuning framework following the philosophy of ``Learning What Not to Learn''. Specifically, ADAPT uses a dual-prompt mechanism with a target prompt and a pool of decoy prompts. During training, the decoy prompts are guided to entrap diverse pseudo-robust features, while the target prompt is constrained to be orthogonal to the decoys in the embedding space to learn robust features. By disentangling the robust features from the pseudo-robust features, ADAPT effectively prevents robust generalization overfitting. We further provide an analysis showing that the orthogonal loss bounds the effect of shifts in pseudo-robust features on unseen classes, yielding a testing error guarantee. Empirically, extensive experiments demonstrate that ADAPT substantially improves the robustness of the target prompt on unseen classes. The code is available at \urlhttps://github.com/cheny02/ADAPT-ACMMM2026.

Abstract:
\beginabstract In low-light image enhancement (LLIE), hybrid loss optimization suffers from gradient conflicts and training oscillationsa challenge often overlooked by conventional batch-averaged strategies that ignore spatial heterogeneity. We propose a sample-level Energy-Conserved Conflict-aware Soft Gating Mechanism (ECSM) to address this issue. ECSM performs instance-wise adaptive weighting by modeling loss-specific gradient responses and inter-objective conflicts for each sample. An energy-conserved redistribution scheme further reallocates weights among objectives, preserving overall optimization strength while aligning each loss with its empirical reliability. Experiments on three benchmark settings and four representative LLIE backbones, evaluated using PSNR, SSIM, and LPIPS, demonstrate improved reconstruction fidelity and more stable optimization without introducing an additional inference module. Code is available at \urlhttps://github.com/werringwu/ECSM. \endabstract

Abstract:
Masked generative models have become a strong paradigm for text-to-motion synthesis, but they still treat motion frames too uniformly during masking, attention, and decoding. This is a poor match for motion, where local dynamic complexity varies sharply over time. We show that current masked motion generators degrade disproportionately on dynamically complex motions, and that frame-wise generation error is strongly correlated with motion dynamics. Motivated by this mismatch, we introduce the Motion Spectral Descriptor (MSD), a simple and parameter-free measure of local dynamic complexity computed from the short-time spectrum of motion velocity. Unlike learned difficulty predictors, MSD is deterministic, interpretable, and derived directly from the motion signal itself. We use MSD to make masked motion generation complexity-aware. In particular, MSD guides content-focused masking during training, provides a spectral similarity prior for self-attention, and can additionally modulate token-level sampling during iterative decoding. Built on top of masked motion generators, our method, DynMask, improves motion generation most clearly on dynamically complex motions while also yielding stronger overall FID on HumanML3D and KIT-ML. These results suggest that respecting local motion complexity is a useful design principle for masked motion generation.

Abstract:
Multi-class Unsupervised Anomaly Detection (MUAD) has attracted increasing research interest for its ability to develop a unified model across multiple classes, significantly reducing the computational cost of training separate models for distinct objects. However, despite the performance gains offered by advanced Transformer-based architectures, the identity shortcut issue persists. These models tend to directly copy inputs to outputs, narrowing the reconstruction error gap between normal and abnormal cases and hindering distinguishability. To address this, we propose ShortcutBreaker, a framework designed to disrupt such shortcuts through two core innovations. First, we introduce a Low-Rank Noisy Bottleneck (LRNB) derived from matrix rank inequality, which theoretically prevents trivial identity mapping by perturbing features across two distinct reconstruction groups. Second, we design a Frequency-Filter Unit (FFU) that integrates frequency spectral filtering and global-local masking. This mechanism further enlarges the anomaly score gap, with less computational cost than standard self-attention. Extensive evaluations on four benchmarks demonstrate the superiority of ShortcutBreaker. The code will be released at https://github.com/TumCCC/ShortcutBreaker.

Abstract:
As telecom fraud increasingly threatens global financial security, the demand for reliable detection models has surged. Advancing these models, however, is severely bottlenecked by two critical challenges: (i) the limited availability of real-world data due to strict privacy regulations on scam recordings; and (ii) poor annotation quality caused by unstructured formats and unstandardized categories. While existing works leverage LLMs and hand-crafted prompts to synthesize fraudulent dialogues, the resulting datasets often lack grounding in documented fraud cases, increasing the risk of factual and procedural drift. To address these challenges, we propose Narrative-to-Dialogue Reverse Reconstruction (NDRR), a framework for constructing case-grounded synthetic fraud dialogues from real-world case narratives. Through three stagesFraud Procedure Extraction, Fact-Anchored Dialogue Reconstruction, and Consistency Verificationwe construct FraudRecon, a large-scale dataset of 80,319 paired audio-text dialogues covering scam and matched benign interactions. It features comprehensive three-level annotations across 12 fraud categories, 6 scam-specific procedural phases, and 42 interaction signals. We also introduce FraudRecon-Bench, a five-task benchmark encompassing Fraud Detection, Category Classification, Phase Identification, Signal Recognition, and Evidence-Chain Reasoning. Extensive experiments on 13 LLMs (11 zero-shot and 2 fine-tuned) demonstrate that FraudRecon-Bench poses a significant challenge: the best zero-shot overall score is 53.3/100, while fine-tuning on FraudRecon improves the overall score by up to 32.8 points on this benchmark. Our code and data-access instructions are available at https://github.com/Henryhe09/FraudRecon.

Abstract:
Few-shot classification of whole-slide images (WSIs) is fundamentally challenging due to extreme data scarcity. Recent approaches enhance supervision by injecting multi-granular textual concept priors. However, they typically rely on generic imagetext alignment, treating all instanceconcept pairs as equally plausible. Such alignment overlooks the diagnostic principles inherent in these textual priors, leading to spurious matches that misguide instance aggregation: (i) cross-magnification regional coherence: tissue morphology remains semantically consistent across magnification levels; and (ii) microenvironment consistency: leveraging the spatial continuity of lesion regions to filter misleading isolated artifacts. To address this gap, we introduce C^2UOT, a novel alignment strategy that reformulates instanceconcept alignment as a principle-constrained unbalanced optimal transport (UOT) problem. C^2UOT integrates two biologically inspired inductive bias modules: (i) enforcing parentchild semantic consistency across magnifications via a guidancepenalty mechanism, and (ii) applying concept-familyaware spatial smoothing to suppress isolated artifacts while preserving lesion boundaries. Experiments across four datasets demonstrate that C^2UOT yields substantial gains by refining the generic alignment, achieving results that exceed SOTA baselines. Code is available at https://github.com/hkustgary/C2UOT.git.

Abstract:
Despite the remarkable progress over the past decades, accurately identifying small objects remains challenging because of their insufficient visual cues. Previous works typically attempt to construct discriminative representation of the small objects. However, the wide range frequency domain noises and label ambiguities have been greatly overlooked, which significantly hinders the accurate localization. To address these issues, we propose a novel small object detection (SOD) detector termed DyFrDet, which is able to precisely localize the small object by dynamically suppressing the background distractions in frequency domain. Specifically, we propose a Dynamic Frequency-aware Feature Pyramid Network (DyFrFPN) to adaptively suppress low-frequency redundancy and excessive high-frequency noises. The DyFrFPN transforms the hierarchical features into frequency domain representation, and introduces a Dynamic Band Predictor (DBP) to preserve the discriminative components for small object identification. Afterwards, we present a novel Label Disambiguation Module (LDM), which leverages probabilistic distributions to explicitly model and alleviate the inherent ambiguity of target labels, yielding efficient improvement in localization precision of the small objects with low-resolution. Extensive experiments demonstrate that DyFrDet achieves state-of-the-art performance across multiple benchmarks, indicating its effectiveness and robustness in various challenging scenarios. Our code is available at \urlhttps://github.com/ManOfStory/DyFrDet.

Abstract:
\beginabstract RGB-Event Video Person Re-Identification (RE-VReID) aims to retrieve specific person across non-overlapping cameras with complementary RGB videos and event streams. % However, existing methods often decouple spatial and temporal modeling, which limits their interaction. % In addition, global-level RGB-Event fusion fails to fully exploit fine-grained discriminative cues. % To address these issues, we propose Paths, a unified framework with spatio-temporal modeling and hierarchical multi-modal fusion for RE-VReID. % Specifically, we first design a Memory-Augmented Backbone (MAB) to maintain modality-specific identity prototypes for stable intra-modal representation learning. % Then, we propose a Prompt-aware Spatio-temporal Transformer (PST) to jointly model spatial and temporal cues within a unified Transformer. % Finally, we introduce a Hierarchical Multi-modal Fusion (HMF) to integrate RGB and event features at global and local levels. % With these modules, our framework can learn robust and discriminative representations for RE-VReID. % Extensive experiments on three public RE-VReID benchmarks including EvReID, MARS and iLIDS-VID, demonstrate the effectiveness of our proposed method. % The code is available at \urlhttps://github.com/Reflection0427/Paths. \endabstract

Abstract:
Hyperspectral imagery provides rich spectral signatures for monitoring Earth surface changes. Existing hyperspectral change detection (HCD) methods require fully retraining between the reference image and new observations, which is computationally prohibitive and impractical for dynamic monitoring. In this paper, we observe that the imaging condition shift is a significant trigger for this demand, and thus propose StyleHCD, an efficient parameter modulation paradigm for HCD, which rebuilds the detector into two parts: a basic detector trained on an initial bi-temporal pair and a Style-conditioned Parameter Injector (SPI) to predict land-cover changes and non-semantic factors, respectively. When new observation data is received, we freeze the basic detector and employ SPI to generate scene-specific parameters from the observed spectral features. Specifically, SPI first leverages a Sensitivity-aware Parameter Selection (SPS) strategy to locate style-sensitive layers and efficiently injects the knowledge via a Parameter Space Recalibration (PSR) mechanism. Moreover, we construct a dynamic HCD benchmark by simulating HSI time-series sequences under varying conditions. Extensive experiments show that inserting StyleHCD into existing HCD models achieves comparable performance to full retraining while reducing the number of trainable parameters to less than 0.004%. Code is released at https://github.com/GC-WSL/StyleHCD.

Abstract:
Diffusion models have achieved remarkable progress in image generation, but their increasing deployment raises serious concerns about privacy and copyright. In particular, fine-tuned models are highly vulnerable, as they are often fine-tuned on small and private datasets. Membership inference attacks (MIAs) are used to assess privacy risks by determining whether a specific sample was part of a models training data. Existing MIAs against diffusion models either assume obtaining the intermediate results or require training the shadow model on auxiliary datasets. In this work, we utilized a critical yet overlooked vulnerability: the widely used noise schedules fail to fully eliminate semantic information in the images, resulting in residual semantic signals even at the maximum noise step. We empirically demonstrate that the fine-tuned diffusion model captures hidden correlations between the residual semantics in initial noise and the original images. Building on this insight, we propose a simple yet effective membership inference attack, which injects semantic information into the initial noise and infers membership by analyzing the models generation result. Extensive experiments demonstrate that the semantic initial noise can strongly reveal membership information, highlighting the vulnerability of diffusion models to MIAs. Code is available at https://github.com/S3IC-Lab/NoiseMIA.

Abstract:
Medical vision language models (VLMs) have demonstrated remarkable multi-modal medical reasoning ability, especially when combined with localization of the anatomical entity. However, acquiring region annotations is costly and burdensome in medical imaging. Although automatic localization provides a practical alternative, the generated masks may contain localization errors and often offer only coarse semantic guidance, failing to capture the fine-grained, question-adaptive visual evidence required for reliable reasoning. Therefore, such generated masks should be regarded only as object priors rather than definitive answer-supporting evidence. In this paper, we introduce RARE, a Reliable and Adaptive Reasoning with grounded Evidence framework that progressively converts self-generated object priors into reliable and question-adaptive visual evidence for answer reasoning. The framework incorporates a counterfactual learning strategy that contrasts factual and counterfactual evidence views, and exploits their evidence differential to prevent error propagation. We further design a Query Mixture-of-Experts (MoE) module that dynamically selects and aggregates the reliable visual evidence most informative for the current question, enabling adaptive evidence encoding under different reasoning demands. In addition, we construct MOG-QA, a large-scale Medical Object-Grounded Question Answering benchmark covering eight imaging modalities, in which each case is organized as an image-mask-QA triplet. Extensive experiments demonstrate that our framework consistently improves medical VQA performance across multiple public datasets and outperforms previous state-of-the-art methods. Code and dataset are at https://github.com/Asphyxiate-Rye/RARE.

Abstract:
Vision Language Model (VLM) typically assume complete modality input during inference. However, their effectiveness drops sharply when certain modalities are unavailable or incomplete. Current research on missing modality primarily faces two dilemmas: Prompt-based methods struggle to restore missing yet indispensable features and degrade the generalizability of VLM. Imputation-based approaches, lacking effective guidance, are prone to generating semantically irrelevant noise. Restoring precise semantics while sustaining VLMs generalization remains challenging. Therefore, we propose a general missing modality restoration strategy in this paper. We introduce an enhanced diffusion model as a pluggable mid-stage training module to effectively restore missing features. Our strategy introduces two key innovations: (I) Dynamic Modality Gating, which adaptively leverages conditional features to guide the generation of semantically consistent features; (II) Cross-Modal Mutual Learning mechanism, which bridges the semantic spaces of the dual models to achieve bi-directional alignment. Notably, our strategy maintains the original integrity of the pre-trained VLM, requiring no fine-tuning of the backbone models while significantly boosting resilience to information loss. Zero-shot evaluations across benchmark datasets demonstrate that our approach consistently outperforms existing baselines, establishing it as a robust and scalable extension that ensures VLM reliability across diverse missing rates and conditions. Our code and models are publicly available at https://github.com/DaiW487/Missing-Modality-Diffusion.

Abstract:
\beginabstract While multimodal federated learning offers a decentralized alternative for image-text retrieval, existing methods face two intertwined issues: 1) a persistent modality gap in pretrained encoders that compromises cross-modal alignment under scarce, non-IID data, 2) the challenge of enabling domain-aware personalization without exposing private client representations. These two issues are closely coupled: unreliable cross-modal geometry hinders domain-aware transfer, while privacy constraints make such transfer harder to achieve. To address them, we propose FedMMDP, a federated multimodal domain-personalization framework that unifies modality-gap reduction, secure domain recognition, and centroid-guided personalized optimization. Unlike prior work, FedMMDP first pretrains a lightweight projector on public multimodal data to improve cross-modal alignment. It then performs secure domain recognition via securely aggregated Lloyd-style K-means, enabling the server-side centroid updates without accessing client representations. These centroids guide personalized local adaptation on heterogeneous clients. Evaluated on ImageNet-derived and IAPR TC-12 benchmarks using CLIP and SigLIP backbones, FedMMDP achieves the state-of-the-art results under the primary \textttpreproj setting, outperforming the second-best baseline by 20.0 R_\mathrmsum points on ImageNet+CLIP. The results support FedMMDP as an effective solution for privacy-preserving, domain-aware multimodal federated retrieval. The supporting code for this paper is available at \urlhttps://github.com/Toony254/FedMMDP. \endabstract

Abstract:
Low-light imaging suffers from coupled degradations of low illumination, motion blur, and defocus blur. Recently, cutting-edge solutions suggest employing a single encoder-decoder network for joint image restoration. In this paper, we provide theoretical analysis and experimental evidence to reveal the sequential inductive bias inherent in this joint enhance-and-deblur task. The enhance-then-deblur paradigm consistently outperforms its reverse counterpart across all compared state-of-the-art (SOTA) models. Motivated by this insight, this paper proposes a dual-domain synergistic network (DDSNet) that jointly restores low-light images with a single encoder-decoder network. The proposed model processes global and local features separately in the spatial and frequency domains during the encoding (illumination enhancement) and decoding (motion and defocus deblurring) stages. Critically, a novel frequency-domain alignment bridge (FAB) is introduced to better preserve the frequency features between the encoder and the decoder. Compared with SOTA methods, DDSNet achieves superior performance in terms of PSNR, SSIM, NRQM, and NIQE on the LOL-Blur and Real-LOL-Blur datasets, indicating the efficacy of exploring sequential inductive bias for joint low-light image enhancement and deblurring. \hrefhttps://github.com/simpleOnly1/DDSNetCode is publicly available.

Abstract:
Recent autonomous driving systems have increasingly shifted from modular pipelines to end-to-end learning paradigms. However, purely end-to-end approaches are prone to shortcut learning or representation collapse, often requiring auxiliary perception or prediction tasks to improve generalization. These auxiliary objectives typically rely on expensive perception annotations, limiting scalability to large-scale unlabeled data. Moreover, many existing designs are largely heuristic, lacking principled theoretical guidance. In this paper, we present S2WAM, a Self-Supervised World Action Model for autonomous driving that learns directly from sequential sensor observations without requiring perception labels. S2WAM learns a latent world representation by reconstructing global multimodal observations from local partial observations. Additionally, we provide a theoretical formulation of S2WAM as a variant of variational autoencoders, offering principled guidance for model design. As a self-supervised framework, S2WAM can be trained on large-scale unlabeled driving data and produces representations that are effective for downstream planning. Experimental results on the NAVSIM benchmark show that our S2WAM achieves competitive performance without perception supervision, while also demonstrating its capability in 4D scene reconstruction and generation. Code is released at \urlhttps://github.com/Say2L/S2WAM.

Abstract:
Image quality assessment (IQA) methods fall short in super-resolution (SR): Full-Reference metrics are impractical without ground-truth references, while No-Reference metrics reflect content rather than reconstruction quality. To address this, we propose DualRefIQA, leveraging large generative models for dual-reference SR-IQA. The method uses two references: the original low-resolution one to preserve content fidelity, and a generated high-quality one to guide perceptual quality assessment. To support this paradigm, we construct a Dual Reference SR Quality Assessment (DSRQ) dataset, a benchmark of over 170,000 triplets with human-annotated Mean Opinion Scores (MOS). Each triplet consists of an LR/HR pair and a generated reference that, augmented from the LR by large models, approximates the HR distribution to facilitate content-quality disentanglement. We design a dual-branch architecture to fully exploit the information provided by both references: a Fidelity Branch that captures local structural fidelity against the aligned LR input, and a Perceptual Branch that leverages cross-attention to assess global texture quality using the generated reference. The two branches focus on disentangled aspects of content and quality, respectively, enabling more accurate evaluation. Extensive experiments show DualRefIQA achieves state-of-the-art correlation with human subjective scores, outperforming existing Full/No-Reference metrics. Our website is available at: \urlhttps://yanshi42.github.io/MM26-DualRefIQA/.

Authors: Yurong Guo (Yanzhao Electric Power Laboratory of North China Electric Power University), He Yufei (Department of Electronic and Communication Engineering, North China Electric Power University), Yonghao Li (Department of Electronic and Communication Engineering, North China Electric Power University), Dongliang Chang (PRIS, School of Artificial Intelligence, Beijing University of Posts and Telecommunications), Ke Zhang (Yanzhao Electric Power Laboratory of North China Electric Power University), Zhanyu Ma (PRIS, School of Artificial Intelligence, Beijing University of Posts and Telecommunications)

Abstract:
Controllable infraredvisible image fusion aims to integrate complementary thermal and structural information with flexible region-aware modulation, producing fused images that adapt to diverse user requirements. However, existing methods typically rely on predefined discrete control conditions, leading to a sparse space that fails to support fine-grained modulation demands. To address this, we propose ConFusion, a novel framework that learns the continuous fusion space via Gaussian-conditioned spatial-aware modulation, enabling instance-level fine-grained controllable infrared and visible image fusion. ConFusion employs a dual-branch architecture to disentangle modality-invariant and modality-specific representations under joint reconstruction and text-guided semantic alignment. Gaussian-conditioned instance modulation variables coupled with Grounded SAM-based instance masks guide instance-level fine-grained modulation through the Mask-Guided Specific Feature Modulator, while the Text-Driven Invariant Feature Enhancer improves semantic consistency and enhances fusion. During inference, the multimodal large language model parses user intents into instance-level modulation variables to guide image fusion.Extensive experiments show that ConFusion achieves state-of-the-art performance across multiple metrics in fusion quality, while supporting fine-grained controllable image fusion. Our code is available at https://github.com/HeyufeiAnto/Confusion

Abstract:
We introduce Composed Video Moment Retrieval (C-VMR), a novel task that localizes temporal segments in untrimmed videos given a composed query of a reference image and a modification instruction. Unlike conventional VMR, which assumes self-contained textual queries, or Composed Video Retrieval (CoVR), which performs coarse-grained cross-video retrieval, C-VMR requires the model to infer an implicit target visual state by composing the reference and instruction, and localizes it within the same video. To support this task, we construct C-ActivityNet, the first benchmark dedicated to C-VMR, comprising 62.1K compositional query triplets across 13.5K videos. To address semantic compositional uncertainty in heterogeneous image-text fusion and temporal scale uncertainty across different edit types, we propose STREAM (Semantic Transformation and Routing-Enabled Adaptive Modulation), a Transformer-based end-to-end framework with two complementary modules. Transformation-Driven Representation Editing (TDRE) explicitly factorizes the multi-modal editing process into spatial (``where'') and channel (``what'') routing, effectively mitigating modality interference. Routing-Aware Granularity Modulation (RAGM) then leverages the routing priors produced by TDRE to adaptively calibrate temporal boundaries according to the semantic scope of each edit instruction. On C-ActivityNet, STREAM achieves 21.33 at R@1 (IoU=0.7), a +149.8% relative gain over the direct-transfer baseline (adapted Moment-DETR with naive composed-query fusion). We further validate generalization through zero-shot evaluation on Charades-C-VMR, a C-VMR evaluation set derived from Charades-STA, where STREAM yields a +25.5% relative gain over the strongest baseline at R@1 (IoU=0.7), demonstrating that structured compositional reasoning is essential for precise temporal localization. Code is available at: \hrefhttps://github.com/HuangJi1019/STREAMhttps://github.com/HuangJi1019/STREAM.

Abstract:
While the fusion of 4D millimeter-wave radar and camera holds significant potential for 3D object detection in autonomous driving, the extreme sparsity of radar points and the deficiency of depth information in images lead to severe cross-modal spatial misalignment. Existing two-stage methods might suffer from the irreversible loss of height information during view transformation, and the deficiency of semantic information within proposals during the refinement stage. To address these challenges, this paper proposes an elevation-aware fusion detection framework, designed to maximize the synergistic effects between 4D radars spatial perception and the camera's semantic richness. First, at the scene-level proposal generation stage, we propose a global fusion module integrated with multi-granularity height feature optimization. This module leverages Deformable Cross-Attention (DCA) to associate radar and image semantics across the global scene and effectively mitigates the height collapse issue during 3D-to-BEV transformation by aggregating elevation features at different granularities, thereby achieving high-precision initial proposal regression. Second, at the region-level feature refinement stage, we design a refinement module featuring selective semantic feature injection. By progressively mapping fine-grained image semantics into 3D bounding box grids, this module compensates for the semantic deficiency within proposals while preserving the spatial geometric structures of 4D radar, enabling precise secondary refinement of object boundaries. Experimental results on VoD and TJ4DRadSet benchmarks demonstrate that ElevC2F achieves state-of-the-art performance, reaching 62.29% and 41.50% 3D mAP respectively, and outperforming existing methods by significant margins of up to +1.41% and +3.53%. Our code and supplementary materials will be available at https://github.com/fancj2017/ElevC2F.

Abstract:
Large language models (LLMs) are increasingly deployed as purpose-specific agents to handle domain-specific tasks such as customer service and code generation. These agents are expected to comply with not only generic safety guardrails but also purpose-specific restrictions tailored to their designated roles. Such additional restrictions enlarge the attack surface, particularly to prompt injection (PI) attacks. To defend against such attacks, existing detection methods primarily rely on analyzing input-output patterns, yet yield limited effectiveness. To address this limitation, we turn to analyzing the hidden activation space and discover that LLMs inherently retain latent policy-violation (PV) concepts when prompted with requests beyond their designated purpose. Particularly, PV concepts capture the semantics of conflicts between user queries and predefined restrictions, implicitly reflecting LLMs' intrinsic awareness of recognizing policy violations. Building on this insight, we propose PVDetector, a training-free framework that detects PI attacks during LLM inference by measuring hidden-state alignment with PV concepts, which are derived offline from the contrastive pairs of policy-violating and policy-compliant prompts. Experiments across multiple LLMs and datasets show that PVDetector achieves <1% false negative rate with minimal auxiliary overhead, consistently outperforming state-of-the-art methods. Our code is available at \urlhttps://github.com/Claresigle/PVDetector.

Abstract:
Vision--language models can face asymmetric visual and textual shifts at deployment. These shifts expose a multimodal failure mode in which an unreliable branch remains overconfident, dominates fusion, and causes entropy-based test-time adaptation to sharpen an incorrect prediction. We model this behavior as doubly stochastic posterior mixing and cast adaptation as constrained de-mixing. Majorization-Guided Multimodal Test-Time Adaptation (MG-MTTA) freezes both encoders and updates only a lightweight fusion module. Running-anchor consistency estimates relative branch drift, while cross-modal conflict regulates modality dominance before entropy sharpening. The analysis gives sufficient conditions for entropy reduction to preserve the clean decision and an explicit threshold at which a biased modality reverses the fused ranking. Across visual, textual, and joint shifts, MG-MTTA improves ImageNet top-1 accuracy from 57.97% to 66.51% under textual shift and from 21.68% to 26.27% under joint shift, while reducing wrong-more-confident failures. The gains are concentrated under textual and joint shifts, where branch reliability is asymmetric. Project page: \urlhttps://mg-mtta.github.io/.

Abstract:
Few-shot whole slide image (WSI) classification remains highly challenging due to the extreme scarcity of annotated slides, the gigapixel resolution of WSIs, and the hierarchical semantic gap between global tissue architecture and local cellular morphology. Although recent vision-language and prompt-based methods have shown promise in low-data regimes, most of them rely primarily on slide-level alignment or static generic descriptions, making them inadequate for modeling the hierarchical diagnostic structure of pathology. In this paper, we present MIND, a framework for Multi-scale Integration with Nested Descriptions, for few-shot WSI classification. MIND is built upon a cognitively motivated hierarchical alignment paradigm that explicitly follows the coarse-to-fine diagnostic process of pathologists. Specifically, we first construct low- and high-magnification WSI bags and build Nested Descriptions from pathology prior knowledge, organizing them into three semantic levels: slide-level diagnostic patterns, entity-level tissue structures, and cell-level fine-grained morphological cues. These descriptions are encoded as semantic anchors and hierarchically aligned with multi-scale visual features. At the slide level, a global descriptor summarized from low-magnification features adaptively filters the entity anchors that are most relevant to the current WSI. At the entity level, the selected anchors are grounded onto low-magnification visual evidence via cross-attention, yielding slide-conditioned entity features. At the cell level, these grounded entity features further guide the aggregation of high-magnification patches together with cell-level semantic anchors, producing discriminative fine-grained features. Finally, MIND jointly optimizes slide-level prediction and cell-level prediction under an entity-grounding regularizer, resulting in robust multi-scale visual-semantic features for classification. Extensive experiments on CAMELYON16 and TCGA-NSCLC under 4-shot, 8-shot, and 16-shot settings demonstrate that MIND consistently outperforms strong MIL and prompt-based baselines. Code is available at \urlhttps://github.com/junjianli106/MIND.

Abstract:
Real-time text-driven joint audio-video avatar generation requires jointly synthesizing portrait video and speech with high fidelity and precise synchronization, yet existing audio-visual diffusion models remain too slow for interactive use and often degrade noticeably after aggressive acceleration. We present Hallo-Live, a streaming framework for joint audio-visual avatar generation that combines asynchronous dual-stream diffusion with human-centric preference-guided distillation. To reduce articulation lag in causal generation, we introduce \emphFuture-Expanding Attention, which allows each video block to access synchronous audio together with a short horizon of future phonetic cues. To mitigate the quality loss of few-step distillation, we further propose \emphHuman-Centric Preference-Guided DMD (HP-DMD), which reweights training samples using rewards from visual fidelity, speech naturalness, and audio-visual synchronization. On two NVIDIA H200 GPUs, Hallo-Live runs at 20.38 FPS with 0.94 seconds latency, yielding 16.0× higher throughput and 99.3× lower latency than the teacher model Ovi. Despite this speedup, it retains strong generation quality, reaching comparable VideoAlign overall score and Sync Confidence score while outperforming other accelerated baselines in the overall quality-efficiency trade-off. Qualitative results further show robust generalization across photorealistic, multi-speaker, and stylized scenarios. To the best of our knowledge, Hallo-Live is the first framework to combine streaming dual-stream diffusion with preference-guided distillation for real-time, text-driven audio-visual generation. Code and models are publicly available at \urlhttps://github.com/fudan-generative-vision/Hallo-Live.

Abstract:
Multimodal Large Language Models (MLLMs) have shown remarkable capabilities, yet they are prone to visual hallucinations: generating text that contradicts visual evidence. A prominent mitigation method is Contrastive Decoding (CD), which enhances visual guidance by penalizing logits derived from a visual-degraded context. Nevertheless, this approach is fundamentally limited as it relies only on the final output logits, ignoring intermediate layers' visual patterns which are crucial for alleviating hallucinations. To delve into this issue, we conduct a diagnostic dissection to trace the visual patterns within the intermediate layers. We quantify visual guidance strength at each layer via measuring the divergence between the predicted distributions from a full multimodal context and those from a visual-degraded context. Our analysis yields a critical insight: the visual guidance strength is adaptive across layers. In particular, the strength peaks in the middle layers and is surprisingly attenuated or even absent at the final layer. With such findings, we propose Layer-Adaptive VIsual Guidance (LAVIG), a novel training-free decoding algorithm to mitigate hallucinations in MLLMs. LAVIG directly adopts the layerwise visual guidance strength to adaptively re-weight visual tokens for the subsequent layer's self-attention computation. As a model-agnostic strategy, LAVIG can be seamlessly integrated into various MLLMs. Comprehensive experiments demonstrate that LAVIG sets a new state-of-the-art on major hallucination benchmarks, evidently reducing factual errors with marginal computational overhead. Code is available at: https://github.com/HiDream-ai/LAVIG/.

Abstract:
Emotion recognition is a fundamental task in affective computing, and most current works consider multimodal fusion to integrate comprehensive information for better performance. Among the modalities, physiological signals present a specialty, as on the one side, they are hard to alter, providing more objective evidence for emotion understanding; on the other side, they require delicate equipment to be attached to the body, which is inconvenient and interfering for the monitored persons. Existing multimodal emotion research concerns more of non-physiological modalities, such as video, audio, and text, whereas physiological signals remain insufficiently explored. To bridge this gap, we introduce a new multimodal emotion dataset with remotely sensed physiological signals by millimeter-wave radar and facial videos. The new dataset is targeted to facilitate two research fields: 1) multimodal remote physiological signal measurement, and 2) physiological-behavioral fusion for emotion recognition with only remote sensors. Furthermore, beyond clip-level emotion annotations, we also provide fine-grained intensity labels, enabling more detailed temporal emotion analysis. Extensive benchmarks were established on the new dataset, including uni- and multimodal approaches for remote physiological signal measurement and emotion recognition. It is expected that the dataset and benchmarks provide a new testbed and open up new possibilities for future research. The dataset is available at: \urlhttps://github.com/helloworld-YY/RemoPhysEmo. \endabstract

Abstract:
\beginabstract Recent large audio language models (LALMs) process extended multimodal sequences but incur high inference cost. Token compression can reduce this cost by pruning redundant tokens. Existing methods typically average token-importance scores over all attention heads, implicitly assuming that heads contribute equally across tasks. Our analysis instead reveals distinct head behaviors across audio domains: only a sparse subset is strongly audio-selective, and the useful heads differ between semantic and acoustic tasks. We therefore propose \ours, a training-free, head-aware pruning method that dynamically routes head weights using the input's semantic--acoustic characteristics, preserving task-relevant tokens under aggressive compression. \ours applies to diverse LALMs. Experiments on AudioMarathon and MMAU-Pro demonstrate state-of-the-art (SOTA) compression performance. At 60% audio-token pruning, \ours retains 99.4% and 97.6% of the unpruned performance on Qwen2.5-Omni-3B and Qwen2.5-Omni-7B, respectively. \endabstract

Abstract:
Audio-driven 3D facial animation aims to synthesize realistic and temporally coherent motions from speech. Despite notable progress in lip synchronization, weakly correlated dynamics, including eyebrow movements, eye blinks, and head motion, which are essential to photorealistic facial animation, remain difficult to model faithfully and often appear static or unnaturally repetitive. We attribute this limitation to three factors: (a) insufficient conditioning for weakly correlated dynamics; (b) the limited ability of deterministic regression to capture diverse motion patterns; (c) data bottlenecks from unreliable upper-face pseudo-labels and limited dataset diversity. To address these issues, we propose SubtleTalk, a framework for generating natural and controllable weakly correlated facial dynamics via multi-condition modeling and residual flow matching. First, to compensate for the limited guidance of speech alone, we introduce interpretable controls, including prosody, regional intensity, and Valence-Arousal (VA) signals, to explicitly capture the timing, magnitude, and affective variation of weakly correlated dynamics. Second, to overcome the limited expressiveness of deterministic regression, we build residual flow matching based on a stable speech-driven motion prior, allowing the model to capture stochastic deviations beyond deterministic prediction. Third, to alleviate the data bottleneck, we construct SubtleTalk-Face, a large-scale 3D facial animation dataset comprising about 3,900 identities and 74 hours of data, built via a simple and scalable pseudo-labeling pipeline and featuring improved upper-face tracking and frame-level VA annotations. Extensive experiments demonstrate that our method significantly improves the realism and diversity of weakly correlated facial dynamics while preserving accurate lip synchronization. Our project page is available at \urlhttps://molly-ding.github.io/SubtleTalk/.

Authors: Khaled Abud (MSU Institute for Artificial Intelligence), Aleksey Yakushev (Trusted AI Research Center RAS), Aleksandr Akimenkov (Trusted AI Research Center RAS), Irina Serzhenko (Independent researcher), Kirill Aistov (MSU Institute for Artificial Intelligence), Egor Kovalev (MSU Institute for Artificial Intelligence), Dmitry Obydenkov (Trusted AI Research Center RAS), Sergey Lavrushkin (MSU Institute for Artificial Intelligence), Anastasia Antsiferova (Trusted AI Research Center RAS), Dmitriy Vatolin (MSU Institute for Artificial Intelligence), Yury Markin (Trusted AI Research Center RAS), Kirill Lukyanov (Trusted AI Research Center RAS)

Abstract:
Digital image watermarking is increasingly critical in media contexts, as emerging regulations and industry practices require marking AI-generated content and ensuring traceable sources to prevent manipulation or misuse. Recent advances in invisible watermarking methods highlight the need to update existing benchmarking practices to reflect current techniques and evaluation criteria. We address this by introducing WARP a unified framework and benchmark for evaluating the robustness of invisible watermarks. WARP incorporates 32 recent classical, deep, and generative watermarking methods, as well as 34 different erasing techniques, ranging from traditional distortions to more sophisticated adversarial, purification, and re-embedding attacks. It provides standardized, reproducible, and easily scalable protocols for evaluating perceptual quality, watermark readability, and attack resilience. Using WARP, we extensively evaluate current invisible watermarking techniques, collecting the largest robustness benchmark in the field. Results identify the most robust approaches under both distortion and adversarial conditions, and reveal consistent relationships between watermarking methods and the attack strategies most effective against them. Our experiments also highlight that some of the watermarking methods considered are highly vulnerable to reembedding, even if they are robust to standard distortions. The code is made available at \urlhttps://github.com/ispras/wibe.

Abstract:
Self-supervised hyperspectral and multispectral image (HSI-MSI) fusion aims to reconstruct high-resolution (HR)-HSIs without ground-truth data. However, existing methods struggle with severe artifacts due to complex real-world degradation, including spatially-varying displacement and spatially-varying blur. We propose SRD-Rectify, a novel self-supervised fusion algorithm that leverages a pre-trained RGB Super-Resolution Diffusion model as a robust Rectifying prior to address these challenges. Unlike HSI-specific priors, which are limited by data scarcity, this diffusion model, trained on large-scale low-resolution (LR)-HR RGB datasets, learns a universal prior that is robust to severe degradation. To bridge the modality gap, we refine a spectral autoencoder to project HSIs into the representation space of the diffusion model. Inference is carried out through a guided diffusion process, where the diffusion model corrects degradation in the LR-HSI and generates high-frequency details, while the HR-MSI acts as a likelihood constraint to maintain the spatial consistency. This process effectively rectifies displacement and blur while preserving the spectral and spatial structures from the LR-HSI and HR-MSI during fusion. Extensive experiments show that SRD-Rectify achieves state-of-the-art fidelity under complex degradation. The code is available at https://github.com/Dmsw/SRD-Rectify.

Abstract:
UAV vision-language navigation (UAV-VLN) focuses on enabling an aerial agent to follow natural-language instructions in open 3D environments from egocentric visual observations. Current approaches suffer from three coupled issues: weak grounding of instruction-relevant landmarks in visual observations, insufficient exploration of long-horizon history, and unstable decisions under local traps or repeated exploration. To address these issues, we propose a unified semantic-to-decision framework. First, we present an instruction-grounded semantic enhancement module that injects object-level semantics and relative spatial cues into the current observation state. Subsequently, we develop a relevance-aware dynamic temporal aggregation strategy that reweights the full history buffer while converting a few high-relevance frames into structured landmark prompts for the decoder. Finally, we devise a topology-aware decision method that combines local-optimum cognition with group-relative policy optimization under progress, goal, semantic, and path-compliance rewards. Experiments on the widely used AerialVLN and OpenFly benchmarks clearly demonstrate that our method achieves state-of-the-art performance. Our code is available at \urlhttps://github.com/mksasx/S2D-UAV-VLN.

Abstract:
Camera relocalization currently struggles with a dilemma between accuracy and adaptability: scene-dependent methods offer high precision but fail to generalize, while scene-independent alternatives require extensive training data and often suffer from suboptimal accuracy. In this work, we aim to bridge this gap by achieving both high accuracy and strong generalization. Our key idea is to base localization on transferable feature-level invariants rather than image appearance, enabling accurate relocalization in unseen scenes even when trained on a single scene. We propose FASIReloc, a feature-anchored scene-independent relocalization framework built on feature Gaussian primitives that encode transferable 3D structures. FASIReloc adopts a coarse-to-fine pipeline: a saliency-driven sampling strategy selects stable and representative Gaussian anchors, and a lightweight scene-independent network establishes reliable 2D-3D correspondences for initial pose estimation. The pose is then refined through dense feature alignment for improved accuracy. Experiments on 7 Scenes and Cambridge Landmarks demonstrate state-of-the-art performance in both scene-dependent and scene-independent settings. In the scene-independent setting, FASIReloc reduces median translation and rotation errors by 51% and 73% on 7 Scenes, and 45% and 45% on Cambridge Landmarks, compared to prior best methods. Notably, these results are achieved when training on only a single scene, highlighting strong data efficiency and cross-scene generalization. The code is available at: \hrefhttps://github.com/junhaowei0524/FASIRelochttps://github.com/junhaowei0524/FASIReloc.

Abstract:
Target speaker extraction (TSE) can be guided by heterogeneous cues such as video, enrollment speech, text, spatial information. However, current multi-cue evaluation remains fragmented across cue settings, preprocessing pipelines, and fusion interfaces. To enable fair and systematic comparison, we establish UniCue, a unified benchmark for multi-cue TSE. UniCue provides a speech-in-video mixture dataset with paired cues from four modalities and speaker-disjoint splits, standardizes cue construction and preprocessing, and introduces modality-specific stress settings that explicitly probe cue availability and cue reliability. To reduce interface bias, we instantiate UniCueNet, a shared reference conditioning interface on a SepFormer-style separator backbone. The interface projects heterogeneous cues into a common latent space, organizes them into a unified cue memory, conditions the separator through mixture-guided attention over the cue memory, and injects cue information via layer-wise gated injection. We additionally validate the interface on a second separator family and observe that, although absolute performance varies across carriers, the main subset-wise trends remain consistent. Experiments over all cue subsets and selected stress conditions reveal clear patterns of cue usefulness, cross-cue complementarity, and robustness, establishing a reproducible testbed for multi-cue TSE research. Code and data will be released. Demo: https://uni-cue.github.io/UniCue-demo/

Abstract:
High-fidelity general audio coding at ultra-low bitrates is crucial for applications ranging from low-bandwidth communication to audio-language modeling. However, existing neural audio codecs struggle under extreme bitrate constraints, exhibiting degraded performance: weak semantic alignment leads to semantic distortion, while limited generative capacity compromises acoustic fidelity. To address these challenges, we propose TeleSound?Codec, a scalable general audio codec that preserves semantic integrity and fine acoustic details together via language?grounded supervision and scalable generative modeling. Specifically, to achieve strong semantic alignment, we propose to train a semantic encoder under the guidance of a pre-trained large language model. This language-grounded supervision grounds codec latents in human-interpretable linguistic spaces, enhancing semantic integrity and facilitating downstream applications. To improve generative capacity, we further propose training a flow-matching-based generative decoder jointly with an acoustic encoder that captures residual acoustic details. This scalable generative modeling design enables reconstruction quality to scale with computation. Experiments demonstrate that, when scaling the model size to 1.8B parameters, TeleSound-Codec achieves high-fidelity reconstruction of 32kHz general audioincluding speech, music, and other sound typesat bitrates as low as 0.275kbps. We present demos on: https://aisaka0v0.github.io/codec\_demo/.

Abstract:
\beginabstract Reasoning in Multimodal Large Language Models (MLLMs) requires both fine-grained visual perception and rigorous logical deduction. Explicit text-based Chain-of-Thought (CoT) is computationally expensive and prone to visual hallucinations, while existing latent reasoning methods typically require costly training. Furthermore, directly adapting training-free LLM reasoning mechanisms to the multimodal setting yields unstable performance. We identify that this failure stems from their reliance on token-level entropy, which fundamentally conflates perceptual ambiguity (e.g., unclear visual details) with logical uncertainty (e.g., complex reasoning steps). To overcome this bottleneck, we present a novel training-free inference strategy for MLLMs that explicitly decouples perception and reasoning. We propose a novel metric, the vision-to-text attention ratio, to dynamically gauge the model's cognitive focus. Guided by this metric, our proposed framework, Attention-Guided Switching (\methodname), adaptively triggers latent reasoning for perceptual tokens to preserve high-fidelity visual information in the continuous space, while enforcing explicit text generation for logical tokens to maintain structural anchoring. Extensive experiments demonstrate that our method achieves state-of-the-art performance, significantly improving both accuracy and inference efficiency by reducing autoregressive steps and latency. Code is released at \urlhttps://github.com/swordAndSnow/MM26-AGS. \endabstract

Abstract:
Multi-view anomaly detection is crucial for industrial inspection to identify defects hidden from single viewpoints. Yet, despite their strong success in single-view applications, reconstruction-based methods remain largely underexplored in multi-view settings. A key challenge in extending these methods lies in generating view-consistent pseudo-anomalies. To address this, we propose PAR, a reconstruction-based framework for enhanced multi-view anomaly detection. First, we introduce a feature perturbation strategy coupled with a consensus aggregation module that integrates these perturbed features across viewpoints to mitigate view-inconsistent anomalies and construct highly aligned pseudo-anomaly representations. Second, we employ a view-specialized multi-decoder architecture; by assigning a dedicated decoder to each view, we increase the model's representational complexity, enabling the generation of highly feasible, localized anomaly maps rather than less precise, view-agnostic reconstructions. Last, we incorporate Fourier attention into the decoders to emphasize global frequency-domain structures over local token evidence, effectively preventing identity mapping. Extensive experiments on the Real-IAD dataset demonstrate that PAR consistently outperforms state-of-the-art methods in sample-level, image-level, and pixel-level anomaly detection. Our code is available at https://github.com/SpatialAILab/PAR_mvad.git

Abstract:
Light field images (LFIs) capture rich angular and spatial information, enabling immersive scene visualization. However, their substantial storage requirements necessitate compression, which inevitably introduces perceptual distortions. Despite the emergence of numerous advanced compression methods, existing quality assessment frameworks for LFIs remain narrow in scope, leaving the perceptual impact of diverse compression distortions largely unexplored. To address this, we conduct subjective experiments and present a new perceptual quality assessment database comprising 1000 distorted LFIs from 20 scenes, generated using 10 compression methods across five quality levels, with mean opinion scores collected from 29 qualified observers under controlled conditions. Building on this database, we propose a novel objective blind metric based on multi-modal and multi-task learning. The metric extracts 2D and 3D multi-modal patches from 4D LFIs, feeding them into image-domain and video-domain branches for preliminary quality assessment. A distortion-aware multi-task learning module then integrates cross-modal features to produce the final quality score. Extensive experiments on the proposed database demonstrate that the proposed metric outperforms representative quality assessment metrics in various aspects. The database and metric code are available at https://github.com/oldblackfish/HUST-LFI.

Abstract:
Despite the remarkable photorealism achieved by large-scale text-to-image diffusion models, they exhibit a persistent failure mode known as Prior Attribute Bias. When prompted with counter-intuitive concepts that contradict training data statistics (e.g., ``a gray strawberry'' or ``a square apple''), models frequently ignore the explicit textual condition in favor of the implicit canonical prior. Existing approaches attempt to mitigate this through two strategies: enforcing spatial alignment via computationally expensive test-time optimization, or manipulating text embeddings. Both remain ineffective against strong structural priors. In this work, we identify that a core limitation lies in the semantic representation: the cross-attention Value projection acts as a ``prior filter'', embedding canonical traits directly into the object's feature representation. Consequently, even with perfect spatial alignment, the object's inherent features conflict with the target attribute. To address this, we introduce LinearV, a training-free framework that decouples attribute binding into two complementary processes: (1) GenPointer enforces spatial alignment by injecting the object's positional information into the attribute's attention mechanism; and (2) VDebias resolves the semantic conflict via targeted semantic steering in the Value Space, explicitly neutralizing the canonical prior. To rigorously evaluate this problem, we further introduce the Counter-Intuitive Attribute Benchmark (CIAB), a curated set of 240 prompts spanning color, shape, and texture. Extensive experiments across multiple diffusion backbones demonstrate that LinearV significantly outperforms state-of-the-art methods. Notably, LinearV achieves a relative gain of over 11.8% in target alignment, measured by Disentangled BLIP-VQA, over the strongest baselines in challenging structural categories, confirming its effectiveness in deep semantic modifications. The CIAB benchmark will be available at \hrefhttps://github.com/Taidvt/LinearVhttps://github.com/Taidvt/LinearV.

Abstract:
A fundamental intent asymmetry plagues modern 3D asset creation: while state-of-the-art 3D toolchains demand precise, executable parameters, ordinary users typically provide vague, underspecified instructions. Current 3D agents treat this ambiguity as noise, defaulting to blind execution under a single-turn assumption. To address this limitation, we introduce \textscClare, a clarification-aware and evolutionary 3D agent that treats intent asymmetry not as an execution error, but as an opportunity for strategic dialogue. By decoupling the generation pipeline into four specialized cognitive roles, \textscClare intercepts and resolves underspecified instructions before invoking computationally expensive 3D tools to seamlessly execute tasks across five diverse domains: text-to-3D generation, single-view reconstruction, multi-view reconstruction, point cloud editing, and post-processing. Crucially, rather than relying on rigid manual rules, \textscClare self-evolves its clarification policy via simulated multi-turn interactions. By optimizing a Multi-turn Reward, the agent internalizes the delicate balance between interaction efficiency and task completion. To rigorously test this, we construct 3D-Clarify, a comprehensive benchmark comprising 620 interaction scenarios with systematically injected ambiguity, missing information, and mistaken details. \textscClare achieves state-of-the-art performance, with 60.40% and 43.34% success rates on single-step and multi-step tasks, respectively, more than doubling existing baselines. Both quantitative and qualitative results demonstrate that proactive clarification is the missing key to robust 3D execution. Code is available at \urlhttps://github.com/xyzhu1225/CLARE.

Abstract:
Histopathological examination primarily relies on hematoxylin and eosin (H\&E) and immunohistochemistry (IHC) staining. Although IHC provides critical molecular information, it is costly and requires specialized expertise. Stain transfer provides an efficient alternative by computationally generating IHC from H\&E images, but remains challenged by unified and interpretable modeling for heterogeneous biomarkers under pixel-unaligned supervision. We propose DMCoStain, a novel Data-Model Co-optimization framework for Stain transfer. It iteratively co-refines training data and model capability, improving staining accuracy and interpretability in both pathological and structural consistency. To refine training data in a clinically meaningful manner, it incorporates the Multimodal Expert-Guided Finer Selection (MEGFS) strategy, built upon a pioneering IHC-positive-expression (IPE) vision-language model (VLM) that emulates pathologist reasoning. To support MEGFS, we construct ImmunoInstruction, the first large-scale IPE instruction-following dataset with 150K VQA samples. Extensive experiments on multiple tissues and biomarkers demonstrate that DMCoStain achieves state-of-the-art (SOTA) accuracy. This paradigm offers strong practical value, and MEGFS also functions as a specialized evaluation tool for future model development. Dataset, code, and more details are in https://github.com/SikangSHU/DMCoStain.

Abstract:
Learning effective multimodal entity representations is fundamental for reasoning tasks such as multimodal knowledge graph completion (MMKGC). However, existing methods often suffer from semantic over-smoothing within modalities and ineffective noise filtration across modalities, particularly under sparse or ambiguous conditions. To overcome these limitations, we propose PrismF, a unified framework that synergizes multi-perspective enhancement with progressive fusion to extract stronger signals from diverse inputs. PrismF enhances fine-grained intra-modal semantics through a multi-perspective mechanism that decomposes each modality into complementary views and constrains them with a decoupling loss to reduce representation collapse. Furthermore, it improves cross-modal integration through a progressive fusion strategy that dynamically calibrates inter-modal interactions, enabling the model to emphasize informative signals while suppressing noisy or unreliable ones. Extensive experiments on three public benchmarks show that PrismF achieves the strongest overall performance, including relative improvements of 4.04% in MRR and 11.17% in Hits@1 on KVC16K. Our code can be found at https://github.com/HubuKG/PrismF.

Abstract:
Unsupervised low-light image enhancement remains a challenging ill-posed problem, primarily due to the difficulty in balancing physical illumination recovery and semantic content preservation. Existing physical model-based methods often suffer from color deviation due to their ``semantic blindness'', while purely data-driven approaches lack interpretability and robust physical constraints. To bridge this gap, we propose a novel Semantic-Physical Synergistic Network (SPS-Net), a dual-prior driven framework that explicitly integrates physical laws with high-level semantic knowledge. Specifically, we first introduce a Physics-Aware Fusion Module (PAFM), which encodes Retinex-based illumination and reflectance priors into robust physical features via parallel channel and spatial attention mechanisms. To inject semantic understanding, we design a SemanticPriorNet leveraging a pre-trained CLIP model to extract deep, robust semantic representations. Central to our restoration process is the Dual-Guided Reflection Branch (DGRB), which progressively reconstructs the reflectance map using two task-specific attention mechanisms: Illumination-Guided Attention (IGA) to adaptively suppress noise based on light intensity, and Semantic-Guided Attention (SGA) to rectify color bias and recover semantically plausible details under semantic guidance. Furthermore, we propose a tailored CLIP-Guided Loss, incorporating both semantic consistency and prompt-based quality constraints, to supervise the network in an unsupervised manner. Extensive experiments on multiple public datasets demonstrate that SPS-Net outperforms state-of-the-art unsupervised methods, achieving superior visual quality and semantic fidelity. The source code and pre-trained models are publicly available at https://github.com/gyz0224/SPS-Net.

Abstract:
Estimating dish calories solely from visual data is an ill-posed problem due to the absence of scale references and hidden ingredients. While Multimodal Large Language Models (MLLMs) offer rich semantic understanding, they are prone to hallucinations and lack the reliability required for safety-critical dietary assessment. To address this, we introduce MARC (Multimodal Anchored Residual Correction), a framework that reformulates calorie estimation as a conditional residual prediction task. Rather than predicting absolute values directly, MARC grounds its inference on a retrieval-based USDA anchor. Specifically, we model the refinement process as a contextual routing problem where a policy network adaptively selects the most effective strategy, ranging from visual scaling to compositional rectification, to predict the precise residual term. To enhance reliability, we incorporate a lightweight gated calibration mechanism to mitigate overconfidence and employ conformal prediction for rigorous uncertainty quantification. Furthermore, we release the CCM, a large-scale Chinese Cuisine Multimodal dataset containing over 32,000 dishes, including an expert-verified gold-standard subset. Extensive experiments on CCM and the public Nutrition2.7K dataset demonstrate that MARC achieves state-of-the-art performance, significantly outperforming existing visual regression models and general-purpose MLLMs while offering valid confidence intervals for trustworthy dietary assessment. Our code is available at \hrefhttps://github.com/Agentyzu/MARChttps://github.com/Agentyzu/MARC.

Abstract:
Long-horizon egocentric video is a rich substrate for wearable AI assistants, but object-centric questions such as where an item was moved, when it last changed state, or why it was relocated remain difficult because caption- and transcript-based memories rarely preserve persistent object identity or structured spatial change. Existing long-video QA methods mainly emphasize temporal grounding and clip retrieval, while prior 3D scene-graph methods typically assume stronger geometry than free-motion wearable RGB video provides, including point clouds, RGB-D input, posed views, sparse reconstruction, or reconstructed scenes. R4DSG introduces a relative 4D scene graph memory for long egocentric video. Instead of storing raw graph sequences, R4DSG converts video into compact queryable memory entries indexed by time, place, persistent objects, anchor-relative change, and local interaction context. The main idea is to separate stable anchors from dynamic objects, maintain persistent object identity across frames, and represent object state through anchor-relative transitions rather than a globally aligned world model. Built on recent RGB-only advances in promptable video segmentation, temporal propagation, and relative 3D lifting, the method produces a retrieval-ready memory directly usable for long-horizon question answering. Evaluation on a 255-question object-related subset from EgoLifeQA shows, under question-only retrieval, a 6.7-point overall gain over EgoRAG-Text and a 12.5-point gain on when questions, which highlights the value of temporally organized object memory. These results position relative 4D scene graphs as a practical memory substrate for wearable assistants, AR systems, and embodied multimedia agents. GitHub Page: https://dualtransparency.github.io/R4DSG/.

Abstract:
Recent advances in AI-Generated Image (AIGI) detection have largely focused on model architectures, loss design, and training strategies, while overlooking a more fundamental factor: how supervision is structured. Existing detectors typically train on benchmark-defined data that implicitly blends semantic and artifact signals, leaving supervision design unexplored. In this paper, we examine supervision structure through controlled experiments and show that such blended supervision is not merely suboptimal but fundamentally counterproductive. Replacing it with disentangled, orthogonal supervisionpaired with inductively matched modelssuffices to achieve strong cross-generator generalization. We formalize this as Orthogonal Supervision Disentanglement and instantiate it in AlignGemini, a two-branch detector in which a VLM receives purely semantic supervision and a vision expert receives purely artifact supervision. With disentangled supervision coupled with inductively matched models, even the simplest training recipe achieves a significant improvement of 9.5% across five in-the-wild benchmarkstrained on images from only two generators, without complex training strategies or diverse generator coverage. We further introduce AIGI-Now, a diagnostic benchmark covering eight contemporary generators, including five closed-source commercial models, with disentangled semantic and pixel evaluation subsets for fine-grained analysis of detector capabilities. Our code is publicly available at https://github.com/GzZning/Align-Don-t-Blend.

Abstract:
Vision-Language Models (VLMs) are the perceptual backbone of embodied AI, but their energy footprint on edge hardware remains poorly understood. Existing efficiency efforts focus predominantly on reducing visual tokens, implicitly treating visual processing as the dominant energy cost. We overturn this implicit assumption through the first systematic energy profiling of on-device VLM inference, spanning five models across three architecture families, four input resolutions, and two hardware platforms (NVIDIA RTX 3070 and Jetson Orin NX). Our analysis yields three findings. First, average inference power is a model-intrinsic constant, invariant to input resolution, image complexity, and prompt type, with less than 5% variation across all conditions. This means that all energy variation across inputs must arise from variation in inference time, not from variation in power draw. Second, each output token costs 11 to 39 more wall-clock time than each input token due to the compute-bound and memory-bound asymmetry between prefill and decode, making output token count the dominant driver of both latency and energy. Third, image complexity, measured by the number of objects in an image, induces up to 4.1 energy differences at identical resolution. This variation arises not from increased visual processing cost, but from differences in output length. These findings expose a fundamental limitation of visual token pruning: even removing all visual tokens saves at most 10% of total energy for fixed-token models. Across models spanning 1 billion to 8 billion parameters, controlling output length saves up to 97% of total energy, with the energy dominance of decoding growing stronger at larger model scale. In short, the true energy bottleneck in edge VLM inference is not what the model sees, but how much it says. Code is available at https://github.com/Junfei-Z/seeing-is-free.

Abstract:
Implicit emotion understanding addresses a fundamental yet underexplored challenge: observable expressions do not always faithfully reflect underlying emotional states, requiring models to infer latent emotions from limited multimodal signals. Existing approaches largely assume direct correspondence between overt cues and true emotions, while alternatives often rely on physiological signals or extensive context, limiting practicality. To address these limitations, we present (i) a new task formulation under real-world constraints; (ii) a benchmark built on the Multimodal Implicit Emotion (MIE) dataset and a unified evaluation protocol; and (iii) EmotionReasoner, an LLM-based framework for reasoning beyond surface cues. MIE covers video, audio, and text, with annotations for both explicit and implicit emotions. A context-grounded verification mechanism reduces labeling ambiguity and improves annotation reliability. The benchmark evaluates explicit and implicit emotions through grouped metrics and fine-grained category-level analysis. EmotionReasoner performs hierarchical reasoning from global perception and local anomaly detection to structured evidence organization and implicit emotion inference. Experiments show that current multimodal large language models exhibit systematic deficiencies in implicit emotion understanding, whereas EmotionReasoner substantially improves performance and robustness. The appendix and source code are available at \urlhttps://github.com/WKH186/MIE.

Abstract:
Source-Free Video Domain Adaptation (SFVDA) is critical for cross-domain action recognition when source data is inaccessible. Existing methods primarily focus on prediction-level pseudo-label denoising, but universally rely on random clip sampling, which tends to capture domain-specific backgrounds and thus amplifies distribution discrepancy at the input level. In this paper, we shift the focus to input-level sampling and reveal that motion-saliency-driven keyframe sampling serves as an implicit domain alignment mechanism. Based on this insight, we construct a Motion Saliency-Aware (MSA) source model by embedding a confidence-guided keyframe sampling strategy into source training, enabling the model to learn more transferable and action-discriminative representations. However, during target-domain deployment, domain discrepancy introduces sampling bias, which degrades pseudo-label quality. To address this issue, we propose a Sharpness-Guided Adaptation (SGA) framework and introduce keyframe sharpness as a core metric. By leveraging keyframe sharpness, the model can perceive domain discrepancy under source-free conditions and adaptively calibrate the self-training objective, thereby suppressing erroneous predictions caused by domain shift. Furthermore, we extend keyframe sharpness to multimodal scenarios to dynamically modulate alignment strength, alleviating the temporal misalignment between discriminative peaks of RGB and optical flow and enabling more robust cross-modal feature alignment. Extensive experiments on UCF-HMDB and EPIC-Kitchens demonstrate that our method achieves state-of-the-art (SOTA) performance across multiple SFVDA benchmarks. Our code is available at https://github.com/SWUFE-XYW/MSA-SGA-SFVDA.

Abstract:
Video generation and editing have made remarkable strides with the advent of Diffusion Transformer (DiT) architectures and unified frameworks such as VACE. However, preserving a character's holistic identity, including facial features, clothing, and body characteristics, across diverse video synthesis tasks remains a significant open challenge. Existing identity-preserving approaches rely predominantly on face recognition embeddings, suffering from the "floating head" artifact under extreme dynamics and failing to capture full-body persona attributes. Moreover, they are tailored to either generation or editing, lacking the versatility to operate within a unified framework. We present OmniPersona, a lightweight, plug-and-play framework built atop the VACE Video Condition Unit paradigm that achieves holistic persona preservation across all video creation and editing tasks. Our approach introduces three core innovations: (1) Multimodal Holistic Persona Representation (MHPR), a dual-branch encoder fusing frequency-decomposed facial structure features with semantic body-clothing features into unified Persona Tokens; (2) Persona-Conditioned Spatiotemporal Decoupled Context Adapter (PST-CA), which restructures the Context Adapter by explicitly decoupling spatial appearance injection from temporal motion control via task-aware adaptive gating; and (3) Motion-Aware Persona Propagation (MAPP), which warps cached identity features along estimated motion trajectories to maintain temporal coherence during large displacements. Extensive experiments demonstrate that OmniPersona significantly outperforms state-of-the-art methods in identity preservation, motion naturalness, and temporal consistency across generation, editing, and composition tasks, while introducing only approximately 1.2% additional trainable parameters. Code: https://github.com/YMlinfeng/OmniPersona.

Abstract:
Automatic music transcription (AMT) aims to convert audio recordings into symbolic note representations. Existing methods model notes as discrete activations on a quantised grid and supervise offset with single-frame point labels, leaving offset predictions poorly anchored to the acoustic evidence of note duration. We propose GaussNote, a differentiable note refinement model based on Gaussian ellipsoid superposition, to couple note duration directly to the spectrogram energy it produces. Given the coarse predictions of any pretrained AMT system, GaussNote lifts each note into a continuous Gaussian ellipsoid on the time-frequency plane, parameterised by a centroid encoding pitch and onset, axis-aligned spreads encoding duration and spectral width, and a rotation angle encoding time-frequency coupling. We design a note feature encoder to capture complementary acoustic, geometric, and contextual information for each note from the local CQT patch, the initial ellipsoid parameters, and note-level context. We introduce a Gaussian parameter adapter block (GPAB) to address asymmetric note parameter correctability via parameter-wise gated residual updates. A differentiable renderer maps the refined ellipsoids back to the time-frequency plane, providing dense supervision over every note's full temporal extent. We provide theoretical guarantees on the optimality of the Gaussian representation, the perceptual validity of the reconstruction loss, and the convergence of the optimisation. Experiments on the MAPS benchmark demonstrate state-of-the-art note-with-offset transcription performance, with consistent improvements in offset accuracy over existing methods. Code is available at \urlhttps://github.com/jyatgithub/Gaussnote

Abstract:
\beginabstract Recent years have witnessed a burgeoning interest in 3D anomaly detection. However, reconstruction-based approaches are vulnerable to the identity shortcut, which narrows the reconstruction-error gap between normal and anomalous inputs and degrades detection performance. This issue gives rise to a critical dilemma. Faithfully reconstructing normal geometry from latent features requires both sufficient fine-grained local detail and effective suppression of anomalous cues. Aggressive latent compression suppresses anomalous cues but sacrifices local geometry, whereas U-Net skip connections preserve such geometry but may also transmit anomalous cues and reintroduce the identity shortcut. To address this trade-off, we propose ProtoSkip, which uses Sinkhorn-based optimal transport to purify skip features into balanced prototypes, preserving geometric detail while suppressing anomalous cues and mitigating the identity shortcut. To facilitate anomaly repair, we employ a Point Cluster Graph Convolution (PCGC) encoder that aggregates surrounding geometric context without relying on over-focused attention. When combined with PCGC, ProtoSkip achieves remarkable performance. On the Real3D-AD dataset, our method outperforms the MC3D-AD baseline by 3.9 points in object-level AUROC and 8.4 points in point-level AUROC. On Anomaly-ShapeNet, ProtoSkip achieves a competitive object-level AUROC of 85.4. Code is available at \urlhttps://github.com/tripmasterwang/Protoskip.git. \endabstract

Abstract:
Fine-grained cross-modal understanding in drone views is essential for aerial vision-language navigation. However, the inherent wide field of view and overhead perspective of drone scenarios impose dual challenges on vision-language understanding. At the macro level, overwhelming background clutter in visual representations leads to Cross-Modal Focus Misalignment, where the model prioritizes global environmental similarities over specific object details. At the micro level, Visual Isomorphism creates ambiguity, where candidates share similar geometric structures yet differ only in subtle attributes. To address these challenges, we propose the Granularity-Aware Region Alignment and Semantic Prototype (GRASP) learning framework, enhancing discriminative capability through two synergistic strategies. Specifically, we introduce Region-Focused Alignment (RFA) to promote object-centric cross-modal alignment while suppressing background interference. Concurrently, to tackle visual isomorphism, we propose Semantic Perturbation Enhanced Matching (SPEM), which leverages a foreground-purified Semantic Prototype Codebook (SPC) to construct semantically perturbed negatives for fine-grained semantic discrimination. Extensive experiments on the GeoText-1652 benchmark and the unseen ERA dataset demonstrate that GRASP achieves competitive performance in drone-view fine-grained image-text retrieval, validating its effectiveness for cross-modal understanding in aerial scenarios. Our code implementation is available at \hrefhttps://github.com/UCAS-JC/GRASPhttps://github.com/UCAS-JC/GRASP.

Abstract:
Spoken Language Understanding (SLU) is the core component of task-oriented dialogue systems and a pivotal link in achieving seamless human-agent interaction. While traditional SLU can effectively extract user semantics for closed-set tasks after in-domain supervised fine-tuning, it faces significant challenges in leveraging in-context learning for open-domain tasks due to its ambiguous rule definitions. This work proposes Spoken Function Calling (SFC), a novel semantic understanding perspective that optimizes semantic understanding with structured rule definitions, to evolve beyond traditional closed-set SLU. Specifically, we curate and extend a suite of spoken functions based on traditional SLU datasets, construct a multi-agent system to synthesize the SFC-Bench dataset, evaluate the performance of Large Language Models (LLMs) and Large Audio Language Models (LALMs), and enhance the SFC capabilities of LALMs through post-training. Experiments demonstrate that SFC outperforms traditional SLU, substantially enhancing the semantic extraction accuracy for LLMs and LALMs.\footnoteThe code and dataset are available at \urlhttps://github.com/QwenAudio/FunResearch/tree/main/SpokenFC.

Abstract:
Existing salient object detection (SOD) methods in aerial imagery are fundamentally constrained by prohibitive annotation costs of fully-supervised paradigms. To bridge this gap, we present the first systematic investigation of semi-supervised SOD for aerial imagery. We propose S3OD, a SAM-driven Semi-supervised SOD framework built upon the Segment Anything Model 3 (SAM3), which seamlessly bridges foundation models and task-specific learners. Specifically, we equip SAM3 with a lightweight Saliency Focus Adapter (SFA), which employs a coarse-to-fine contextual design to enable spatially adaptive modulation for fine-grained saliency extraction. To foster robust and discriminative representation learning from unlabeled data under extreme label scarcity, we propose a dual-perspective perturbation learning strategy comprising Instance-Aware Perturbation (IAP) at the data level and Feature Contrastive Perturbation (FCP) at the feature level, jointly yielding a robust semi-supervised teacher. Furthermore, for efficient deployment, we present a Saliency Ranking Distillation (SRD) strategy that transfers cross-scale structural ranking knowledge from the SAM3 teacher to a lightweight SegFormer student, achieving a favorable performance-efficiency trade-off with only 10.19G MACs and over 170 FPS. Extensive experiments on three datasets demonstrate that S3OD consistently outperforms 12 state-of-the-art semi-supervised approaches under both 5 and 10 percent labeled settings, delivering absolute gains of 2--4 percent in F-measure. The code repository is https://github.com/lyf0801/S3OD.

Authors: Zixiang Peng (Institute of Information Engineering, Chinese Academy of Sciences), Yongxiu Xu (Institute of Information Engineering, Chinese Academy of Sciences), Qin-Yi Zhang (Institute of Automation, Chinese Academy of Sciences), Jiexun Shen (Institute of Information Engineering, Chinese Academy of Sciences), Yi-Fan Zhang (Institute of Automation, Chinese Academy of Sciences), Hongbo Xu (Institute of Information Engineering, Chinese Academy of Sciences), Yubin Wang (Institute of Information Engineering, Chinese Academy of Sciences), Gaopeng Gou (Institute of Information Engineering, Chinese Academy of Sciences)

Abstract:
Unified Multimodal Large Models (UMLMs) integrate understanding and generation capabilities within a single architecture. While unified architectures expand multimodal capabilities, their safety implications remain important yet underexplored. Existing safety benchmarks predominantly focus on isolated understanding or generation tasks, failing to evaluate the holistic safety of UMLMs when handling diverse tasks under a unified framework. To address this, we introduce Uni-SafeBench, a comprehensive benchmark featuring a taxonomy of six major safety categories across seven task types. To ensure rigorous assessment, we develop Uni-Judger, a framework that effectively decouples contextual safety from intrinsic safety. Based on comprehensive evaluations across Uni-SafeBench, we find that the original safety alignment of the underlying LLM is not consistently preserved in current unified models. Moreover, open-source UMLMs exhibit much lower safety performance than multimodal large models specialized for either generation or understanding tasks, particularly on the generation side. We release all resources at https://github.com/pengzixiang2002/Uni_SafeBench to systematically expose these risks and foster safer AGI development.

Abstract:
Transformer-based visual geometry models have shown strong potential for camera pose estimation and 3D scene understanding from video streams. Among them, Visual Geometry Grounded Transformer (VGGT) provides a powerful framework for joint camera and geometry modeling, but its training typically depends on large amounts of ground-truth supervision, which limits its applicability to unlabeled and unseen large-scale environments. In this paper, we present a self-supervised framework to adapt VGGT for large-scale visual localization from unlabeled driving videos. Instead of relying on conventional pairwise supervision, we extend self-supervision to sequence-wise geometric constraints that exploit richer temporal and multi-view consistency within long video streams. Specifically, multiple source frames are sampled in each sequence and geometrically projected onto different target frames to reinforce cross-view feature alignment and long-range spatial consistency. We further introduce a joint optimization objective that couples photometric consistency with geometric constraints, enabling the model to learn physically grounded camera motion and scene structure without hard labels. With this training strategy, the pretrained geometric backbone of VGGT is retained, while the camera and depth prediction heads are effectively adapted to large-scale localization scenarios under the proposed self-supervised objective. Experiments on challenging large-scale outdoor benchmarks demonstrate that the proposed method consistently improves long-range localization accuracy and trajectory stability in dynamic and adverse environments. Our code is available at https://github.com/Xnavi313/GPA-VGGT.

Abstract:
Autoregressive (AR) video diffusion models enable long-form video generation but remain expensive due to repeated multi-step denoising. Existing training-free acceleration methods rely on binary cache-or-recompute decisions, overlooking intermediate cases where direct reuse is too coarse yet full recomputation is unnecessary. Moreover, asynchronous AR schedules assign different noise levels to co-generated frames, yet existing methods process the entire valid interval uniformly. To address these AR-specific inefficiencies, we present SCOPE, a training-free framework for efficient AR video diffusion. SCOPE introduces a tri-modal scheduler over cache, predict, and recompute, where prediction via noise-level Taylor extrapolation fills the gap between reuse and recomputation with explicit stability controls backed by error propagation analysis. It further introduces selective computation that restricts execution to the active frame interval. On MAGI-1 and SkyReels-V2, SCOPE achieves up to 4.73x speedup while maintaining quality comparable to the original output, outperforming all training-free baselines. Our code is publicly available at https://github.com/hsc113/SCOPE.

Abstract:
Medical image segmentation remains challenging due to inherent imaging noise and spatial detail degradation caused by downsampling. Existing methods often suffer from the degradation of fine-grained details in deep network layers, while applying self-attention mechanisms to spatially large feature maps incurs prohibitive computational complexity. Furthermore, the lack of inductive biases in standard self-attention limits the representation capacity in data-limited medical segmentation scenarios. To address these challenges, we propose DyLoC-LA, an efficient framework designed to suppress noisy representations and preserve fine-grained spatial details. Specifically, DyLoC-LA comprises a Position Estimation Denoiser (PED) and a Dynamic Local-Global Aggregation (DLGA). The PED operates during the downsampling process to suppress noise and localize target regions. Within the DLGA, a Dynamic Local Perception Module (DLPM) adaptively extracts spatially contiguous local features, while a Context-Gated Linear Attention (CGLA) mechanism captures long-range dependencies, introduces inductive biases, and alleviates feature homogenization. Extensive experiments demonstrate that DyLoC-LA achieves state-of-the-art performance across ten datasets and four imaging modalities. Moreover, the proposed framework exhibits strong zero-shot generalization on five unseen datasets, validating its cross-dataset robustness. Code is available at \urlhttps://github.com/hao-ux/DyLoC-LA.

Abstract:
Simulation-to-Real (Sim-to-Real) video translation is essential for bridging the domain gap in autonomous driving, yet existing methods built on controllable video diffusion face a fundamental Consis\-tency-Realism Dilemma: low-level conditioning signals (e.g., edges, blurred images) ensure structural control but ``bake in'' synthetic textures, whereas high-level priors (e.g., depth, semantics) enable photorealism but discard the structural detail needed for consistent guidance. % We observe that Vision Foundation Model (VFM) features, specifically DINOv3, inherently encode both low-level structure and high-level semantics within a unified latent space, making them a natural bridge that sidesteps this trade-off entirely. % Building on this insight, we present \modelfull (\model), a controllable video diffusion framework that conditions generation on DINO features through three targeted designs. Notably, \model is trained exclusively on real-world driving videos without requiring any simulated data or sim-real paired samples. Minor Components Pruning applies PCA-based projection with stochastic Random Channel Tail Drop to suppress the high-frequency texture details responsible for synthetic artifact leakage, while preserving essential structural cues. Spatial Resolution Enhancer compensates for DINO's 16× spatial downsampling by processing inputs at higher resolution and aligning them to the diffusion backbone via a learnable Spatial Alignment Module, recovering fine-grained boundary control. Causal Temporal Aggregator replaces naive keyframe sampling with causal-convolution-based downsampling that preserves historical motion context, mitigating motion blur and enhancing temporal stability. Experiments show state-of-the-art fidelity (sKID, sFID) and temporal consistency, with competitive photorealism and structural consistency; see our \hrefhttps://albertchen98.github.io/DwD-project/

Abstract:
Speaker diarization technology is still challenged by poor cross-domain generalization of deep learning models, largely due to the limited availability of large-scale data resources. To address this issue, we present MISP-M^3SD, a large-scale multimodal, multi-scenario, and multilingual dataset for robust speaker diarization. Specifically, MISP-M^3SD contains more than 770 hours of synchronized audio-visual recordings collected from in-the-wild online videos, covering 14 scenarios and 16 languages. To enable efficient, cost-effective, and scalable dataset construction, we develop a largely automated pipeline including multilingual multi-scenario media acquisition, data cleaning and preprocessing, and cross-modal consistency-guided annotation. Specifically, we compare the predictions of audio-only, visual-only, and audio-visual diarization branches, and route low-confidence cross-modal inconsistency cases to selective manual verification. Experiments in both audio-only and audio-visual settings show that MISP-M^3SD is an effective training resource for cross-dataset generalization, and also provides a strong initialization for efficient scenario-specific adaptation with limited in-domain data. All related resources are publicly available at \urlhttps://github.com/coalboss/MISP-M3SD.

Abstract:
We study the problem of detecting whether consensus in a dyadic interaction reflects genuine information integration or a superficial agreement. While recent studies have linked interpersonal physiological synchrony to teamwork and decision quality, current multimedia datasets and models provide limited support for modeling consensus authenticity in conflict-driven interactions. To address this gap, we introduce CODEC (COnsensus DEtection in Conflict), a dyadic multimodal dataset collected with a hidden-profile paradigm that naturally induces disagreement, negotiation, and consensus formation. The dataset includes synchronized wearable physiological and behavioral streams from 13 dyads, together with fine-grained temporal annotations of interaction states. We also present a dynamics-informed late-fusion model that combines subject-aware prompt tuning, cross-modal attention guided by a cross-recurrence (CRQA) prior, and supervised contrastive learning to classify true versus false consensus from continuous multimodal segments. Under leave-one-group-out evaluation, the proposed model achieves a weighted F1 score of 0.85 and consistently improves over early-fusion and sequential baselines. Our results suggest that multimodal physiological-behavioral signals offer a promising direction for analyzing consensus quality beyond overt audiovisual behavior alone. The dataset can be accessed through \urlhttps://github.com/CODEC-MM/Dataset.

Abstract:
Expressway video anomaly detection is important for traffic safety, but remains challenging across diverse scenes, particularly for far-field vehicles with subtle abnormal motion. Vision-Language Models (VLMs) provide strong semantic reasoning capabilities, yet processing full frames can dilute evidence from distant targets and introduce substantial computational overhead. To address these challenges, we propose VIBES, an asynchronous framework that uses Bayesian inference to guide focused VLM reasoning. Specifically, an online kinematics-guided Bayesian inference module continuously estimates a context-dependent normal-motion distribution from vehicle trajectories and updates its probabilistic boundaries. Deviations from these boundaries produce asynchronous triggers that localize candidate anomalies in time and space. Instead of processing continuous full-frame video, the VLM reasons only over selected frames and localized visual regions associated with the triggers, reducing irrelevant visual content and unnecessary inference. Extensive experiments show that VIBES improves far-field anomaly detection and semantic interpretation while achieving real-time processing efficiency across diverse expressway conditions. The code is available at https://github.com/maoxiaowei97/VIBES.

Abstract:
\beginabstract Domain-Incremental Object Detection (DIOD) requires a detector to continuously adapt to new visual domains while retaining performance on all previously encountered domains. To resolve the stability--plasticity trade-off, existing approaches usually assign and train a new set of parameters for each new domain while freezing the backbone and the parameters for earlier domains. Although freezing the backbone eliminates disturbance to learned domains, it severely restricts the models ability to learn from new ones. Conversely, allowing the backbone to be freely updated enables it to capture new information but often degrades the representations needed for old domains. To tackle this dilemma, we propose a Partial Backbone Unfreezing-Retro-Calibration framework (PBU-ReCab) for DIOD. To mitigate the adverse effects of backbone drift, we derive a closed-form retrospective calibration rule that guarantees strict consistency. It directly acts on adapters parameters in an offline manner, without needing gradient computation or access to historical data. Moreover, we design two complementary mechanisms: Master Gradient Projection on LoRA Subspace (MGP-LS) and Adapter-Conditioned Distillation (ACD). The former projects backbone gradients into the null space of previous adapters while the latter promotes the feature alignment by comparing outputs from the evolving backbone and a frozen snapshot. Besides, we replace the standard classifier with a fixed Equiangular Tight Frame head. By coordinating these mechanisms, the model can more effectively adapt to new domains while preserving the knowledge from previous ones. Comprehensive experiments on the Pascal VOC series and the BDD100K series dataset demonstrate that PBU-ReCab achieves superior performance compared to existing approaches. %Existing approaches within a single training phase: freezing the backbone eliminates forgetting but creates a representational ceiling on visually distant domains, while unconstrained fine-tuning absorbs new knowledge freely but destroys representations that historical adapters depend on. %The core issue is not whether the backbone should evolve, but that its evolution must be structured enough to be correctable afterward. %After training, Retrospective Calibration (Retro-Calib) compensates for residual drift by solving a Ridge-regularized least-squares correction on each historical adapter's parameters in closed form, requiring no data access and negligible computation. % During training, Master Gradient Projection on LoRA Subspace (MGP-LS) projects backbone gradients away from directions that historical adapters are sensitive to, while Adapter-Conditioned Distillation (ACD) penalizes multi-scale feature deviation under historical adapter routing; together they confine backbone drift to a bounded, structured regime without suppressing plasticity. \endabstract

Abstract:
\beginabstract Cross-modal retrieval is a fundamental task in multimedia understanding, which aims to retrieve samples in one modality that share the same or similar semantics by querying with another modality. Existing methods typically treat labels merely as supervisory signals or use label information to enforce semantic consistency between cross-modal features, while overlooking the high-order semantic information implicit in the relationships among multiple labels and their combinations. To explore accordingly, this paper proposes a novel scene-aware cross-modal retrieval method. This method combines multiple single tags into a tag group as a scene and explicitly models the higher-order semantic relationships guided by the scene. Furthermore, to mitigate the interference of background noise, a self-attention mechanism is introduced within each scene to dynamically learn the semantic contribution weights of different cross-modal samples under the same scene. Meanwhile, considering that different scenes have varying importance across the entire dataset, we further leverage global label co-occurrence statistics to assign overall weights to scenes, so as to measure the relative importance of different scenes at the global level. Extensive experimental results demonstrate that the proposed method achieves significantly superior performance compared with existing approaches on three widely used cross-modal retrieval benchmark datasets. \endabstract

Abstract:
Diffusion models have achieved remarkable success in synthesizing complex static and temporal visuals, a breakthrough largely driven by Classifier-Free Guidance (CFG). However, despite its pivotal role in aligning generated content with textual prompts, standard CFG relies on a globally uniform scalar. This homogeneous amplification traps models in a well-documented ``detail-artifact dilemma'': low guidance scales fail to inject intricate semantics, while high scales inevitably cause structural degradation, color over-saturation, and temporal inconsistencies in videos. In this paper, we expose the physical root of this flaw through the lens of differential geometry. By analyzing Tweedie's Formula, we reveal that CFG intrinsically performs a tangential linear extrapolation. Because the natural data manifold is highly curved, this uniform linear step introduces a severe orthogonal deviation. To keep the generation trajectory safely bounded, we formulate a theoretical upper bound for spatial and adaptive guidance. Based on these geometric insights, we propose Spatial Adaptive Multi Guidance (SAMG), a training-free and virtually zero-cost sampling algorithm. SAMG dynamically computes point-wise conditional guidance energy, applying a conservative minimum scale to high-energy boundary regions to preserve delicate micro-textures, while deploying an aggressive maximum scale in low-energy regions to maximize semantic injection. Extensive experiments across diverse image (SD 1.5, SDXL, SD3.5 Medium) and video (CogVideoX, ModelScope) architectures demonstrate that SAMG effectively resolves the detail-artifact dilemma, achieving superior semantic alignment, structural integrity, and temporal smoothness without any computational overhead.

Abstract:
Segmentation with Vision Transformers typically ignores depth structure, while existing depth-incorporation methods either depend on depth sensors, treat depth as a passive input, or optimise fragmented surrogate losses disconnected from the evaluation metric. We observe that pixels at similar depths overwhelmingly share semantic identity---a property we call \emphiso-depth semantic coherence---and propose Unified Depth-Aware Segmentation (UniDAS), which threads this prior through feature extraction, attention, and optimisation. UniDAS comprises three synergistic components: (i)~\emphUniMoDE, which derives a latent depth prior from RGB features alone and routes tokens to depth-specialised experts, enabling depth-adaptive feature transformation; (ii)~\emphUniDAA, which groups window tokens into depth-coherent clusters and computes attention over cluster representatives, expanding the effective receptive field while reducing per-window cost; and (iii)~\emphUniDAR, a critic-free reinforcement fine-tuning strategy that jointly maximises a composite reward capturing semantic accuracy, boundary precision, and depth consistency---without additional inference-time parameters or depth annotations. Experiments on ADE20K, Cityscapes, COCO-Stuff, NYUv2, and SUN\,RGB-D across semantic, instance, and panoptic segmentation demonstrate that UniDAS establishes new state-of-the-art results.

Abstract:
Audio fingerprinting converts audio to much lower-dimensional representations, allowing distorted recordings to still be recognized as their originals through similar fingerprints. Existing deep learning approaches rigidly fingerprint fixed-length audio segments, thereby neglecting temporal dynamics during segmentation. To address limitations due to this rigidity, we propose Variable-Length Audio FingerPrinting (VLAFP), a novel method that supports variable-length fingerprinting. To the best of our knowledge, VLAFP is the first deep audio fingerprinting model capable of processing audio of variable length, for both training and testing. Our experiments show that VLAFP outperforms existing state-of-the-arts in live audio identification and audio retrieval across three real-world datasets.

Abstract:
The recent focus of response-based knowledge distillation (KD) possesses inherent constraints and limited effectiveness, underscores the necessity of developing more advanced approaches. However, existing methods fail to capture the distinct characteristics of local knowledge, i.e., relationality and divergence, leading to poor discrimination and generalization in student models. To this end, we propose Guided Scale Decoupled Distillation (G-SDD) to provide comprehensive guidance capability for student models via the elaborated context-adaptive guidance mechanism. To enhance guidance, we propose region-relational cross-attention learning to strengthen contextual perception among regions. Unlike conventional distillation that ignores inter-regional dependencies, our cross-attention distillation loss explicitly guides contextual feature learning. This enables the model to capture richer spatialsemantic relationships for more comprehensive context understanding. To mitigate divergence, we propose a dynamic soft-weighting mechanism that adaptively adjusts distillation weights based on sample informativeness. Unlike fixed weights that ignore varying information divergence, our similarity-based module adaptively reweights regions in a unified manner. This enables the student model to focus on informative samples and enhance deeper learning. Extensive experiments show that G-SDD surpasses state-of-the-art response-based methods.

Abstract:
Existing Any-to-Any MLLMs (Multimodal Large Language Models) generate only pairwise modalities ('Text + X') within a single response. We propose SpiderPTG for Any-to-Many Modalities Generation (AMMG) that enables the generation of arbitrary modality combinations ('Text + Xs'). First, we introduce the PTG (Planning-Thinking-Generation) structure, where Planning decomposes complex tasks into multi-steps, Thinking generates adaptive signals to guide decoders, and Generation produces coherent multimodal outputs. Second, we design an Interference-proof architecture to address signal interference and semantic degradation in multi-decoder control. Third, we propose a Pseudo-to-Real training paradigm, which progressively transitions the model from learning basic PTG structures on pseudo data to mastering high-quality content generation on real multimodal data. We further construct a comprehensive AMMG dataset with three progressively-enhanced subsets (Any2Many-Pseudo, Any2Many-PTG-Pseudo-10k, and Any2Many-PTG-2k), enabling models to develop sophisticated planning and thinking capabilities. Extensive experiments demonstrate that SpiderPTG outperforms existing methods in Any-to-Any and Any-to-Many generation scenarios.

Abstract:
Multi-view learning seeks to learn more comprehensive representations by exploiting the complementarity and consistency across diverse modalities or views. However, existing multi-view fusion strategies treat intra- and inter-view fusion as independent stages, without simultaneously considering the evolution within views and the dependency across views. Such an asynchronous fusion paradigm inevitably constrains cross-view interactions due to conflicting view-specific structural inductive biases. As a result, information flow is prone to distortion and compression along intermediate pathways, confining the model to learn within a restricted solution space. To address this, we propose Synchronous Multi-view Neural Diffusion (SynMDiff), which conceptualizes the multi-view feature space as a unified dynamical system driven by a diffusion process. By modeling the diffusion flow across arbitrary dyadic feature interactions in a joint space, SynMDiff enables the concurrent and adaptive intra- and inter-view information fusion. While a direct implementation of this synchronized mechanism incurs prohibitive computational costs, we further introduce an energy-based topological sampling strategy and an Ego-Net style centralized training architecture, ensuring both efficiency and scalability during learning and inference. Due to its conceptual elegance and computational efficacy, evaluations on real-world datasets demonstrate that SynMDiff outperforms the baselines by a large margin.

Abstract:
Existing image inpainting methods have achieved promising performance in recovering textures and semantics. However, these approaches struggle to preserve clear instance boundaries and coherent spatial layouts, when missing regions span multiple depth layers, due to the entanglement of structural cues. In this paper, we find that depth cues provide an effective geometric prior for identifying structurally compatible and visually plausible content for missing-region completion. Motivated by this finding, DeGIT is proposed as a depth-guided geometry-consistent inpainting Transformer for high-fidelity missing-region reconstruction. Specifically, DeGIT adopts a two-stage inpainting framework: the first stage produces a coarse yet visually and semantically plausible completion, while the second stage leverages depth cues as explicit structural guidance to calibrate the spatial layout and integrates the refined structures with the coarse prediction. We further introduce a strict reference geometry criterion to preserve cross-layer spatial hierarchy during refinement. Experimental results on Places2 demonstrate that DeGIT outperforms recent state-of-the-art methods in both quantitative results and visual quality.

Abstract:
Generating coherent motion sequences from natural language is an important step toward intuitive motion specification and interaction. Although recent diffusion-based text-to-motion (T2M) models have shown strong generation ability, they still suffer from semantic errors and temporal instability, which limits the quality and reliability of synthesized motions. In this work, we analyze T2M denoising from a frequency-domain perspective and show that it follows a hierarchical coarse-to-fine process: low-frequency components recover the global semantic structure, while high-frequency components refine local motion details. Motivated by this observation, we propose Free-T2M, a frequency-enhanced T2M framework with stage-specific frequency-domain consistency alignment. We further introduce a frequency-domain temporal-adaptive module that dynamically adjusts the contribution of different frequency bands across denoising steps. This design improves both semantic consistency and fine-grained motion fidelity. Extensive experiments on HumanML3D and KIT-ML show that Free-T2M significantly improves generation quality and robustness over strong diffusion baselines. In particular, when applied to StableMoFusion, our method reduces FID from 0.152 to 0.060 on HumanML3D. These results demonstrate that frequency-aware modeling is a promising direction for robust and reliable language-conditioned motion generation.

Abstract:
Face anonymization aims to remove identity information while preserving the utility of face images. Despite recent progress in identity protection, existing methods often fail to preserve gaze under viewpoint changes, leading to inconsistent gaze behavior across anonymized faces. We propose 3DGFA, the first face anonymization framework that enforces cross-view 3D gaze consistency. Specifically, 3DGFA introduces a world-coordinate multi-view gaze consistency objective to preserve coherent 3D gaze behavior across anonymized faces generated from different viewpoints. To improve anonymization quality, we propose a plug-and-play identity-attribute fusion module that can be inserted into arbitrary image encoders to effectively integrate anonymous identity cues with original facial attributes. We further develop a hybrid anonymous identity selection strategy to improve attribute preservation and output diversity. We also introduce a two-stage training scheme that equips the network with 3D awareness while keeping the training cost manageable. Extensive experiments across multiple datasets demonstrate that 3DGFA achieves a superior balance between privacy protection and utility preservation, while showing strong robustness, generalization, and practical effectiveness.

Abstract:
All-in-One video restoration (AVR) aims to recover clean videos from unknown mixtures of degradations. However, existing AVR methods typically assume simplified degradation settings, making them inadequate for the complex, time-varying corruptions encountered in real-world videos. To better characterize such degradations, we propose Degradation-as-Events (DaE), which formulates degradations as superposed events. DaE modulates the degradation intensity according to scene cues such as texture, motion, and illumination, thereby producing realistic, content-dependent corruption trajectories. This event-based formulation introduces a challenging restoration problem, in which degradation events vary over time and are strongly entangled with scene content. To address this challenge, we propose the Content-Degradation Disentanglement Network (CDDNet), which restores videos by explicitly reasoning over latent degradation events. CDDNet uses a dual-encoder architecture to disentangle content and degradation, and introduces Event Prototype Attention (EPA) to summarize dynamic mixed degradations into a compact set of event prototypes (EPs) without explicit degradation labels. An Event Prototype Transformer (EPT) decoder then performs clean-token intervention over these EPs, enabling robust restoration even under abrupt and severe quality fluctuations. Extensive experiments demonstrate that CDDNet outperforms state-of-the-art methods by 1.9 dB in PSNR and achieves consistently better visual quality across multiple benchmarks.

Abstract:
Moir\'e artifacts stem from frequency aliasing between display pixels and camera sensors, causing spatially non-uniform and temporally unstable distortions. Dual-camera smartphones mitigate this via synchronized streams: wide-angle frames provide high-frequency detail despite heavy moir\'e, while ultra-wide-angle frames are cleaner but suffer from blur and inconsistencies. In this paper, we propose DualVDM, the first \underlineDual-camera \underlineVideo \underlineDe\underlineMoir\'eing framework for smartphones. The framework transforms ultra-wide-angle frames into high-fidelity references via an ultra-wide-angle view enhancement (UWE) module and employs a dual-branch fusion backbone with kernel-aligned adaptive gated fusion (KAGF) module to recover fine-grained details from the wide-angle stream. To suppress moir\'e while preventing over-smoothing, we introduce the cross-camera frequency blending (CCFB) module for frequency correction, the difference-guided moir\'e correction (DMC) module for discrepancy-guided refinement, and the frequency-aware modulation (FAM) module for adaptive modulation, together with moir\'e confidence maps (MCMs). To reduce temporal flickering, we further introduce a lightweight temporal attention refinement module. Furthermore, we collect the first large-scale paired dataset to facilitate this novel and practical task. Extensive experiments demonstrate that DualVDM achieves clear gains in frame quality over strong baseline methods while maintaining competitive temporal stability.

Abstract:
Cross-modal hashing (CMH) has emerged as a prominent technique for large-scale multimedia retrieval due to its storage efficiency and query speed. However, existing CMH methods predominantly rely on the ideal assumption of complete, paired data and labor-intensive manual annotations. In real-world scenarios, data is frequently incomplete and unsupervised. These dual challenges render traditional supervised or paired-unsupervised approaches ineffective. To address this issue, this paper proposes a novel Unsupervised Incomplete Two-stage Cross-modal Hashing (UITCH) framework. Leveraging CLIP-based feature representations, UITCH decomposes the problem into two sequential stages: Semantic-Preserving Modality Reconstruction and Fusion-Guided Contrastive Hashing. In the first stage, we introduce a Cross-modal Mutual Information constraint based on neural estimation to guide the imputation of missing modality features, ensuring the recovery of intrinsic semantic distributions rather than superficial geometric patterns. In the second stage, we propose a Mutually Enhanced Affinity Matrix Fusion strategy integrated with a Cross-modal Transformer. By mining the consensus between intra-modal topology and cross-modal alignment, this strategy constructs a high-confidence affinity matrix that serves as a reliable pseudo-supervisory signal to guide the hashing optimization. Extensive experiments on three standard benchmarks (MIRFlickr, NUS-WIDE, MS COCO) demonstrate that UITCH significantly outperforms state-of-the-art baselines, validating its efficacy in handling data incompleteness without supervision.

Abstract:
Co-speech gesture video generation aims to create realistic videos of speakers where facial expressions and body movements are naturally synchronized with the speech audio. An effective system must reconcile two often-conflicting objectives: a) real-time performance, which is crucial for interactive applications, and b) high-fidelity synthesis, ensuring photorealism and natural motion. While impressive progress has been made, most methods excel at one at the expense of the other. To bridge this gap, we present a two-stage framework that achieves zero-shot high-fidelity synthesis while operating in real-time. Our approach strategically employs lightweight 2D upper-body keypoints as an efficient intermediate representation, decoupling the task into two stages. First, a motion generation module predicts precise 2D keypoints from the audio input, explicitly capturing its rhythm and semantic cues. Second, a rendering module employs an implicit warping strategy to synthesize the final video. Conditioned on the predicted poses, it uses a cross-modal attention mechanism to warp a source speaker image to the target pose. Our proposed method supports real-time generation of 512 ×512 resolution videos at up to 30 frames per second on a single RTX 3090. To facilitate research, we also introduce TED-GVD (TED Gesture Video Dataset), a new large-scale dataset with 150 hours of high-quality upper-body speech videos and corresponding skeleton annotations. Extensive experiments demonstrate that our approach achieves visual realism and synchronization accuracy comparable or superior to state-of-the-art methods while offering substantial speedups over diffusion-based methods.

Abstract:
Cross-view geo-localization is critical for drone-based geo-localization in GNSS-denied environments. Existing methods mainly rely on batch-local hard sample mining, but are less effective at handling false positive samples that remain competitive across training stages, resulting in recurrent false matches. To address this issue, we propose Ambiguity-Guided Relearning (AGR), a framework that explicitly captures and reuses recurring ambiguity patterns during training, in contrast to batch-local methods that focus only on instantaneous difficulty. Specifically, an ambiguity metric is defined based on retrieval ranking structure, by which the competitiveness of candidate matches is quantified using the gap statistics among top-ranked results. Based on this metric, Recurrent Ambiguous Memory (RAM) is constructed to retain recurring relations throughout training. Furthermore, Memory-Conditioned Optimization (MCO) is introduced, in which the original contrastive learning formulation is preserved while highly ambiguous cases are explicitly regularized. By exploiting historical ambiguity patterns accumulated during training, the proposed framework encourages more robust separation from recurrently competitive samples. Extensive experiments on multi-view geo-localization benchmarks demonstrate the effectiveness and strong extensibility of the proposed relearning paradigm.

Abstract:
Remote sensing change detection (CD) aims to identify where land-cover semantics change across time, but most existing methods still assume a fixed label space and therefore cannot answer arbitrary user-defined queries. Open-vocabulary change detection (OVCD) instead asks for the change mask of a queried concept. In the fully training-free setting, however, dense concept responses are difficult to compare directly across dates: appearance variation, weak cross-concept competition, and the spatial continuity of many land-cover categories often produce noisy, fragmented, and semantically unreliable change evidence. We propose Consistency-Regularized Open-Vocabulary Change Detection (CoRegOVCD), a training-free dense inference framework that reformulates concept-specific change as calibrated posterior discrepancy. Competitive Posterior Calibration (CPC) and the Semantic Posterior Delta (SPD) convert raw concept responses into competition-aware queried-concept posteriors and quantify their cross-temporal discrepancy, making semantic change evidence more comparable without explicit instance matching. Geometry-Token Consistency Gate (GeoGate) and Regional Consensus Discrepancy (RCD) further suppress unsupported responses and improve spatial coherence through geometry-aware structural verification and regional consensus. Across four benchmarks spanning building-oriented and multi-class settings, CoRegOVCD consistently improves over the strongest previous training-free baseline by 2.24 to 4.98 \mathrmF1_C points and reaches a six-class average of 47.50% \mathrmF1_C on SECOND.

Abstract:
Current noisy correspondence learning (NCL) pipelines typically treat correspondence quality as a binary variable, neglecting the abundant category of weakly-noisy correspondences. Two persistent issues are introduced: (i) over-exclusion, where partially informative pairs are discarded as negatives, shrinking the effective data manifold, and (ii) under-alignment, where residual noise from weakly mismatched pairs propagates through gradient updates, degrading representation fidelity. To address these challenges, this work proposes a unified forwardreverse diffusion framework called ``DiffNCL'' to amplify and subsequently refine weakly-noisy correspondences for robust noisy correspondence learning. In the forward diffusion, synchronized stochastic perturbations inject Gaussian noise into paired visualtextual embeddings, and step-wise similarities are aggregated to highlight the diffusion discrepancy of weakly-noisy mismatches. During reverse diffusion, two complementary consistency objectives, i.e., intra-modal structural consistency and cross-modal semantic consistency, progressively refine weakly-noisy correspondences while preserving primary semantic structure for subsequent training cycles. To better model weakly-noisy conditions, we adopt a controlled perturbation strategy that introduces local semantic variations while maintaining core semantic consistency. Extensive experiments on benchmark datasets, including Flickr30K, MS-COCO, and CC152K, demonstrate strong robustness and competitive performance under several noisy correspondence settings for cross-modal retrieval.

Abstract:
Hand gesture detection plays a vital role in human-computer interaction and intelligent interface systems, yet is inherently challenged by fast motion, extreme-lighting scenarios. Conventional color and depth sensors struggle to meet these demands due to limited frame rates, high processing latency, and low dynamic range. While event cameras offer a promising solution with microsecond temporal resolution and high dynamic range, dedicated event-based hand-gesture detection frameworks remain critically underdeveloped. To facilitate research in this field, we introduce the first large-scale event-based hand gesture detection dataset comprising 10 gesture categories with over 50K samples captured using DAVIS346 (346260) and EVK4-HD (1280720) sensors. We further propose GNN-STA (spatiotemporal aggregation), a hybrid GNN-CNN network that optimizes the accuracy-speed tradeoff. Specifically, the GNN extracts local spatiotemporal features, while a lightweight CNN module captures long-range dependencies and high-level semantics, complemented by cross-branch fusing to enhance overall representation. Extensive experiments conducted on our proposed datasets demonstrate the superior accuracy-speed tradeoff of GNN-STA. For instance, GNN-STA achieves a state-of-the-art 95.4% mAP at 241 FPS on EvGestureDet346 (vs. MvHeatDET 94.2% mAP at 54FPS), and 93.2% mAP at 104 FPS on the high-resolution EvGestureDet1280 (vs. Dino 91.4% mAP at 26FPS). Beyond validating our framework, these comprehensive evaluations against various detection approaches establish our dataset as a robust benchmark, providing a standardized foundation for future comparisons and research in event-based hand gesture detection.

Abstract:
Pansharpening aims to reconstruct a high-resolution multispectral image by combining the spatial details of a panchromatic image with the spectral information of a low-resolution multispectral observation. Although recent deep methods show strong performance, most of them rely on generic visual backbones and static fusion strategies, which limits interpretability and weakens adaptability across heterogeneous remote-sensing scenes. We propose a Text-aware Physical Prior Network to address the current issues. First, the scene prototype most relevant to the input image is retrieved from a predefined text pool using a frozen CLIP model. Then, the selected prototype modulates both construction parameters and channel-wise weights to generate adaptive physical guidance, which is further projected into dense features for visual refinement and multiscale fusion. Rather than directly injecting language tokens into the fusion backbone, the proposed framework introduces scene-level semantics into pansharpening. Experiments on GF-2, QB, and WV-3 show consistently strong performance under both reduced-resolution and full-resolution settings. Ablation studies further confirm the effectiveness of text retrieval, text-adaptive prior modulation, and guide-conditioned fusion.

Abstract:
Federated learning has emerged as a promising paradigm for privacy-preserving collaborative learning across distributed clients. However, the data owned by different clients are usually highly heterogeneous and severely Non-Independent and Identically Distributed (Non-IID), which leads to client drift and unsatisfactory global model performance. Although existing federated learning methods have attempted to alleviate the Non-IID problem, most of them mainly operate at the model level and still rely on frequent communication. In this paper, we propose a novel data-centric federated learning framework called FedGAT, to address severe data heterogeneity under limited communication budgets. On the client side, we improve the quality and diversity of condensed knowledge by preserving local geometric structure and prioritizing insufficiently learned yet non-redundant samples. On the server side, we further enhance global discrimination under heterogeneous data distributions by adaptively aligning class prototypes. Extensive experiments on eight medical and natural image datasets against ten state-of-the-art methods demonstrate that FedGAT achieves superior accuracy, outperforming the strongest baseline by up to 16.48 percentage points.

Authors: Xiaolong Sun (State Key Laboratory of Human-Machine Hybrid Augmented Intelligence, IAIR, XJTU), Wang Zhitao (2012 Laboratories, Huawei Technologies), Sanping Zhou (State Key Laboratory of Human-Machine Hybrid Augmented Intelligence, IAIR, XJTU), Hongcheng Huo (University of Toronto), Siyuan Cheng (2012 Laboratories, Huawei Technologies), Yonghao Dong (School of Software Engineering, XJTU), Liushuai Shi (State Key Laboratory of Human-Machine Hybrid Augmented Intelligence, IAIR, XJTU), Yuxiang Chen (University College London, University of London), Le Wang (State Key Laboratory of Human-Machine Hybrid Augmented Intelligence, IAIR, XJTU)

Abstract:
Trajectory planning is a fundamental task of end-to-end autonomous driving, necessitating both high precision and causal consistency. However, existing paradigms struggle to balance these objectives: autoregressive models are hindered by quantization errors inherent in discrete vocabularies, while diffusion models lack the sequential decision-making logic essential for causal coherence. Although recent masked diffusion models offer flexible trajectory decoding, they remain constrained by vocabularies. In this paper, we propose CoMaD, a novel Continuous Masked Diffusion framework that fully de-discretizes the masked diffusion process to balance planning precision and causal coherence. CoMaD redefines trajectory generation as an iterative refinement within a continuous state space, effectively eliminating cumulative quantization errors while preserving decoding flexibility. It begins by establishing continuous token modeling via a diffusion process, removing the dependency on categorical distributions. Building upon this, we design a new remask schedule by replacing traditional categorical confidence with a learned trajectory scoring head for iterative refinement. To further ensure robustness of masked diffusion, we integrate a noise-injection and denoising scheme for unmasked tokens during training, effectively mitigating the training-inference gap. Extensive experiments on NAVSIM-v1 and v2 benchmarks demonstrate that CoMaD significantly outperforms state-of-the-art methods in trajectory planning. Furthermore, qualitative analysis reveals that CoMaD autonomously determines optimal decoding sequences across diverse scenarios, showcasing superior interpretability and causal consistency.

Authors: Raphalle Lemaire (UNICAEN, ENSICAEN, CNRS), Azamat Kaibaldiyev (UNICAEN, ENSICAEN, CNRS), Elonore Mariette (UNICAEN, Inserm), Dbora Viglieri (UNICAEN, Inserm), Jeremie Pantin (UNICAEN, ENSICAEN, CNRS), Alexis Lechervy (UNICAEN, ENSICAEN, CNRS), Fabrice Maurel (UNICAEN, ENSICAEN, CNRS), Gal Dias (UNICAEN, ENSICAEN, CNRS), Gatane Blaizot (UNICAEN), Veronique Agin (UNICAEN, Inserm), Nicolas Poirel (Universit Paris Cit, CNRS), Eric Bui (UNICAEN, Inserm, CHU Caen), Herv Platel (UNICAEN, Inserm), Denis Vivien (UNICAEN, Inserm, CHU caen), Youssef Chahir (UNICAEN, ENSICAEN, CNRS)

Abstract:
This paper introduces MONET (Museum Observatory of Neuro-affective Eye-Tracking), a novel multimodal dataset collected over a nine-month period in a naturalistic museum setting, built to observe the effect of artwork on well-being. The core contribution of this work lies in the simultaneous capture of visual attention and affective responses from 148 visitors viewing original artworks, thereby ensuring an ecological validity often missing from traditional laboratory screen-based studies. The experimental design follows a unique two-phase protocol comparing solo and paired (two visitors simultaneously) visits, under both mediated and unmediated conditions. For each of the twelve paintings, the dataset provides individual saliency maps derived from mobile eye-tracking, valence and arousal scores obtained via an affect grid, and comprehensive demographic metadata. Exploratory findings reveal that while painting content remains the primary driver of inter-individual gaze variability, paired viewing leads to a significant convergence in affective responses between partners. Furthermore, the use of the Kullback-Leibler Divergence identifies a statistical link between spatial gaze distribution and emotional states, offering new pathways for modeling attention and affect in unconstrained environments. The dataset and supplementary materials are available at \urlhttps://git.unicaen.fr/raphaelle.lemaire/monet, and the MONET dataset is distributed under a Data Use Agreement (DUA) license.

Abstract:
\beginabstract An image conveys meaning through both its visual content and emotional tone, jointly shaping human perception. We introduce Controllable Emotional Image Content Generation (C-EICG), which aims to generate images that remain faithful to a given content description while expressing a target emotion. Existing text-to-image models ensure content consistency but lack emotional awareness, whereas emotion-driven models generate affective results at the cost of content distortion. To address this gap, we propose EmoCtrl, supported by a dataset annotated with content, emotion, and affective prompts, bridging abstract emotions to visual cues. EmoCtrl incorporates textual and visual emotion enhancement modules that enrich affective expression via descriptive semantics and perceptual cues. To align with human preference, we further introduce an emotion-driven preference optimization with specifically designed emotion reward. Comprehensive experiments demonstrate that EmoCtrl achieves faithful content and expressive emotion control, outperforming existing methods. User studies confirm EmoCtrl's strong alignment with human preference. Moreover, EmoCtrl generalizes well to creative applications, further demonstrating the robustness and adaptability of the learned emotion tokens. \endabstract

Abstract:
As deep models continue to scale, post-training quantization has become an important tool for efficient deployment. Compared with uniform quantization, mixed-precision quantization offers greater compression potential by assigning different bit-widths according to the quantization tolerance of different layers. However, existing methods mainly rely on layer-wise sensitivity ranking or discrete search for bit allocation, while making limited use of the directional differences of quantization perturbations in the local loss landscape. To address this limitation, we propose Compression Error Topography (CET), which models mixed-precision quantization as structured perturbations in parameter space and relates quantization perturbations to loss variation through a local second-order approximation. CET further leverages the spectral structure of the Hessian to characterize curvature variations across perturbation directions, identify low-curvature geometric subspaces that are more compression-friendly, and use them to guide layer-wise mixed-precision allocation. Extensive experiments across vision, language, and generative models demonstrate that CET achieves competitive compressionaccuracy trade-offs, with particularly strong performance in mixed-precision and ultra-low-bit regimes.

Abstract:
Vision-Language Navigation (VLN) requires an embodied agent to follow natural-language instructions and navigate in previously unseen environments. Despite recent progress in continuous multimodal navigation, long-horizon VLN remains challenging since it requires preserving useful history and maintaining decision consistency under partial observability. In this paper, we address long-horizon VLN by memory editing and propose a dual-memory framework, MemEdit-VLN. Specifically, MemEdit-VLN improves long-horizon VLN through Semantic Memory Consolidation, which compacts redundant semantic observations, and training-free Recallable Spatial Memory Selection, which preserves geometry-grounded spatial history for later navigation decisions. To better evaluate memory-intensive navigation, we construct the R2R Long-Horizon (R2R-LH) dataset from standard VLN-CE episodes, with longer horizons and stronger cross-stage dependencies. Experiments on R2R-LH and standard short-horizon VLN-CE benchmarks demonstrate that MemEdit-VLN outperforms state-of-the-art methods, achieving a remarkable 71.7% relative improvement in success rate for long-horizon VLN.

Abstract:
Inverse rendering aims to recover geometry, materials, and illumination from images, enabling faithful novel-view synthesis and relighting. Achieving these goals for glossy objects remains a longstanding challenge, mainly due to strong view-dependent effects and inter-reflections. Despite substantial progress in NeRF-based and 3DGS-based methods, faithfully capturing such effects while recovering accurate surface geometry is still difficult, often resulting in high-frequency surface noise and structural artifacts. To address these issues, we propose GlossySurfels, a two-stage joint optimization framework based on 2D Gaussian surfels. We begin with fast initialization through physically-based rendering on each primitive, then perform deferred rendering with 2D Gaussian ray tracing restricted to mirror-reflection direction to explicitly model inter-reflections. To capture high-frequency specular lighting beyond spherical harmonics, we introduce Gaussian light probes as a stable representation for simulating multi-bounce specular transport. For more realistic environment reflections, we further extend the effective cubemap coverage. Finally, to refine geometry reconstruction, we propose the normal prior, depth filtering, and material-based relocation to optimize scene normals, depth, and primitives distribution. Extensive experiments demonstrate state-of-the-art performance for glossy objects in photorealistic novel view synthesis, geometry reconstruction, and relighting under novel illuminations, outperforming existing inverse rendering approaches. Our code and data will be made publicly available upon acceptance.

Abstract:
Video multimodal large language models (Video-MLLMs) have shown strong capability in video understanding, yet their adaptation to sequentially evolving domains remains underexplored. In real-world deployments, video data often arrives continuously from heterogeneous domains, requiring the model to acquire new domain-specific knowledge without overwriting previously learned capabilities. Existing continual learning methods typically rely on shared adaptation spaces, which can induce severe cross-domain interference and catastrophic forgetting. We propose Distribution-Aware Expert Routing (DAER), a parameter-efficient framework for continual Video-MLLM adaptation over evolving domains. DAER maintains domain-isolated lightweight experts while keeping the pretrained Video-MLLM backbone frozen, thereby decoupling domain-specific adaptation from the general multimodal knowledge of the pretrained model. To enable fine-grained specialization, we introduce an intra-domain distribution-aware routing mechanism that matches each input to expert-level prototype reservoirs using Maximum Mean Discrepancy (MMD). To address the absence of task identities at inference time, we further propose an inter-domain routing mechanism that performs prototype matching in a discriminative subspace for robust domain identification. In addition, we introduce adaptive domain merging to improve parameter scalability and adopt a two-stage optimization strategy to stabilize expert specialization during continual learning. We evaluate DAER by curating a domain-incremental benchmark built from ten VidQA datasets covering diverse visual environments and reasoning demands. Experiments on two strong Video-MLLM backbones show that DAER consistently outperforms prior methods. On InternVideo2.5-8B, for example, DAER improves the average accuracy from 64.38% to 67.59% over the strongest prior baseline while reducing backward forgetting from 1.81 to -0.14. These results demonstrate that combining domain-isolated expert adaptation with distribution-aware routing provides an effective and scalable solution for continual Video-MLLM adaptation under evolving domains.

Abstract:
As speech generation models become increasingly realistic and widely accessible, concerns about the misuse, attribution, and governance of synthetic speech continue to grow. Watermarking provides a practical way to make synthesized speech traceable and verifiable. Most existing speech watermarking methods embed watermark information into signal-level representations, such as waveforms or spectrograms. Under sufficiently strong distortions, the embedded watermark may be weakened or destroyed, leading to degraded detectability. In this paper, we propose SSTMark, a training-free speech watermarking framework that operates at the semantic level through text watermarking. Unlike conventional signal-level watermarking methods, SSTMark encodes watermark information into the semantic content conveyed by generated speech, and detects the watermark from the recovered linguistic content. Experiments on AudioMarkBench demonstrate that SSTMark exhibits the strongest average robustness. Compared with the state-of-the-art baselines at a fixed false positive rate of 1%, SSTMark improves the average detection rate by 4.6% and 16.9% on signal-processing edits and compression edits, respectively.

Abstract:
Multi-Modal Large Language Models (MLLMs) increasingly support complex reasoning across language, vision, and structured inputs, yet their reasoning processes, especially for models with explicit intermediate thinking (MLLMs-T), remain under-evaluated. Existing benchmarks largely emphasize perception or final-answer correctness, providing limited insight into how multi-modal reasoning succeeds or fails. To address this gap, we introduce MMMR, a benchmark for evaluating multi-modal reasoning with explicit thinking. MMMR contains 1,083 high-difficulty questions across six reasoning domains and a modular Reasoning Trace Evaluation Pipeline (RTEP) that assesses relevance, consistency, and structured error types beyond answer accuracy. Experiments show that MLLMs-T generally achieve stronger overall performance than non-thinking MLLMs, but even leading models exhibit substantial reasoning failures, including inconsistency and overthinking. These results reveal persistent gaps between answer correctness and reasoning quality, making MMMR a scalable foundation for diagnosing and improving multi-modal reasoning systems.

Abstract:
Recent advances in 3D Gaussian Splatting (3DGS) have revolutionized novel view synthesis and inspired extensions toward 3D scene understanding. However, integrating both appearance and semantic information into 3DGS exposes a key challenge: capturing fine-grained appearance requires highly dense primitives, whereas semantic information is inherently sparse and structured in the 3D scene. This mismatch complicates the joint modeling of detailed textures and consistent semantic structure. In this paper, we present an appearancesemantic balanced 3D Gaussian representation that enhances both semantic scene understanding and high-fidelity visual representation. In particular, (i) a Decoupled Appearance Learning (DAL) mechanism that disentangles a stable, view-consistent base appearance from fleeting, view-dependent effects, thereby mitigating the impact of illumination variations; (ii) a Multi-Attribute Consistency Consolidation (MACC) strategy that jointly considers multi-metrics to remove redundant primitives while preserving sufficient diversity intra-class variations; (iii) an Adaptive Semantic-driven Densification (ASD) mechanism that redistributes primitives toward semantically informative regions, improving semantic boundary sharpness and structural coherence. Extensive experiments show that the proposed framework delivers improved reconstruction and segmentation performance, underscoring the benefit of jointly exploiting semantic priors and appearance characteristics for efficient 3D scene understanding.

Abstract:
Vision-Language Navigation in Continuous Environments (VLN-CE) requires agents to follow natural language instructions in unseen 3D spaces. The dominant training paradigm relies on imitation learning from static expert trajectories, which optimizes action-sequence imitation without modeling the step-wise local intent behind each decision. This leaves agents unable to understand why each action is taken, limiting generalization to unseen environments and hindering purposeful exploration and self-correction. Existing solutions remain limited by unconstrained reasoning representations, expert-confined exploration, and the wholesale discard of failed self-generated trajectories. To address these, we propose INTACT, an intent-centered self-improvement framework that treats step-wise local intent as an explicit, structured, and optimizable variable throughout self-training. At its core, PLIA-CoT decomposes local intent into four schema-constrained dimensions, transforming free-form reasoning into a controllable semantic interface. Building on this representation, our framework forms a closed self-training loop that uses local intent to guide exploration, filter self-generated trajectories, and optimize policy learning from both successful and informative failed experiences. Experiments on R2R-CE and RxR-CE demonstrate that INTACT raises Success Rate from 32.5% to 58.8% over three self-training rounds, outperforming methods with large-scale external data using only 40.5% of their training samples, without any additional human annotation.

Abstract:
We present SubjectAnchor, a Subject-Aware Memory-to-Video paradigm for multi-shot storytelling in which the current shot is generated by conditioning on explicit visual memories extracted from previous shots. The objective is to preserve subject identity and scene consistency across cuts while retaining the controllability of shot-wise prompting. Built on Wan2.2-I2V-A14B, SubjectAnchor contains three key components: subject-related memory construction, subject-aware temporal rotary position encoding, and memory-aware attention partition. For each target shot, the method constructs a compact memory bank by tracing each required subject to its historical appearance and retrieving the most relevant precomputed keyframes. These memory frames are encoded into the model input as explicit visual conditions, while different subjects are assigned to separated negative temporal slots to reduce identity interference. In addition, memory-aware attention partition regulates the interaction between memory tokens and generated content within a shared backbone. This formulation preserves the appearance anchoring of explicit visual memory while remaining compatible with script-driven shot-by-shot generation. Experiments show that SubjectAnchor improves cross-shot identity consistency over representative memory-based and holistic baselines while maintaining competitive visual quality.

Abstract:
This study presents the hardware and software architecture of a transformative system for illuminating line drawings and letterforms. These mid-air illuminations are indoors and might be animated. The hardware contribution is a drone equipped with servo-actuated rod joints and a dense, addressable LED strip that enables arbitrary orientation, a LightBender. The software contributions are threefold. First, a Blender add-on enables users to register LightBenders, author graphics and animations represented by swarms of LightBenders, and deploy the swarm for illumination through one-click functions. Second, users may import SVG files into either the Blender add-on or a standalone LB-Author tool to illuminate line drawings directly from vector graphics. Third, a distributed execution framework using a motion capture system coordinates the flight, actuation, and lighting of individual LightBenders, enabling the swarm to illuminate line drawings synchronously. We present results from an IRB-approved human subject study (n=21) to evaluate the impact of LightBender misalignment on the perceived illuminations. Obtained results demonstrate that the system's 10.1 mm maximum misalignment is perceptually acceptable, with a median quality rating of 8 on a 0-10 scale.

Abstract:
Accurate turn-taking prediction from speech is essential for responsive conversational systems. Existing approaches typically rely either on purely acoustic modeling or on semantic information derived from automatic speech recognition (ASR) transcripts. Acoustic models preserve prosodic cues but have limited access to linguistic completeness, whereas text-based models discard fine-grained acoustic information. Parallel audio-text systems retain both sources of information but generally process them through separate pathways. To address these shortcomings, we introduce a unified audio-only approach for turn-taking prediction that operates directly on speech waveforms and jointly considers acoustic and semantic information by mapping token-like acoustic representations into a pretrained large language model (LLM). The model requires only acoustic input at inference time and is optimized directly for turn-yield prediction rather than treating turn-taking as an auxiliary objective. We study two variants: TurnAudio~CLS, a binary classifier, and TurnAudio~GEN, a generative formulation that predicts turn-switch tokens alongside transcriptions. Experiments on Switchboard and Fisher show strong performance in the controlled audio-only turn-yield setting, with substantial reductions in false positives relative to the evaluated acoustic and text-based baselines. Under the same GPU evaluation setup, TurnAudio~CLS also achieves lower measured per-sample latency than VAP. TurnAudio~GEN provides a word-level comparison with text-based pipelines but is not intended as a low-latency predictor. These results indicate that audio-conditioned LLM representations can provide a useful semantic prior without requiring explicit transcription at inference time.

Abstract:
Computer-generated holography promises photorealistic 3D displays via precise wavefront control. While recent pupil-aware frameworks embed the human pupil function into propagation models to ensure natural viewing, they still face two key challenges. First, they do not fully bridge the gap between coherent holographic reconstruction and incoherent human visual perception. When coherent wavefronts are constrained by a finite aperture, binary transmittance introduces abrupt spectral cutoff, leading to ringing artifacts at out-of-focus planes. Second, existing methods either approximate pupil dynamics using discrete states or require computational intensive operations to model continuous variations, limiting real-time adaptation. To address these challenges, we propose a unified framework based on continuous pupil-conditioned modeling and efficient holographic reconstruction. First, we introduce a soft pupil prior derived from the Stiles-Crawford effect and integrate it into the conditioning space as a weighting function to compensate ringing artifacts. We further design a pupil-conditioned network with a pupil-condition embedding module and a complex-valued Mamba phase generator to model continuous pupil dynamics and efficiently reconstruct phase holograms. Finally, we build a holographic display prototype using a high-speed digital micromirror device as a programmable pupil amplitude modulator. Extensive experiments on both simulation and real-world data show that our method outperforms prior pupil-adaptive holography, achieving a 4 times speedup with reduced ringing artifacts.

Abstract:
While 3D human mesh modeling has been extensively studied, enabling convenient and precise editing of shape attributes according to user intent remains a significant challenge. Existing methods typically rely on predefined editing operations, limiting both flexibility and fine-grained controllability. In this paper, we propose Text2Reshape, a text-driven framework for human shape editing that, for the first time, enables structured and quantitatively controllable manipulation from free-form textual descriptions. We formulate shape editing as a structured graph-based transformation problem using a part-aware graph representation, where inter-part dependencies are explicitly modeled to support controllable deformation. To address the long-standing conflict between localized control and global consistency, we introduce two task-specific static graph priors tailored for local and global editing, respectively. These priors are integrated with text-conditioned dynamic graphs, enabling both disentangled local manipulation and coordinated global adjustment within a unified framework. Furthermore, we propose a measurement-aware shape scaling strategy that injects anthropometric semantics into the latent manipulation space, allowing quantitative textual instructions to be translated into metrically accurate geometric deformations. To facilitate this task, we construct TS-Editing, a new benchmark dataset with paired human meshes, automatically generated textual annotations, and LLM-based realism assessment, enabling scalable training and evaluation. Extensive experiments demonstrate that Text2Reshape achieves flexible, diverse, and metrically accurate shape editing with strong semantic alignment to user-provided descriptions.

Abstract:
Multimodal agents that think with images iteratively manipulate visual evidence and invoke tools across many steps. Existing reinforcement learning methods reduce trajectories to scalar rewards, forcing the policy to discover reusable tool-use patterns from scratch on every new task; memory-based alternatives retain past experience, yet they rely on test-time retrieval, without updating the policy to absorb reusable patterns from that experience. Our key insight is that multimodal reasoning trajectories should be distilled into reusable skills that co-evolve with the policy during training, rather than being consumed as rewards or retrieved from a static store. To this end, we propose SPyCE (Skill-Policy Co-evolution), a framework that distills trajectories into a hierarchical skill library and updates it throughout reinforcement learning. Execution skills capture local visual operations, while workflow skills encode high-level priors that orchestrate tool use. During training, the policy model conditions on retrieved skills to guide its rollouts, while the skill library evolves using valuable rollouts generated by the policy. This creates a closed loop in which improved policies yield better skills, and the evolving skill library, in turn, provides stronger priors for policy rollouts. Experiments across eight benchmarks demonstrate that SPyCE consistently outperforms both RL-based and memory-based baselines. Further analysis reveals that both the hierarchical skill design and the co-evolution mechanism are critical to our design. These results suggest joint skill-policy optimization as a promising paradigm for building capable multimodal agents.

Abstract:
Multi-modal clustering (MMC) learns complementary and consistent representations from heterogeneous data to improve clustering accuracy in an unsupervised manner. Information bottleneck (IB) has gained increasing attention in MMC because it offers a principled way to balance information compression and task-relevant information preservation. However, existing IB-based MMC methods still rely mainly on local optimization during information preservation, without a unified optimization between modality-specific features and global cluster assignments. This causes two problems: cross-modal consistency information cannot be fully exploited, and modality-specific information that is truly beneficial to clustering is difficult to identify. Inspired by the proxy mechanism in computer networks, we propose a novel proxy information bottleneck (PIB) framework. We treat the globally fused representation as proxy semantics and use it as a bridge between modality-specific features and global cluster assignments, thus providing a unified optimization in the information preservation stage. Specifically, we first use variational encoders to learn latent representations for each modality, which achieves information compression for each modality. Then, we adaptively construct a more reliable global shared representation through cross-modal alignment and feed it into a clustering network to obtain global cluster assignments. Based on this, we further design a new conditional mutual information term to discover modality-specific but task-beneficial information under cross-modal guidance. The objective is optimized with variational inference and a Jensen-Shannon divergence estimator. Extensive experiments conducted on multiple benchmark datasets demonstrate the superior effectiveness of the proposed PIB method compared with state-of-the-art approaches.

Abstract:
Existing shadow removal methods largely rely on paired or unpaired training data, which are costly to obtain and often hinder generalization. Although zero-shot approaches alleviate data dependency, they typically lack strong image priors and physical constraints, resulting in unstable or unrealistic results. In this paper, we propose a physics-guided diffusion framework for zero-shot shadow removal. Specifically, we formulate shadow removal as a conditional generation problem and decompose it into a prior-driven generation term and a physics-constrained data-consistency term. To model the shadow-free image prior, we introduce a cross-shadow region attention mechanism within a pre-trained diffusion model, which enforces illumination consistency between shadow and non-shadow regions. Furthermore, to incorporate data-consistency constraints, we devise a two-stage progressive guidance strategy that incrementally introduces illumination-invariant structural guidance and physics-based shadow model guidance during the generation process. This design circumvents the challenges associated with direct estimation of physical model parameters. Extensive experiments demonstrate that the proposed method exhibits strong generalization capability and consistently outperforms state-of-the-art approaches.

Abstract:
Visual graph reasoning requires answering graph-theoretic questions directly from graph images, where graph topology and state are conveyed visually rather than given in symbolic form. Despite recent progress of vision-language models (VLMs), current approaches to visual graph reasoning still fail on simple visual graph problems. This reveals a fundamental limitation of existing approaches: they prioritize final-answer supervision over the intermediate recovery of an explicit graph representation that preserves graph topology and state from visual input. To address this limitation, we propose VGCompiler, a compilation-centric paradigm for visual graph reasoning via knowledge compilation. VGCompiler organizes reasoning around two compilers: a representation compiler that recovers a structure-preserving intermediate graph representation from visual input, and an operation compiler that compiles query intent under the recovered graph state into an executable graph operation. Specifically, we build VGCompiler on Qwen3-VL-8B and train it with reinforcement learning guided by a layered reward over executability, compiled graph validity, representation quality, and operation quality. VGCompiler uses a frozen observer to summarize graph and question conditions into lightweight signatures, enabling archive retrieval and code reuse across similar regimes. Experiments on three benchmarks GVLQA, VisionGraph, and VGCURE, show that Qwen-VGCompiler, built on an 8B backbone, surpasses the strongest closed-source VLM baseline by 28.9% and the strongest code-based baseline by 23.7%, while maintaining high efficiency. We further evaluate VGCompiler on three real-world domains, including metro routing, logistics delivery, and network fault assessment, where it generalizes across heterogeneous visual graphs and domain-grounded tasks.

Abstract:
Rectified flow models are increasingly used for text-to-image generation, a core multimedia content generation task, thanks to efficient, near-deterministic ODE sampling, but aligning them with human preferences or prompt-following objectives still commonly relies on reinforcement learning with stochastic rollouts. Existing extensions of GRPO to flow-based generation inject noise into the reverse dynamics and optimize terminal rewards with step-wise clipped surrogates, which introduces two key mismatches: an ODE-SDE terminal marginal mismatch between training and inference, and a reward-granularity mismatch caused by repeatedly reweighting a single terminal reward across long denoising trajectories. To address this, we propose Traj-MCPO, a trajectory-wise policy optimization framework for rectified flows under terminal rewards. Starting from the marginal terminal objective, we derive an equivalent trajectory-form policy gradient and build a trajectory-wise clipped surrogate with group-relative advantages. At the core of Traj-MCPO is a length-normalized trajectory likelihood ratio, which aggregates step-wise likelihood changes into a single horizon-normalized quantity so that all timesteps contribute uniformly through an averaged log-likelihood gradient. Building on this formulation, trajectory-wise clipping directly constrains the average log-likelihood shift between the updated and behavior policies, thereby reducing rollout drift and improving consistency with inference-time ODE sampling. Experiments on text-to-image post-training with reward models show that Traj-MCPO stabilizes optimization and consistently improves prompt following and image quality over step-wise GRPO baselines.

Abstract:
Post-training mixed-precision compression has largely been driven by magnitude-based objectives, such as reconstruction error, while leaving perturbation direction largely uncontrolled. As a result, compressed models often exhibit scattered and unstable performance shifts: similar levels of compression can lead to markedly different outcomes depending on the induced perturbation. This suggests that compression stability depends not only on perturbation magnitude, but also on perturbation direction relative to the local loss landscape. We propose DiRa, a unified, model-agnostic, and training-free framework for direction-aware post-training compression. DiRa is motivated by a simple view: compression decisions should be evaluated not only by the size of the induced perturbation, but also by its local effect on the loss. To achieve this, DiRa introduces neighborhood-aware local loss surrogates, using a first-order surrogate in N_1 and a second-order surrogate in N_2. Under this unified formulation, DiRa can be instantiated for both tensor decomposition and mixed-precision quantization. Experiments on diverse vision and language models show that DiRa yields more stable compression and competitive or superior accuracy--compression trade-offs compared with existing post-training baselines, while often remaining near-lossless under practical compression settings.

Abstract:
World models enable long-horizon planning by internally generating and evaluating imagined trajectories, making them a promising foundation for generalist agents. However, this imagination-driven decision process also introduces new security risks. Existing backdoor attacks typically aim to manipulate local features, one-step predictions, or instantaneous policy outputs. While such objectives may suffice for weaker reactive models, they are often ineffective against world models, where the learned dynamics prior and planning process can absorb or wash out the effects of shallow perturbations. More importantly, we find that world models exhibit a distinct backdoor vulnerability rooted in the long-tailed ranking structure of imagined trajectories, where disrupting the ordering of a few decision-critical trajectories can systematically hijack planning. To exploit this vulnerability, we propose TRAP, a backdoor attack framework for world models that targets imagined trajectory ranking. TRAP combines a tail-aware ranking loss to focus optimization on decision-critical trajectories with dual gating mechanisms that stabilize optimization and regulate when and where the attack penalty is applied. Under trigger conditions, TRAP alters the relative ranking of imagined trajectories to redirect planning outcomes, while largely maintaining the normal ranking structure on clean inputs. Experiments on DreamerV3 and TD-MPC2 across diverse tasks show that TRAP consistently induces sustained behavioral deviations and significant performance degradation, highlighting the need for dedicated security evaluation of world-model-based agents.

Abstract:
Unified visual tokenizers aim to align generative and understanding representations within a single framework. However, jointly optimizing high-frequency visual details and high-level semantic information introduces substantial quantization errors, particularly when non-semantic artifacts are forcefully discretized. This results in high-variance stochastic noise during training, severely undermining optimization stability and convergence. To address this challenge, we propose ModUni, a unified tokenizer integrated with a learnable Pre-Quantization Gating mechanism. This module functions as an adaptive feature manifold rectifier, learning input-dependent sparsity to filter out stochastic noise and enable non-linear manifold rectification prior to quantization. This leads to a latent space that better captures high-level semantics while preserving fine-grained details, as reflected by improvements in both semantic understanding and reconstruction metrics. Consequently, ModUni establishes new state-of-the-art performance on ImageNet-1K, achieving 0.25 rFID in reconstruction and 76.8% zero-shot accuracy. Furthermore, ModUni demonstrates exceptional scalability, achieving competitive results on multimodal understanding and class-conditional generation tasks.

Abstract:
Dataset distillation aims to compress large-scale datasets into compact synthetic subsets that enable efficient model training with comparable performance to the original data. Recent advances show that diffusion models are highly effective for this task due to their strong generative capability. Existing diffusion-based approaches typically perform class-wise clustering and use the obtained cluster prototypes as diffusion inputs to generate distilled samples. However, the uniformly distributed original samples in a category inevitably lead to entanglement near cluster boundaries, pulling the cluster prototypes closer and reducing the inter-cluster separation. Consequently, the diffusion model receives overly similar latent inputs, resulting in synthetic data with limited diversity. To address this issue, we propose Disentangled Embedding Optimization (DEO), a simple yet effective framework that replaces static clustering with trainable embeddings. These embeddings are optimized to match the Gaussian statistics (i.e., mean and covariance) of the original data distribution, while a diversity regularizer is designed to encourage separation among these embeddings to improve the intra-class variability. Extensive experiments demonstrate that our proposed DEO can generate more diversified and representative distilled data than existing cluster-based diffusion methods, achieving state-of-the-art performance and superior generalization across multiple datasets.

Abstract:
Bridging the domain gap across heterogeneous driving datasets remains a central bottleneck for scalable Semantic Scene Completion (SSC), as existing benchmarks differ substantially in sensor configurations, camera setups, and semantic label spaces, severely limiting cross-dataset generalization. We present Meta-Occ, a unified generative framework that bridges multiple SSC datasets through controllable multi-view diffusion. We construct a shared occupancy space by aligning coordinate systems, semantic labels, and camera configurations across four major driving datasets, yielding over 90K unified samples for scalable cross-domain learning. To ensure geometric consistency, we develop a 3D-aware diffusion framework that injects volumetric structural priors via a dedicated control branch and models cross-view correspondence through geometry-aware deformable attention. We further introduce a depth-aware sampling strategy that adaptively modulates the reverse diffusion schedule according to scene depth, mitigating resolution-dependent degradation of distant regions. Extensive experiments demonstrate that Meta-Occ produces geometrically consistent multi-view images. Our synthesized datasets consistently improve downstream SSC performance for both monocular and stereo methods when used as training data, yielding substantial gains in cross-dataset generalization and highlighting its practical value for scalable SSC learning.

Abstract:
3D Gaussian Splatting has achieved remarkable progress in real-time scene rendering, yet existing methods are predominantly tailored to a single camera model, typically the pinhole camera, which limits their applicability to images captured by heterogeneous devices. In this paper, we introduce CDR-GS, the unified 3DGS framework explicitly designed for cross-device rendering. Our key insight is to leverage 360 images as a universal intermediate representation that inherently encompasses the full field of view (FoV) of diverse camera models, thereby eliminating the need for error-prone block-wise rendering and mapping. A differentiable cross-device mapping then transforms the intermediate 360 images into target camera domains, forming a unified rendering pipeline supervised directly by original cross-device images. To mitigate the excessive GPU memory overhead of activating nearly all Gaussians during 360 rendering, we propose a Gaussian selection strategy that exploits prior Gaussians and target FoV to retain only high-contribution Gaussians. We further design an adaptive-FoV mapping method that substantially reduces textureless regions when projecting cross-device images into pinhole images for prior Gaussians generation. To facilitate future research, we establish CDR-Bench, a benchmark comprising three datasets and adapted baselines. Extensive experiments demonstrate that CDR-GS achieves state-of-the-art rendering quality and efficiency across diverse camera models.

Abstract:
Text-to-Image (T2I) diffusion models are widely recognized for their ability to generate high-quality and diverse images based on text prompts. Despite recent advances, these models are still prone to generating unsafe images containing sensitive or inappropriate content, which can be harmful to users. Current efforts to prevent inappropriate image generation for diffusion models are easy to bypass and vulnerable to adversarial attacks. How to ensure that T2I models align with specific safety goals remains a significant challenge. In this work, we propose a novel, training-free approach, called Prompt-Noise Optimization (PNO), to mitigate unsafe image generation. Our method introduces a novel optimization framework that leverages both the continuous prompt embedding and the injected noise trajectory in the sampling process to generate safe images. Extensive numerical results demonstrate that our framework achieves state-of-the-art performance in suppressing toxic image generations and demonstrates robustness to adversarial attacks, without needing to tune the model parameters. Furthermore, compared with existing methods, PNO uses comparable generation time while offering the best tradeoff between the conflicting goals of safe generation and prompt alignment.

Abstract:
Universal multimodal embedding (UME) maps heterogeneous inputs into a shared retrieval space with one model. Recent methods generate explicit chain-of-thought (CoT) rationales before extracting embeddings to resolve complex query intent, but this adds substantial inference cost and funnels rich multimodal evidence through a textual bottleneck. We propose PLUME, a latent reasoning framework that replaces verbalized CoT with a short autoregressive rollout of continuous latent states. A semantic-anchor-guided transition adapter steers the rollout along input-dependent reasoning trajectories under a fixed computation budget. A progressive explicit-to-latent curriculum uses verbalized reasoning only as a temporary training scaffold, then transfers it into hidden-state computation so no CoT is generated at inference. On the 78-task MMEB-v2 benchmark, PLUME outperforms strong explicit-CoT UME baselines while reducing hundreds of generated reasoning tokens to fewer than 10 latent steps and delivering a 30.3× inference speedup. It is especially effective when evidence is dense and structurally complex, as in video and visual document retrieval. These results show that structured latent computation preserves the benefits of intermediate reasoning without explicit rationale generation, offering a stronger accuracy--efficiency tradeoff for practical retrieval.

Abstract:
Image-based indoor scene relighting remains challenging due to the complex interplay between cluttered geometry and local illumination, requiring precise modeling of light position, color, and intensity. Existing data-driven methods implicitly learn this relationship via paired multi-illumination datasets. Nevertheless, this data is costly and fails to scale, which is essential for accurate light-source-level control. Conversely, inverse-rendering methods reduce the data dependency by incorporating physical priors; however, they lack the robustness of intrinsic estimation in challenging conditions. In this paper, we present FreeLit, a paired-free framework for controllable indoor relighting that explicitly manipulates light-source location, color, and intensity. Instead of relying on paired supervision, we construct a physics-guided illumination prior from intrinsic scene properties, generating a structured lightmap along with a pseudo-relit image to guide diffusion-based synthesis. To address instability in intrinsic estimation, especially in low-light scenes, we introduce a relighting-guided intrinsic stabilization strategy that enforces illumination-invariant reflectance through structure-aware distillation and consistency constraints. Furthermore, we propose controllability-oriented evaluation metrics to quantify alignment with user-specified illumination color and intensity. Experimental results demonstrate that FreeLit achieves stable, physically consistent, and controllable relighting, with improved robustness in low-light indoor scenes, without requiring paired supervision.

Abstract:
Recent advances in multimodal emotion recognition (MER) have shifted the focus from classification to explainable reasoning. However, reasoning complexity varies across cases and is often overlooked. Existing adaptive reasoning methods fail to explicitly model such cognitive difficulty, resulting in a see-saw trade-off between performance and efficiency. Therefore, we propose a Cognitively-Inspired Adaptive Reasoning framework (CIAR), which operates in three stages: (1) Cold-start fine-tuning to establish stable representations for adaptive reasoning; (2) Adaptive cognitive difficulty modeling captures two complementary aspects: modality conflict, inspired by cognitive load theory, quantifies the extra effort required under conflicting emotional cues across modalities; and modality reliance, inspired by cognitive bias theory, reflects difficulty induced by biased reasoning when removing a modality unexpectedly increases confidence. Unlike prior discrete difficulty measures, CIAR models difficulty as a continuous signal; and (3) Difficulty-aware reinforcement learning with verifiable rewards to iteratively optimize the reasoning policy. Extensive experiments across multiple MER benchmarks in both open- and closed-vocabulary settings demonstrate that CIAR achieves a superior balance of performance and efficiency. Furthermore, the framework yields reasoning processes that closely align with human cognitive mechanisms, while exhibiting robust out-of-domain generalization and excellent scalability as model capacity increases.

Abstract:
Multimodal Entity Linking (MEL) aims to link ambiguous multimodal mentions to their corresponding entities within a multimodal knowledge base (MMKB). Existing methods focus on designing fine-grained multimodal interaction and fusion mechanisms, and have achieved promising results. However, the void problem in mentions and entities, i.e. the missing textual or visual information, still hinders their performance in real-life datasets. In order to address this issue, we first categorize the void problem as void description, void image, and void entity, and then tackle them via description generation, void image learning, and void entity penalty, respectively. We also propose two optimization techniques, namely hard negative sampling and MLLM reranking, to further enhance the performance of MEL task. Extensive experiments on three public MEL datasets demonstrate that our method consistently outperforms existing approaches. Specifically, our base model without reranking improves H@1 by up to 3.66%, while the complete version extends this improvement to 8.86% over the strongest baseline. Moreover, our base method achieves up to a 6.99% H@1 increase in low-resource settings.

Abstract:
Blind image inpainting aims to recover plausible contents without an explicit corruption mask, requiring the model to jointly solve where-to-fix and what-to-fill. Existing methods typically follow two directions. Some introduce an additional mask predictor to improve contaminated-region localization, but incur extra inference overhead and error propagation. Others retain efficient single-stage inference, yet often lack reliable contamination awareness. To address this issue, we propose Mask Supervision Pre-training (MSPT), a training-only mask supervision paradigm for blind image inpainting. MSPT adopts an asymmetric traininginference design, using explicit mask supervision only during training to inject contamination guidance into a single-stage backbone. Concretely, contaminated-region detection is formulated as a dedicated pre-training task, enabling the backbone to internalize contamination awareness without any auxiliary mask predictor at inference time. We further introduce a lightweight Dynamic Context-aware Head (DCH) to extract contamination cues from deep multi-scale features. In addition, we propose a Region-Adaptive Mask (RA-mask) strategy for masked image modeling, which strengthens what-to-fill learning by applying heavier masking to contaminated regions while preserving clean contextual evidence. Extensive experiments on multiple benchmark datasets demonstrate the effectiveness of MSPT, which achieves strong quantitative and visual results while retaining efficient single-stage inference.

Abstract:
\beginabstract Standard softmax self-attention excels in global modeling but suffers from \mathcalO(N^2) complexity, limiting scalable deployment in vision applications. Linear recurrent attention models offer favorable efficiency, yet their compressed state transition can cause certain information retrieval deficiencies, and their causal nature is not aligned with 2D visual data. In this work, we revisit the hybridization of linear and softmax attention from a fresh layer-stacking perspective. First, through a systematic study on hybridization schedules, we demonstrate that sparsely inserting softmax attention into the middle to late layers effectively mitigates the long range retrieval bottlenecks of linear architectures. Second, guided by the structural demands of layer-wise hybrid backbones, we introduce VDelta, a high speed unidirectional visual linear recurrent block built upon Gated DeltaNet, with improved adaptation of causal recurrence to 2D visual modeling. By combining this hybridization schedule with the VDelta linear layer, we propose SoLA-Vision (Softmax-Linear Attention Vision), a hierarchical vision backbone featuring a fine-grained layer-wise design. By seamlessly integrating our macro and micro level designs, this hierarchical backbone achieves an exceptional trade-off among modeling capacity, computational complexity, and real world inference speed. For classification on the ImageNet-1K dataset, SoLA-Vision outperforms purely linear and other hybrid attention models, while achieving up to 4× higher throughput than conventional multidirectional visual recurrent models. In dense prediction tasks, SoLA-Vision consistently surpasses strong baselines by a considerable margin. \endabstract

Abstract:
Weakly supervised semantic segmentation~(WSSS) with image-level labels has successfully bridged the gap between weak annotations and pixel-wise predictions based on CAMs. Recently, Contrastive Language-Image Pre-training~(CLIP) has been introduced into WSSS to improve the quality of CAMs and yields promising results. Existing methods predominantly rely on a frozen CLIP and repose its pre-trained knowledge can seamlessly generalize to the downstream segmentation datasets. However, such frozen representations are doomed to lack the vital adaptability to overcome inherent distribution shifts between the CLIP's training set and the segmentation datasets without task-specific fine-tuning. In this manuscript, we propose a Self-Supervised Distribution Alignment~(SSDA) strategy to address this issue. Specifically, SSDA leverages the pseudo-labels generated by CLIP Surgery to reconstruct the segmentation dataset into an image-text matching dataset for full-tuning CLIP. During the training process, SSDA introduces a Feature Rectification Loss \mathcalL_FR to supervise the CAMs of the vanilla CLIP by the CAMs of CLIP Surgery, thereby achieving the goal of forcing the patch tokens to attend to the object itself rather than the frequently co-occurring background regions. Furthermore, we propose a SAM-based post-processing (SPP) strategy to generate better pseudo labels by refining the CAMs produced by the CLIP. SPP utilizes SAM-generated attention as a spatial guidance to suppress interference from irrelevant regions during affinity propagation, while simultaneously employing refined pseudo labels as prompts for SAM to yield more fine-grained masks. Experiments show that our method achieves the state-of-the-art performance on the PASCAL VOC 2012 dataset and competitive results on the MS COCO 2014 dataset.

Abstract:
Composing independently trained adapters within a shared diffusion backbone provides a modular approach to multi-character generation, but naive joint deployment often causes identity mixing, cross-character attribute leakage, and unstable scene composition. We study this interference from a parameter-space perspective and hypothesize that it arises partly from conflicts between overlapping dominant subspaces in shared layers. To address this issue, we propose SDO, a Subspace Deconflicting Operator for multi-adapter composition. SDO reconstructs layer-wise low-rank updates from the selected adapters, extracts compact subspace signatures, measures pairwise conflict through output-subspace overlap, and applies a permutation-equivariant transformation that suppresses harmful shared directions while retaining identity-specific characteristics. The resulting representations are mapped back to standard adapter updates and can be directly incorporated into existing diffusion inference pipelines. Experiments demonstrate that SDO consistently improves identity fidelity and compositional stability, with particularly clear gains as the number of jointly composed adapters increases.

Abstract:
Robotic grasp detection aims to learn reliable grasp poses from 3D point clouds. Existing methods primarily train neural networks to estimate grasp poses but often overlook the dataset biases stemming from imbalanced scene distributions. Specifically, due to the underrepresentation of densely cluttered and severely occluded scenes, models are prone to learning spurious correlations between 3D point cloud features and grasp poses. To address this issue, we introduce causal inference into grasp detection and propose a Scene Causal Intervention for Grasping (SCIG) network. We first build a tailored causal graph to formalize the confounding effect introduced by scene context. Building on this analysis, we construct a scene confounder dictionary that captures diverse scene prototypes through clustering of pre-trained scene features. Leveraging this dictionary, SCIG then implements a back-door intervention to mitigate such confounding, effectively reducing scene-induced biases. Importantly, our method is designed as a plug-and-play, model-agnostic plugin that can be seamlessly integrated into existing grasping frameworks. Extensive experiments on the GraspNet-1Billion dataset demonstrate that SCIG substantially improves generalization, particularly in minority and complex scenes. Furthermore, cross-dataset evaluations on the GraspClutter6D dataset validate the effectiveness of our causal approach for zero-shot 3D grasp detection.

Abstract:
Text-guided diffusion image editing aims to modify semantic attributes of an image while preserving its identity, layout, and background. However, na\"ively switching the text condition during sampling often causes global drift, as denoising dynamics propagate changes across tokens and can disrupt unedited regions. To address this issue, we propose Asynchronous Token Decoding Edit (ATDEdit), an inference-time framework that views each sampler step as a parallel update of a globally coupled token matrix and enables token-indexed condition switching with differentiated update policies. Instead of applying synchronous target-conditioned updates to all tokens, ATDEdit estimates editable locations using token-wise conditional surprisal and applies target-conditioned corrections to the selected token set. It supplies source key/value memory at keep-token positions and projects selected keep-token latent rows back to their source values; these operations promote background preservation but do not constitute a pixel-level invariance guarantee. This approach combines local editing and background preservation without external or user-provided spatial masks and without model fine-tuning. On PIE-Bench, ATDEdit achieves the strongest reported preservation metrics, including 27.44~dB PSNR and 0.055 LPIPS, while retaining competitive semantic alignment.

Abstract:
Text-guided image editing is an important capability for controllable content creation, and recent flow-based generative models make training-free editing increasingly practical. However, inversion-free editors such as FlowEdit remain expensive because both the source and target branches must be evaluated at every step, while the source-side query is re-sampled with different noise at each step, making naive sparse reuse unreliable. We introduce StraightEdit, an efficient low-step framework for inversion-free FlowEdit. StraightEdit keeps the target branch densely evaluated for semantic editing accuracy, reduces source-side computation through anchor-based sparse evaluation, uses Noise Correction (NC) to align the cached source prediction with the noise realization, and applies Rotational Correction (RC) to stabilize the resulting low-step editing update under large step sizes. It achieves comparable or better edit quality with fewer function evaluations while preserving structural and background fidelity in low-step settings. Experiments on PIE-Bench with FLUX and ReFlow show that StraightEdit improves the speed and quality trade-off over prior inversion-free baselines.

Abstract:
Robust image watermarking requires embedded signals to survive diverse distortions while remaining imperceptible. Existing deep methods follow the embedding--distortion--extraction pipeline with only end-to-end supervision, lacking explicit coordination between embedder and extractor. This creates a dual mismatch: the embedder writes without knowing how it will be read, while the extractor passively adapts to embedding behavior, forcing reliance on extensive distortion enumeration for robustness. We argue that robustness fundamentally requires write--read alignment between embedding and extraction. To this end, we propose DualAlign, a framework that achieves robustness through Dual-Level Explicit Alignment rather than distortion enumeration. At the feature level, an Adaptive Feature Modulation Module (AFMM) decouples spatial structure from modulation strength, determining where to embed and how strongly. These decisions provide the representational basis for robust watermarking. At the coordination level, a Collaborative Interaction Mechanism (CIM) exchanges modulation-state cues between embedder and extractor during training. These two components operate in mutual reinforcement: CIM's strategic synchronization is grounded in AFMM's feature-level robustness, while the guidance from CIM further refines AFMM's modulation. Extensive experiments demonstrate that DualAlign outperforms state-of-the-art methods even when trained with only two representative distortions, surpassing approaches employing extensive enumeration (10+ types). This reveals that explicit alignment is a critical mechanism for robustness, offering an alternative to exhaustive distortion enumeration.

Abstract:
Embodied reasoning requires models to perceive task-relevant objects and spaces in physical environments and maintain consistent visual grounding throughout multi-step reasoning. However, current vision-language models rely on text-only or coordinate-augmented chain-of-thought, where entity references remain implicit and ambiguous. This may cause the reasoning process to decouple from visual evidence, entity references to drift across steps, and a causal disconnection between the reasoning trajectory and the final answer, with these problems further amplified in multi-view scenarios due to cross-view appearance changes. To address these issues, we propose Pinned Chain-of-Thought (PinCoT), a structured reasoning paradigm that pins every reasoning step to visual evidence. PinCoT introduces the concept of reasoning anchor, which binds each task-relevant entity to a structured visual anchor with entity name, unique identity, view index, and spatial grounding, enabling consistent entity tracking across reasoning steps and views. We build a fully automated data generation pipeline to construct PIN-170K, a high-quality PinCoT-formatted reasoning dataset. We then train RoboPIN through three-stage post-training that progressively injects embodied knowledge, structured reasoning ability, and process-supervised alignment, with rewards that directly constrain both anchor localization and identity consistency during reasoning. On 14 benchmarks covering embodied spatial reasoning, multi-view reasoning, and pointing, RoboPIN with only 4B parameters surpasses 7B level open-source embodied models on average, achieving a 12% average improvement over the strongest 7B baseline, Mimo-Embodied. Further analysis shows that PinCoT improves grounding accuracy and cross-step identity consistency, validating the effectiveness of process supervision.

Abstract:
We present SentiAvatar, a framework for building expressive interactive 3D digital humans, and use it to create SuSu, a virtual character that speaks, gestures, and emotes in real time. Achieving such a system remains challenging, as it requires jointly addressing three key problems: the lack of large-scale high-quality multimodal data, robust semantic-to-motion mapping, and fine-grained frame-level motion-prosody synchronization. To solve these problems, first, we build SuSuInterActs (21K clips, 37 hours), a dialogue corpus captured via optical motion capture around a single character with synchronized speech, full-body motion, and facial expressions. Second, we pre-train a Motion Foundation Model on 200K+ motion sequences, equipping it with rich action priors that go well beyond the conversation. We then propose an audio-aware plan-then-infill architecture that decouples sentence-level semantic planning from frame-level prosody-driven interpolation, so that generated motions are both semantically appropriate and rhythmically aligned with speech. Experiments show that SentiAvatar achieves state-of-the-art on both SuSuInterActs (R@1 43.64%, nearly 2× the best baseline) and BEATv2 (FGD 4.941, BC 8.078), producing 6\,s of output in 0.3\,s with unlimited multi-turn streaming.

Abstract:
Text-to-audio retrieval has made significant progress with shared embedding models such as CLAP and Pengi, yet they often struggle with fine-grained semantic alignment due to the inherent modality gap between text and audio. In this work, we propose FORTE, a unified framework that integrates structured logical reasoning with parameter-efficient cross-modal alignment to improve retrieval precision. Our approach first transforms queries into first-order logic and refines them via a constrained search that preserves semantic invariance while introducing discriminative attributes. The refined representation is then aligned with audio embeddings using a lightweight projection module, followed by a predicate-aware re-ranking step that enforces logical consistency at inference. Extensive experiments on AudioCaps and Clotho demonstrate consistent improvements over strong baselines, particularly in challenging fine-grained scenarios. Our results highlight the effectiveness of combining symbolic reasoning with representation learning for cross-modal retrieval.

Abstract:
Point-based interactive segmentation, exemplified by the Segment Anything Model (SAM), has become a widely adopted interface for object selection. However, most existing methods assume that point prompts are accurately placed on the target, an assumption that often breaks in real-world settings such as cross-view collaboration and AR/VR interaction. In these scenarios, calibration errors, parallax, and projection noise introduce systematic drift and ambiguity, giving rise to the fuzzy point challenge. To address this gap, we formulate Semantic-Guided Fuzzy Point Segmentation, where an imprecise point is paired with a semantic prompt (text or a visual exemplar) to disambiguate the target. Building on SAM2, we propose FuzzySAM, a robust framework that corrects prompt drift via a plug-and-play Semantic-Guided Adapter. The adapter employs conditional deformable attention to predict sampling offsets around each fuzzy point and aggregate target-consistent features, effectively re-aligning the models attention to the intended object. Experiments on Ego-Exo4D and HANDAL-X show that FuzzySAM improves robustness to projection drift and outperforms strong baselines, providing a practical solution for segmentation under spatial uncertainty.

Abstract:
Training-free video anomaly detection (VAD) with multimodal large language models (MLLMs) offers a promising paradigm for open-world anomaly localization. However, localizing brief anomalies in long videos remains highly challenging due to anomaly dilution, where weak anomaly evidence is overwhelmed after being absorbed into coarse temporally coherent units dominated by normal context. Existing window-based methods suffer from context mixing, while graph-based approaches typically rely on coarse event-level representations, limiting their ability to preserve fine-grained anomalous moments. To address this, we propose S2MGraph-VAD, a training-free framework for anomaly localization via scene-to-moment graph reasoning. It first decomposes a video into scene-level units with stable contextual semantics, and then refines each scene into fine-grained moments to preserve abrupt local deviations without breaking scene coherence. A Scene-to-Moment Graph links scene context and moment-level anomaly cues for sparse MLLM querying and topology-aware dense localization. Experiments on UCF-Crime and XD-Violence demonstrate strong frame-level localization performance, with particularly clear gains on short-duration and context-diluted anomalies. These results show that preserving anomaly-bearing local moments within stable scene context is critical for training-free anomaly localization in long videos.

Abstract:
Most existing hand motion generation benchmarks for handobject interaction (HOI) focus on static objects, leaving dynamic scenarios with moving targets and time-critical coordination largely untested. To address this gap, we introduce the DynaHOI-Gym, a unified online closed-loop platform with parameterized motion generators and rollout-based metrics for dynamic capture evaluation. Given the difficulty of conducting reproducible, large-scale real-world evaluations for dynamic capture, we ground our study in simulation, which enables controllable target motion, standardized rollouts, and scientifically comparable benchmarking. Built on DynaHOI-Gym, we release DynaHOI-Gym-10M, a large-scale benchmark with 10M frames and 180K hand capture trajectories, whose target motions are organized into 8 major categories and 22 fine-grained subcategories. We also provide a simple observe-before-act baseline ( ObAct ) that integrates short-term observations with the current frame via spatiotemporal attention to predict actions, achieving an 8.1% improvement in location success rate.

Abstract:
Precise spatial modeling in the operating room (OR) is essential for intraoperative awareness, hazard avoidance, and surgical decision-making. Although existing approaches exploit multimodal data to learn spatial relationships, many depend on sensing modalities that are difficult to deploy in real clinical environments and remain limited in explicit 3D reasoning under constrained sensing conditions. Meanwhile, models trained primarily on readily available 2D data often fail to capture the fine-grained geometric and semantic structure of complex OR scenes. To address these limitations, we introduce OR-Expert, a large vision-language framework for 3D spatial reasoning with RGB-only inference. OR-Expert internally derives depth, panoptic segmentation, and point-cloud cues from RGB images and encodes them as structured spatial representations. Its Spatial-Enhanced Feature Fusion Block aligns these pseudo-modalities with RGB and textual features in a shared token space, enabling joint semantic, geometric, and language reasoning. The unified end-to-end MLLM therefore supports detailed spatial understanding without requiring external depth, segmentation, or point-cloud sensors, or additional expert annotations at inference time. Experiments on multiple operating-room benchmarks demonstrate that OR-Expert achieves state-of-the-art performance and generalizes effectively to unseen surgical scenes and downstream spatial reasoning tasks.

Abstract:
Quantitative spatial reasoning in visual-language models (VLMs) aims to infer spatial distances and directional relationships among objects in 3D space from a 2D image and a natural language query. Despite recent progress, VLM spatial reasoning remains brittle under distribution shifts, largely due to the high cost of 3D supervision. As a result, models often produce inconsistent or contradictory predictions when faced with novel object configurations or rephrased spatial queries, revealing a misalignment between learned representations and underlying geometry. To address this, we propose TTL-SR, a geometry-aware Test-Time Learning framework for quantitative Spatial Reasoning that leverages geometric consistency constraints and unlabeled test data to adapt models to target domains. Specifically, TTL-SR augments the input query with geometrically coupled auxiliary queries, filters unreliable predictions via adaptive geometric triggering to construct structured token-level pseudo-labels, and updates model parameters under a geometry-aware multi-objective loss using only test data. Experimental results demonstrate that TTL-SR significantly boosts spatial reasoning performance, yielding 6.47% and 9.41% accuracy gains for Qwen3-VL-4B-Instruct and SpatialRGPT-VILA-1.5-8B on Q-Spatial-ScanNet dataset, respectively.

Abstract:
In this paper, we propose a dynamic evolutionary diffusion approach dubbed DiffusionEvo, generating trajectories capable of rapidly responding to sudden driving interactions. Conventional generative methods fail to respond to perceptual changes during the generation process, causing a perception-generation lag between the dynamically changing scenario and the trajectory generated by the model, which poses significant challenges for timely responses to sudden interactions. Instead, our DiffusionEvo aims to model the generation process as a dynamic evolution of latent variables over continuous time, enabling the generation process to dynamically adjust based on time-varying scenario perceptions. Additionally, we construct dynamic adaptive anchors from multi-modal perception streams to guide the evolution of latent variables, which precisely characterize the variations in scenario perceptions. Compared with the state-of-the-art methods, the DiffusionEvo achieves 20\,ms response time for sudden driving interaction, demonstrates 5× improvement, and runs at 50\,FPS on NVIDIA 4090. Extensive experiments on the NAVSIM and Bench2Drive benchmarks demonstrate that our approach achieves state-of-the-art performance, setting new records with 93.2 PDMS and 67.49 DrivingScore, while significantly outperforming existing methods in responding to sudden driving interactions.

Abstract:
Cross-modal hashing (CMH) with partial labels aims to achieve efficient multimodal retrieval, with each pair annotated with an ambiguous candidate label set. However, most existing methods ignore the reliability of model predictions, which can lead to incorrect hash codes under partial-label supervision. To address this limitation, we propose Robust Evidential Cross-Modal Hashing (RECH), a unified evidential learning framework that explicitly models prediction uncertainty and incorporates it into both label disambiguation and cross-modal consistency learning. RECH consists of two tightly coupled modules. First, an Evidential Disambiguation Learning (EDL) module represents class predictions as Dirichlet distributions and performs label disambiguation based on the accumulated evidence for candidate labels, effectively avoiding overconfident predictions under insufficient or conflicting evidence. Second, an Uncertainty-Guided Consistency Learning (UGCL) module leverages evidential uncertainty derived from evidential fusion to adaptively regulate cross-modal alignment and hash discretization, thereby preventing unreliable semantic signals from dominating the hashing space. By jointly integrating these two modules, RECH achieves robust label disambiguation and stable cross-modal hashing with partial labels. Extensive experiments demonstrate that the proposed RECH achieves state-of-the-art performance across multiple benchmark datasets.

Abstract:
The widespread application of face recognition technology in multimedia environments has led to increasingly severe risks of personal privacy leakage. To address this challenge, privacy protection methods aimed at generating adversarial perturbations to prevent unauthorized identification have become a research focus. However, existing methods typically adopt an input-agnostic, general-purpose paradigm and fail to tailor protection strategies for specific faces. Meanwhile, text-guided modules often operate independently, lacking sufficient reference to the semantic information of the image. These limitations cause existing methods to struggle in balancing image quality and protection success rate. In response to this, this paper proposes a Meta-learning-based Tailored Text-enhanced Facial Privacy Protection framework. The core of the method lies in introducing a meta-optimizer that fine-tunes the protection process based on the input face and textual description, thereby tailoring the generation of privacy perturbations. The text encoding process incorporates information from the image, making the resulting text embeddings more targeted to the current input and achieving text enhancement. Furthermore, a pixel-space refiner, co-trained with the meta-optimizer, is introduced to predict and apply subtle adjustments, restoring visual quality while ensuring high privacy protection effectiveness. Experiments on the CelebA-HQ and FFHQ datasets show that, compared to existing methods, the privacy-preserving images generated by this framework achieve better performance in terms of transferability across different models and visual concealment, while maintaining efficient processing speed. We also conducted evaluations on commercial APIs in real-world scenarios, further demonstrating the practical application potential of the proposed method.

Abstract:
Conventional digital cameras suffer from limited shutter speed and dynamic range, making it difficult to capture blur-free and high-dynamic-range (HDR) images in fast-rotating scenes, which complicates panoramic image reconstruction. The high-speed spike cameras offer advantages such as high temporal resolution and high dynamic range, making them suitable for capturing panoramic HDR scenes within a short time. In this paper, we present a new framework for reconstructing panoramic HDR color images from single-second spikes. We first introduce a spike-based panoramic reconstruction method that leverages the differential of spike firing time (DSFT) representation for finer texture preservation and spike-noise suppression, followed by DSFT-enhanced vertical constraint matching and texture optimization for reconstruction. Since spike cameras inherently lack colors, the reconstructed panoramic images remain monochromatic. To incorporate color information, we employ a side-by-side conventional color camera that captures a single color frame instantaneously. This color guidance is subsequently integrated through a fine-tuned image editing model. To train and validate practicality, a comprehensive dataset including synthetic and real-world spikes is presented. Experimental results demonstrate that our method generates high-quality panoramic HDR color images in fast-rotating scenes.

Abstract:
Learning the intrinsic relationships among views and data in the anchor space is crucial for effective multi-view clustering. Existing methods focus on mining multi-view diversity and consistency, yet they fail to capture the delicate dynamic trade-off between them. To tackle this challenge, this paper proposes Balanced Dual-Anchor Multi-View Clustering via Time-Spanning Guidance (BAT-MVC), which adaptively modulates the interplay between complementarity and consistency across both anchor spaces and multi-view representations. Specifically, we first leverage a dual-anchoring strategy to learn underlying consistent information from two differentiated anchor spaces. Building on this, we construct a tensor-based module to jointly optimize inter-view and intra-view dynamics, which incorporates information-theoretic entropy weighting and a signal-to-noise ratio (SNR) strategy to regulate the optimization procedure. Moreover, our framework further pioneers a time-spanning consensus learning method to ensure effective and consistent guidance for the clustering structure. Experimental results demonstrate that BAT-MVC can selectively leverage diversity and consistency to achieve consistently superior clustering performance.

Abstract:
Camera motions can add dynamism to static scenes and are crucial for drone videography. To automate this process, existing drone videography methods often focus on scenario-specific rules like target following, which can be difficult to generalize to other shot types such as landscape shots and acrobatic camera moves. In this work, we formulate drone camera path planning for open-ended videography as an imitation learning problem using online videos. To make this possible, we introduce 1) a drone camera path dataset, 2) an interactive benchmark, and 3) a simple and effective baseline. First, to create the training set, we estimate drone camera trajectories by connecting reconstructed 3D camera poses, and develop a scalable filtering scheme to ensure data quality, producing 99k high-quality trajectories. Second, we introduce an interactive evaluation environment with 38 natural scenes and 7 real city scans, and a set of benchmark metrics for generated camera paths. Third, we present a strong baseline named DVGFormer. Despite its architectural simplicity, the proposed approach can reproduce complex cinematic behaviors such as scenic reveals and obstacle?weaving stunt moves, demonstrating the effectiveness of the proposed imitation learning formulation. Data and code are available.

Abstract:
Integrating asymmetric modalities with high-resolution structural details from panchromatic images and spectral properties from low-resolution multispectral images is a core challenge in representation learning. This is especially true in pansharpening, where the frequency difference between panchromatic (PAN) and low-resolution multispectral (LR-MS) signals often causes a severe coupling conflict. Standard operators in Cartesian coordinates struggle to harmonize these disparate features because the combined update of real and imaginary parts leads to phase drift and spectral blurring. To resolve this contradiction, we propose the Eulerian Neural Operator. This physics-inspired framework redefines asymmetric feature fusion as a continuous functional mapping on polar manifolds. Departing from standard convolutions, our approach uses Eulers formula to decouple feature representations into magnitude and phase components. Specifically, we develop an Eulerian interaction mechanism that splits the fusion process into two specialized paths. In the first path, explicit phase rotation ensures geometric consistency through linear evolution in the circular domain. In the second path, implicit magnitude mapping calibrates spectral distributions via neural manifold approximation. Operating in the frequency domain allows our model to capture global receptive fields and exhibit a resolution-invariant property, enabling zero-shot generalization across varying spatial scales. Extensive experiments demonstrate that this lightweight architecture achieves a superior balance between physical interpretability, computational efficiency, and reconstruction quality.

Abstract:
3D Gaussian Splatting (3DGS) enables fast, photorealistic talking-head rendering, yet accurate lip articulation remains elusive: mouth motion is often over-smoothed and may violate hard articulatory constraints such as bilabial closures, producing the notorious ``leaky mouth'' artifact. A key difficulty is that brief, discrete articulatory events are inferred from a continuous acoustic embedding under a regression objective, which biases predictions toward averaged mouth configurations. While modern self-supervised speech encoders provide rich prosodic and phonetic cues, they do not provide an explicit, frame-aligned linguistic target that reliably disambiguates closure-level events. We propose Phoneme-Driven Gaussian Splatting (PD-GS), which augments a 3DGS talker with time-aligned phoneme tokens obtained from an automatic ASR and forced-alignment pipeline. Our core component, the Linguistic Fusion Module (LFM), adaptively fuses continuous audio context with discrete phoneme embeddings through a learned gate, allowing the model to preserve smooth audio-driven dynamics while strengthening phoneme guidance on articulation-critical segments. PD-GS is trained purely from monocular video using image reconstruction and lip landmark supervision. On HDTF, PD-GS achieves the best lip geometry among the compared baselines (LMD 2.66) and qualitatively reduces closure violations in challenging phoneme sequences, yielding more linguistically faithful neural avatars.

Abstract:
Vision Foundation Models (VFMs), benefiting from large-scale pre-training, provide generic and highly transferable visual priors, making them a promising foundation for Remote Sensing Change Detection (RSCD). However, adapting VFMs to RSCD remains challenging due to the domain gap between natural and remote sensing imagery and the lack of bi-temporal joint adaptation in existing Parameter-Efficient Fine-Tuning (PEFT) strategies. In this paper, we propose a robust and elegant new paradigm for adapting VFMs to multi-temporal RSCD. Specifically, an Adaptive Spatial-Frequency Adapter (ASF-Adapter) is designed to jointly adapt bi-temporal representations with minimal trainable parameters by decoupling spatial and frequency modulation in the feature space. Notably, in the frequency branch, low-frequency components are leveraged for cross-temporal global style alignment, while high-frequency components are selectively refined to preserve informative boundaries and suppress non-semantic noise. Together with strengthened multi-scale structural context modeling, this design improves change-aware representation learning. In addition, a Text-Anchored Semantic Contrastive (TASC) loss is introduced to enhance the semantic consistency of changed regions by transferring semantic priors from vision-language models during training, without incurring additional inference overhead. Extensive experiments on two multi-semantic and three binary RSCD datasets, along with cross-dataset evaluation, demonstrate that our framework achieves state-of-the-art (SOTA) and exceptional generalization.

Abstract:
Degraded infrared-visible image fusion (DIVF) aims to integrate multimodal images under degradation factors. Existing methods typically address a single degradation type, limiting generalizability. Recent approaches have adopted large language models (LLMs) to simulate user instructions, yielding linguistic priors to handle diverse degradations universally. Despite promising results, they still face three issues: (1) semantic homogenization, as monotonic generation rules for LLMs restrict the diversity of generated prompts; (2) semantic ambiguity, as vague user instructions make it hard for the model to identify degradation types; and (3) detail loss, as high-level linguistic priors often overlook low-level cues crucial for preserving fine-grained visual details. In this paper, we propose DDAP, a novel text-guided universal DIVF framework that integrates Dual Degradation-Aware Priors to address diverse degradation types and user instructions. We first outline three guiding principles for dual priors, ensuring their coordination and adaptive perception of different degradations. For the linguistic prior, we introduce three-tier generation rules to enhance semantic diversity in LLM-generated prompts, along with a Linguistic-prior Adaptive Calibration (LAC) module to refine textual embeddings for enhanced interaction. For the frequency prior, we design the Frequency-prior Adaptive Decoupling (FAD) module to capture implicit degradation patterns based on subband distribution, enhancing detail preservation. Finally, we propose a Dual-Prior Fusion module to integrate both priors and apply an auxiliary classification task to further ensure degradation-awareness. Extensive experiments on degraded, compound, and unseen real-world scenarios validate the superiority and generalizability of our DDAP framework.

Abstract:
Multimodal clustering aims to group samples by jointly modeling cross-modal semantic consistency and modality-specific structural relationships. Existing methods typically embed all modalities into a single latent geometry, usually Euclidean, which is often insufficient to capture heterogeneous structural patterns such as hierarchies, cycles, and flat relations. This mismatch can distort neighborhood structure and degrade clustering quality. To address this limitation, we propose MixCurv, a multimodal clustering framework that learns representations on mixed-curvature product manifolds with hyperbolic, Euclidean, and spherical experts. MixCurv introduces an embedding-aware routing mechanism, guided by a local topology prior, to adaptively assign samples to appropriate curvature subspaces. It further employs a per-expert tangent contrastive objective to preserve manifold-specific structural cues during cross-modal alignment, together with a multimodal curvature alignment module that encourages consistent curvature preferences across modalities. Finally, clustering is performed using a geometry-aware Riemannian spectral scheme to preserve the learned non-Euclidean structure during inference. Experiments on five benchmarks covering both multimodal and multi-view settings show that MixCurv consistently outperforms strong baselines and achieves competitive clustering performance.

Abstract:
Despite significant advances in weather forecasting, operational meteorology still heavily depends on human experts to interpret complex weather systems and integrate multi-source meteorological data. Multimodal large language models (MLLMs) have recently attracted considerable attention as potential assistants for these tasks. However, the hallucination problem in MLLMs constitutes a major obstacle to their reliable deployment and remains a largely underexplored challenge in meteorological analysis. In this work, we go back to basics and explore the causes of hallucinations in MLLMs for meteorological analysis. We introduce a new benchmark, i.e. MeteorIR, and design several targeted tasks, ranging from basic visual perception to map reasoning and precipitation cause analysis, for this investigation. We observe that inaccurate basic perception is a major issue, degrading meteorological map reasoning. We also reveal the limitations of MLLMs in integrating multiple perceptual signals for reasoning. Finally, we show that strengthening these map perception and reasoning skills through pre-training is able to significantly improve performance in downstream precipitation analysis tasks. We expect this work to drive improvements in the reliable weather MLLM design.

Abstract:
Human-Object Interaction (HOI) detection, a fundamental problem in visual scene understanding, relies heavily on contextual semantics to disambiguate complex interactions. Although Transformer-based methods have achieved notable progress, they largely remain instance-centric and insufficiently exploit the broader contextual cues that are crucial for human-like reasoning. To address this limitation, we propose MSSU (Multi-stream Semantic Understanding), a novel multi-stream HOI detection framework. MSSU uniquely integrates four complementary streams to capture diverse semantic cues, within which one stream leverages Stable Diffusion to directly inject human-like shape priors into the HOI detector. Furthermore, to effectively synthesize these streams, we design two novel fusion strategies: the Visual Geometry Encoder (VGE) and the Orthogonal Decomposition Fusion (ODF). The former fuses geometric and visual features, thereby enhancing the basic visual representations with relation-aware knowledge. The latter orthogonally fuses textures and shape priors, thus reducing the redundancy between these distinct features. Empowered by this multi-stream design, MSSU achieves state-of-the-art performance on the challenging HICO-DET and V-COCO benchmarks. Extensive experiments demonstrate that MSSU effectively captures diverse contextual semantics, enabling more robust and human-aligned reasoning for HOI detection.

Abstract:
Multimodal large language models (MLLMs) have demonstrated remarkable capabilities in complex reasoning tasks by integrating visual and textual information. However, the autoregressive generation of intermediate reasoning steps leads to rapid token accumulation, resulting in quadratic computational complexity and substantial memory overhead. Existing token compression methods fail to address the unique challenges of multi-hop multimodal reasoning, where token importance varies dynamically across reasoning steps and modalities. We propose Active Token Pruning (ATP), a novel framework that trains an active token pruner to selectively retain semantically critical tokens during multimodal reasoning. Unlike passive pruning strategies, ATP employs reward signals derived from task performance to guide token selection, enabling the pruner to learn task-specific importance patterns. Our method introduces a reinforcement learning-based pruning policy that conditions on reasoning progress, modality characteristics, and cross-modal fusion dynamics. Through extensive experiments on seven spatial reasoning benchmarks and four general multimodal tasks, we demonstrate that ATP achieves 42.3% reduction in FLOPs and 48.1% reduction in KV-cache usage while maintaining or improving accuracy.

Abstract:
Most visual trackers formulate tracking as a frame-wise conditional inference problem, repeatedly using the first-frame template to condition each incoming search frame. To remain efficient, existing methods usually inject target cues sparsely, while stronger target awareness often requires target-conditioned computation throughout the network, increasing per-frame cost. We revisit tracking from a different perspective: after initialization, the target has already been specified, so repeated explicit template-search interaction may be unnecessary. We therefore propose CSTrack, a compile-then-specialize framework that reformulates frame-wise conditional tracking as sequence-level model specialization. CSTrack uses the first-frame template as a compiler signal and generates block-wise operator priors with a Template Prior Compiler. These priors are injected into the backbones depthwise convolution layers to construct target-specialized representations. We further introduce one-shot weight integration, which absorbs the target-conditioned branch into the backbone weights and yields a target-specialized backbone for subsequent frames. Thus, most target-conditioned computation is performed only once during initialization, reducing online inference cost. Experiments demonstrate that CSTrack achieves a favorable accuracy-efficiency trade-off, supporting the effectiveness of the proposed paradigm.

Abstract:
Federated learning frequently encounters the challenge of label noise in real world scenarios. Existing robust methods commonly rely on local statistical partitioning to separate clean and noisy samples. However, this assumption breaks down under the sparse or heterogeneous local data common in federated settings. The resulting local statistical collapse destroys the bimodal loss structure, causing unreliable clean-noisy separation and optimization instability. To address this fundamental limitation, we propose InstaFed, a novel instance aware federated learning framework with formal theoretical guarantees for its core components. Instead of forcing scarce samples into unreliable statistical distributions, InstaFed tracks the longitudinal evolutionary trajectory of individual instances. This achieves highly precise noise decoupling that is independent of local data scale and class balance. Specifically, we utilize cumulative prediction entropy for privacy preserving client partitioning and maintain an exponential moving average of historical confidences to dynamically extract clean samples. Furthermore, we design a selective instance level semantic alignment mechanism that strictly neutralizes the memorization of corrupted labels without restricting the capacity of local models to learn authentic heterogeneous features. Extensive experiments on multiple benchmark datasets demonstrate that InstaFed significantly outperforms existing state of the art methods under various federated label noise settings.

Abstract:
Self-training with teacher-generated pseudo-labels has driven notable progress in cross-domain object detection. However, most existing methods primarily rely on output-level supervision, while overlooking the transfer of intermediate representations. Although feature distillation appears to be a natural solution, directly applying conventional distillation under domain shifts often leads to limited gains. In this paper, we identify two key challenges that hinder effective distillation: (1) deceptive channel noise caused by domain discrepancy, and (2) blind spatial alignment in the absence of target annotations. To address these issues, we propose Domain-Aware Feature Distillation (DAFD). Specifically, DAFD consists of Dual-Perspective Channel Refinement (DPCR), which jointly evaluates channel structural quality and student-teacher spatial agreement to suppress unreliable channel responses. Concurrently, Collaborative Confidence Region Mining (CCRM), which dynamically mines discriminative spatial masks to focus distillation on informative regions. Through domain-aware channel purification and spatial calibration, DAFD enables more reliable feature transfer across domains. Extensive experiments on standard cross-domain benchmarks demonstrate that DAFD achieves state-of-the-art adaptation performance across diverse detectors. We further verify its effectiveness in the more challenging source-free object detection.

Abstract:
RGB-T tracking has attracted increasing attention due to its ability to simultaneously exploit RGB and thermal infrared modalities for robust tracking. Achieving accurate RGB-T tracking requires not only effective cross-modal fusion to exploit modality complementarity, but also efficient inter-frame interaction for temporal modeling. However, many existing methods either rely on complex fusion designs or perform insufficient inter-frame and cross-modal interaction modeling, resulting in limited effectiveness and efficiency. To alleviate the above issues, we propose GateKVT, a novel progressive fusion framework with a gated KV-Cache mechanism for effective inter-frame interaction and cross-modal information integration. Specifically, KV-Cache is adopted to cache and reuse template information, while a lightweight gating mechanism is introduced to effectively regulate the cached information. Based on this design, a multi-modal causal (MMC) attention mechanism is further introduced to model inter-frame and cross-modal interactions. Furthermore, a simple hierarchical progressive fusion (HPF) module is designed to progressively integrate information at both the frame and modality levels before the prediction head. Extensive experiments on four RGB-T tracking benchmarks demonstrate that the proposed method achieves favorable effectiveness and efficiency.

Abstract:
The explosive growth of multimodal data drives increasing retrieval demands for efficiency and scalability, making cross-modal hashing a promising solution. However, offline hashing methods are ill-suited for dynamic online environments due to the cost of full model retraining, while existing online methods often lag behind in retrieval accuracy. This motivates us to seek a framework capable of preserving offline-level accuracy through efficient closed-form updates. The core bottlenecks hindering current approaches lie in computationally redundant iterations that compromise speed, and the underutilization of structural priors that limits precision. To address these challenges, we propose Fast Perceptual Cross-Modal Hashing (FPCH). FPCH first employs random projection to bypass iterative feature compression. Subsequently, it leverages the Discrete Cosine Transform basis as a set of complete orthogonal structural priors. By projecting latent features onto this complete basis, FPCH effectively captures the intrinsic distribution structures across bits, enabling the direct generation of robust hash codes that preserve essential discriminative information. Extensive experiments demonstrate that FPCH achieves superior retrieval accuracy and speed across multiple benchmark datasets, effectively combining offline-level accuracy with online agility.

Abstract:
Multimodal image fusion (MMIF) aims to integrate complementary sensor data into a single representation that preserves \emphintrinsic scene reality while eliminating \emphenvironmental interferences. Most existing approaches rely on \emphblind feature aggregation, which excels at signal accumulation but fails to distinguish essential content from physical degradations. We propose SGPDFuse, which bridges this gap by mapping inputs into a physics-disentangled structural representation via a \emphSemantic-Physical Parametric Bridge (SPPB) built on pretrained vision foundation models, utilizing the Intrinsic-Variation principle to decouple invariant scene attributes from transient environmental factors. To guide this decomposition, we introduce a \emphSemantic Alignment mechanism: we explicitly anchor the fused representation to salient semantic features in the same foundation model feature space via cosine similarity to preserve critical targets, while enforcing physical texture fidelity through Gram-matrix regularization to strictly eliminate unnatural artifacts. Extensive experiments demonstrate that SGPDFuse achieves state-of-the-art performance across infrared-visible, multi-focus, and multi-exposure benchmarks using a single architecture.

Abstract:
The rapid rise of social media platforms has significantly increased information accessibility while also facilitating the widespread dissemination of fake news, causing profoundly negative impacts on individuals and society. Current fake news detection methods are typically tailored for specific modality combinations (e.g., text-only, text-image, or text-video), making them inflexible in adapting to diverse modality combinations actually present on social media platforms. This limitation restricts the practical value of detection models in real-world scenarios. In this paper, we propose Uni-Detector, a universal multimodal fake news detection model. It customizes the Sparse Mixture-of-Experts (SMoE) layer as both the Feedforward Network (FFN) and Multi-head Self-Attention (MSA) modules within the Transformer architecture, tackling the aforementioned barriers. This approach separates the parameter space sufficiently for different modalities. The conditional computation nature of SMoE allows us to organically allocate relevant model components to dynamically accept and process any combination of input modalities. We conducted extensive experiments on Chinese and English multimodal datasets, each comprising three categories: text-only, text-image, and text-video. The proposed Uni-Detector achieves highly competitive performance across all modality combinations on test sets, establishing the Uni-Detector as a more flexible and powerful tool for detecting fake news in real-world scenarios.

Abstract:
Text-to-video (T2V) diffusion models can generate realistic depictions of actions such as kicking, stabbing, and shooting, raising safety concerns that motivate targeted concept erasure. Although concept erasure has been extensively studied for static concepts in text-to-image and T2V models, erasing motion concepts remains largely unexplored. We present a systematic study of motion concept erasure in video Diffusion Transformers (DiTs). Through causal interventions, we show that text-conditioning attention carries concept-specific motion information and supports selective intervention, whereas perturbing temporal positional encoding suppresses both target and non-target dynamics. We further find that directly adapting ESD, a representative weight-level image erasure method, to a video DiT yields modest and uneven motion suppression: reducing its erasure training loss does not by itself remove the concept signal from the difference between the conditional and unconditional predictions, which classifier-free guidance (CFG) then scales at every denoising step. From these findings, we derive three requirements for motion concept erasure: concept specificity, spatial selectivity, and temporal naturalness. Each determines one component of \method (Motion concept Unlearning in Text-to-video gEneration): at each denoising step, \method extracts a concept direction through token neutralization, derives a spatial gate from the direction's intrinsic structure, and subtracts the resulting correction from the velocity output before CFG is applied. \method is training-free and requires no weight modification. Experiments on 20 motion concepts show that \method outperforms representative prompt-level, weight-level, and inference-time baselines on Wan2.1-T2V, and the same formulation transfers to CogVideoX, supporting its applicability across distinct T2V attention architectures.

Abstract:
\beginabstract Many indoor applications, such as mobile navigation and interior redesign, require the 3D model of the scene to structurally reflect the reality. We study the problem of 3D completion for \emphatrium scene, i.e., a large and visually consistent multi-storey indoor space, given its sitemap (composed of the layout of each floor) and possibly an incomplete 3D model of the site. Existing approaches either consider no structural constraint or assume a cellular layout of independent rooms and regions, and hence cannot be extended to the atrium case. We propose SiteBuilder, a novel sitemap-grounded framework to complete the 3D atrium scene. SiteBuilder first transforms the sitemap into a bootstrap mesh to extract the geometry and appearance information, based on which along with the key frames rendered from the incomplete 3D model (if any), it then completes the scene using conditional video diffusion. We conduct extensive experiments on many complex atrium sites, and show that SiteBuilder, through integrated and well-guided generation, produces realistic complete 3D scenes that are globally consistent and structurally correct. \endabstract

Abstract:
Deep hashing is appealing for cross-modal retrieval because of its low storage cost and efficient Hamming-distance search. However, practical settings remain challenged by cross-modal semantic gaps, partial semantic overlap in multi-label data, and semantic ambiguity under incomplete correspondence. To address these issues, we propose Proxy-Guided Masked Contrastive Alignment for Cross-Modal Hashing (PGMCA). PGMCA first constructs complementary semantic representations at two granularity levels: a global anchor pathway captures coarse instance-level semantics, while a concept-aware pathway selectively aggregates tokens into fine grained concept units within a shared concept space. On top of these representations, PGMCA introduces masked concept consistency, which constructs independently masked views from the joint visual--textual concept pool and enforces consistency across them, improving robustness under partial concept observations. To further tighten cross-modal correspondence, a semantic coupling objective aligns image--text pairs at both the global instance level and the local concept level. Finally, trainable semantic proxies provide explicit category-level supervision to the global semantic representations by pulling samples toward relevant categories and suppressing similarity to irrelevant ones, thereby improving discriminative hash learning for multi-label data. Experiments on three public benchmarks show that PGMCA achieves the highest reported mAP across the evaluated settings, demonstrating its effectiveness for large-scale cross-modal retrieval. The code will be released.

Abstract:
Multi-task learning (MTL) with parameter-efficient fine-tuning (PEFT) is fundamentally bottlenecked by the quality of the shared representations. Existing pipelines predominantly rely on models pretrained with classification objectives, which introduce source-task bias and fail to provide universally transferable representations. This has driven a shift toward masked autoencodingbased pretraining for more general representations. However, despite their strong reconstruction capability, such methods still suffer from dimensional collapse, leading to semantically incomplete or poorly structured feature spaces. This issue is further exacerbated during parameter-efficient adaptation, where uniform updates tend to disrupt principal components while failing to reactivate dormant capacities. To address this, we propose a unified framework that systematically enhances the representations for MTL from both pretraining and adaptation perspectives. First, we introduce Semantically Grounded Pretraining (SGP), which couples masked autoencoding with frozen DINOv3 distillation to jointly enforce reconstruction fidelity and semantic consistency, producing well-organized and semantically rich representations while alleviating dimensional collapse. Second, we propose Covariance-aware Subspace Routing LoRA (CoSR-LoRA), a PEFT method that explicitly decomposes feature spaces into a head subspace for preserving dominant pretrained knowledge and a tail subspace for task-specific specialization via low-rank adaptation. This design further mitigates dimensional collapse and maximizes the utilization of pretrained representations in multi-task settings. Extensive experiments show that our approach not only establishes a strong, general-purpose backbone but also enables efficient and effective adaptation, achieving state-of-the-art performance across diverse downstream tasks.

Abstract:
Retraining neural network models for every domain is computationally expensive and often impractical. Existing methods typically rely on transfer learning for cross-domain adaptation, but often encounter either high computational overhead in full fine-tuning or insufficient adaptation in parameter-efficient fine-tuning (PEFT) under distribution shifts. In this work, we observe that features contain generalizable components that can be reused to reduce trainable parameters, as well as shift-specific variations that require targeted adaptation under distribution shifts. Accordingly, we propose Feature Distribution Shift--Guided Efficient Transfer Learning (FDT)an automated, data-driven adaptation framework for efficient transfer under distribution shifts. Based on source--target representation similarity, FDT automatically decides whether to freeze, adapt, or fine-tune for each layer. For CNN backbones, we further introduce a lightweight Distribution-Shift Adapter (DS Adapter) together with an Adaptive Weight Module (AWM) for efficient intermediate adaptation. Extensive experiments on CIFAR-10-C, ImageNet-C and Office-Home demonstrate that FDT consistently outperforms both full fine-tuning and recent PEFT methods. On ImageNet-C, FDT improves Top-1 accuracy over full fine-tuning by 2.45% while reducing trainable parameters by 84%; averaged over popular PEFT baselines (Conv-Adapter, SSF, Mona), it further improves accuracy by 8.68% with only about 5% additional trainable parameters. Additional validation on ConvNeXt-S and ViT-B/16 further shows that the proposed similarity-guided layer-wise strategy remains effective on both modern convolutional and Transformer backbones.

Authors: Minghang Zhu (Fujian Key Laboratory of Urban Intelligent Sensing and Computing), Zhijing Wang (Fujian Key Laboratory of Urban Intelligent Sensing and Computing), Yuxin Guo (Fujian Key Laboratory of Urban Intelligent Sensing and Computing), Chen Liu (Fujian Key Laboratory of Urban Intelligent Sensing and Computing), Yongshu Huang (Fujian Key Laboratory of Urban Intelligent Sensing and Computing), Wen Li (University of Bristol), Sheng Ao (Fujian Key Laboratory of Urban Intelligent Sensing and Computing), Cheng Wang (Fujian Key Laboratory of Urban Intelligent Sensing and Computing)

Abstract:
LiDAR relocalization aims to estimate the global 6-DoF pose of a sensor in the environment. However, existing regression-based approaches often encounter limitations in dynamic or ambiguous scenarios, as they typically prioritize single-frame inference, leaving the potential of spatio-temporal consistency across scans not fully explored. In this paper, we propose a Temporal-aware Localization framework (TempLoc) designed to enhance the robustness of outdoor localization by effectively modeling sequential consistency. Specifically, a Global Coordinate Estimation module is first introduced to predict point-wise global coordinates and associated uncertainties for each LiDAR scan. A Prior Coordinate Generation module is then presented to estimate inter-frame point correspondences by the attention mechanism. Lastly, an Uncertainty-Guided Coordinate Fusion module is deployed to integrate both predictions of point correspondence in an end-to-end fashion, yielding a more temporally consistent and accurate global 6-DoF pose. Experimental results on the NCLT and Oxford RobotCar benchmarks show that our TempLoc outperforms state-of-the-art methods by a large margin, demonstrating the effectiveness of temporal-aware correspondence modeling in LiDAR relocalization.

Abstract:
\beginabstract Infrared and visible images provide complementary cues for low-light image generation, and physics-based methods can better exploit this complementarity. However, existing physics-based methods still rely heavily on paired data. To address these limitations, we propose Multi-View Routing Mixture-of-Experts (MVR-MoE), a framework for physically consistent TeX generation that learns multi-view structural representations. MVR-MoE mines intrinsic structural cues from infrared and visible images to organize them into multi-view representations, while the proposed multi-view hyper connections expand the feature space and preserve both shared evidence and modality-specific details. A hierarchical routing mechanism further regulates feature utilization at both the structural-view level and the expert level, enabling adaptive exploitation of heterogeneous cues. The predicted temperature, emissivity, and texture maps are constrained by the Stefan-Boltzmann law to ensure physically consistent. Extensive experiments on challenging datasets demonstrate that MVR-MoE achieves superior robustness and discriminability while remaining lightweight. \endabstract

Abstract:
Federated multi-view clustering (\FedMVC) has proven to be highly effective in exploring consistency information from heterogeneous multi-view data. However, most existing \FedMVC methods overlook the disparities in the types and quality of views held by different clients, resulting in difficulties for partial-view clients to participate in cross-client collaborative training. Moreover, these methods impose overly stringent alignment constraints on the representations of the same sample across different views at the local client, thereby neglecting the inherent differences among views. Therefore, we propose \Method, a novel \FedMVC framework that enhances collaborative learning in practical scenarios with partial-view imputation. (1) To relax rigid cross-view constraints, a cluster-level alignment strategy guided by pseudo labels is designed to reduce alignment enforcement on high-confidence samples. (2) A low-level semantics generator is proposed and trained on full-view latent features. This mitigates bias in learned embeddings and imputes missing representations for partial-view clients. (3) \Method establishes a global latent subspace based on aggregated feature covariances, serving as a unified semantic guidance that enables partial-view clients to actively engage in collaborative learning. Extensive experimental results validate that \Method exhibits superior performance in federated multi-view clustering.

Abstract:
Sound source localization (SSL) estimates the positions of active acoustic sources, but conventional SSL is query-agnostic and does not decide which source a user actually wants. We study query-conditioned binaural localization, where the target is specified by a text query, an audio query, or both, and the goal is to identify and localize the queried source over time. The key challenge is that query semantics and spatial geometry are not directly aligned, so target-relevant acoustic evidence should be made explicit before spatial reasoning rather than left implicit in an entangled mixture. We therefore propose Q2-Loc, a unified framework that combines the Cross-Modal Completer, query-conditioned extraction, Spectral Relevance Scoring, and grouped audio tokenization before shared Qwen-based reasoning with explicit direction-of-arrival (DoA) and cardinality heads. Given a binaural mixture, Q2-Loc predicts a frame-level DoA heatmap, frame-level source cardinality, and a short text answer. Experiments on QueryLoc-Sim and real recordings from the TAU Spatial Room Impulse Response Database (TAU-SRIR DB) show consistent gains over adapted localization and spatial-audio large language model (LLM) baselines. Ablations, missing-modality tests, and token-region masking further show that explicit query-conditioned grounding is essential for robust spatial reasoning.

Abstract:
Incomplete multi-view clustering (IMVC) seeks to uncover coherent cluster structures from multi-view data in the presence of missing observations. A central challenge is that the available views are often unreliable: heterogeneous noise and view-specific distortions can render paired samples weakly informative, making conventional cross-view alignment and prototype transfer prone to negative transfer. To address this issue, we propose Reliability-Gated Incomplete Multi-View Clustering with Entropy-Adaptive Prototypes (RGEA-IMVC), a deep framework that jointly performs uncertainty-driven reliability estimation, reliability-weighted semantic regularization, and prototype-based completion. RGEA-IMVC leverages view-specific autoencoders with heteroscedastic reconstruction uncertainty to infer instance-level reliability from predicted precision, incorporates a reliability-weighted energy-based semantic alignment module that learns cluster-specific energy functions and suppresses intra-cluster energy dispersion to promote semantic compactness while attenuating unreliable instances, enforces cross-view consistency of instance-to-prototype assignment distributions via a reliability-gated constraint with a balancing regularizer, and imputes missing-view representations through a prototype-driven completion module with entropy-adaptive soft prototype mixing to remain effective under severe incompleteness. Experiments on six benchmarks across a range of missing rates demonstrate that RGEA-IMVC consistently outperforms recent IMVC methods and retains robust performance as missingness increases.

Abstract:
Parameter-Efficient Fine-Tuning (PEFT) has become a promising paradigm for Remote Sensing Image-Text Retrieval (RSITR), enabling the adaptation of large pre-trained models with a small number of trainable parameters. However, remote sensing image-text pairs exhibit heterogeneous structures: compositional caption representations entangle entities, attributes, relations, and quantities, while overhead images contain substantial background clutter. To address these challenges, we propose Dynamic Fine-Grained Perceptive Bidirectional Image-Text Retrieval (DFPR), a PEFT framework tailored for RSITR that introduces structured semantic routing and entity-guided visual focusing during training. To alleviate this semantic entanglement, we introduce a multi-structured mixture-of-experts (MS-MoE). Its dual-gating mechanism adaptively routes text representations to specialized experts, injecting fine-grained domain priors to disentangle caption semantics into explicit components and yield a dedicated entity anchor. Conditioned on this anchor, spectral graph-guided focusing (SG^2F) suppresses background interference through spectral filtering and aggregates salient visual features. To mitigate the inherent optimization imbalance between image-to-text and text-to-image retrieval, a bidirectional calibration constraint mechanism (BCCM) dynamically calibrates the directional loss weights using online feedback and exponential smoothing, promoting more stable and balanced optimization. Extensive experiments on three benchmarks demonstrate that DFPR consistently outperforms the compared PEFT baselines in retrieval accuracy.

Abstract:
Text-guided human motion generation has gained significant interest due to its extensive applications in various fields, such as filming, animation and robotics. In this paper, we are interested in a specific form of text-to-motion generation, referred to as storytelling dual-character motion synthesis, which aims to generate the motions of two characters in an interactive scenario to convey the semantics of a given story described by a text prompt. The story descriptions are predominantly composed of abstract, vague concepts, which distinguishes the storytelling motion synthesis from conventional text-to-motion generation that largely relies on concrete action instructions. To solve this task, we present a diffusion-based model that is able to comprehend abstract story text and map it to compelling two-character animations aligned with the input stories. Our model utilizes a novel architectural design that captures intricate interactions between two characters across multiple temporal scales, enabling the generation of highly expressive, storytelling character motions. For training our model, we construct a large-scale two-character motion dataset, consisting of expressive motion sequences of two interacting characters, alongside story-oriented textual descriptions that cover the essential factors of storytelling. Our experiments on the new dataset show that our method can generate the realistic dual-character motions that faithfully convey the given textual stories, significantly outperforming alternative baselines across various metrics. Moreover, we demonstrate the uniqueness and necessity of our dataset for the area of storytelling motion generation of multiple characters, as compared to exiting human motion datasets.

Abstract:
Natural language-guided drone geolocalization aims to identify the location depicted in a drone-view image from a free-form textual description. This task requires fine-grained vision-language understanding, as free-form descriptions often involve complex compositional semantics, with multiple textual elements corresponding to different visual regions. Existing methods typically construct the image representation either from language-agnostic global features or through caption-level text conditioning, neither of which effectively captures fine-grained correspondences between individual textual elements and their associated visual regions. To address this limitation, we propose STARE, a SpaTially-calibrAted token-to-REgion representation that models description-specific, fine-grained correspondences between textual units and image regions. Specifically, multiple text-derived tokens independently attend to image patch tokens, each capturing spatially distinct visual cues, and the resulting token-grounded features are composed into a compact representation for matching. To ensure that each token grounds to its correct spatial region, bounding-box annotations provide explicit region-level anchors that calibrate the grounding process. On the GeoText-1652 benchmark, STARE achieves strong performance, attaining Recall@1 of 30.2% for image-to-text retrieval and 14.7% for text-to-image retrieval, while delivering 16× faster training and 11× faster inference than prior methods. It also generalizes effectively across standard, fine-grained, and long-text zero-shot imagetext retrieval benchmarks, demonstrating the broader potential of token-to-region representations.

Abstract:
Achieving high-fidelity generation in extremely few sampling steps has long been a central goal of generative modeling. Existing approaches largely rely on distillation-based frameworks to compress the original multi-step denoising process into a few-step generator. However, such methods inherently constrain the student to imitate a stronger multi-step teacher, imposing the teacher as an upper bound on student performance. We argue that introducing preference alignment awareness enables the student to optimize toward reward-preferred generation quality, potentially surpassing the teacher instead of being restricted to rigid teacher imitation. To this end, we propose Reward-Aware Trajectory Shaping (RATS), a lightweight framework for preference-aligned few-step generation without the need of additional data. Specifically, teacher and student latent trajectories are aligned at key denoising stages through horizon matching, while a reward-aware gate is introduced to adaptively regulate teacher guidance based on their relative reward performance. Trajectory shaping is strengthened when the teacher achieves higher rewards, and relaxed when the student matches or surpasses the teacher, thereby enabling continued reward-driven improvement. By seamlessly integrating trajectory distillation, reward-aware gating, and preference alignment, RATS effectively transfers preference-relevant knowledge from high-step generators without incurring additional test-time computational overhead. Experimental results demonstrate that RATS exhibits strong few?step generation capability for visual generation while achieving a favorable efficiencyquality trade?off.

Abstract:
Multimodal Large Language Models (MLLMs) rely on a projector to align visual representations with the language embedding space, making it central to cross-modal understanding. In Multimodal Continual Instruction Tuning (MCIT), however, shifting visual distributions and evolving instruction semantics cause this shared projector to drift, leading to projector-level forgetting, an issue largely overlooked by methods that focus primarily on the LLM backbone. We introduce Progressive Multimodal Alignment (PMA), a framework that enables the projector to adapt continually while preserving previously learned alignment. PMA detects multimodal distribution shifts via a lightweight representation descriptor and progressively expands projector experts only when needed. An expandable router integrates expert outputs based on multimodal features, while the original pretrained projector is retained as a stable alignment anchor. This progressive mechanism balances stability and plasticity with sub-linear parameter growth and serves as a method-agnostic add-on to existing MCIT approaches. Extensive experiments on two recent MCIT benchmarks demonstrate that mitigating projector-level forgetting yields consistent gains over prior state-of-the-art methods when combined with PMA. Moreover, PMA scales across diverse MLLM backbones, demonstrating robust and broadly applicable MCIT performance.

Abstract:
Long-tailed data distributions remain challenging for deep learning: over-represented head classes dominate optimization and suppress learning for tails. This imbalance is not only induced by skewed class frequencies, but also progressively reinforced by training dynamics, leading to class-wise norm disparities and degraded generalization. Existing prior-based methods mitigate static distributional skew via re-sampling or re-weighting, yet they often under-address the dynamic amplification of imbalance that emerges during training. We propose Synergistic Adaptive Training (SAT), a lightweight online training layer that dynamically regulates learning exposure and regularization, complementing prior-based approaches without altering base objectives or static augmentation policies. SAT comprises two cooperative modules: Difficulty-Aware Sampling (DAS), which aggregates instance-level difficulty into class-level priorities to rebalance effective updates; and Exposure-Aware Augmentation (EAA), which adapts augmentation probability and strength based on exposure and prediction confidence to enhance feature-weight alignment and reduce shortcut reliance. Experiments on various datasets show consistent gains under matched computational budgets, improving both head and tail performance. It curbs the dynamic amplification of norm imbalance, stabilizes class-wise representations, and serves as a near-zero-overhead, simple complement to existing long-tailed learning strategies.

Abstract:
Transformer-based detectors have transformed the detection paradigm by leveraging the concept of object queries and set prediction. This end-to-end paradigm, while simplifying the detection pipeline, also introduces additional adversarial vulnerabilities. Prior robustness research has primarily focused on classifiers or CNN-based detectors, in which task heads directly operate on visual features to decode object categories or locations, making the perturbation of feature extraction the primary objective of adversarial attacks. In contrast, we propose query-based attacks specifically targeting the transformer architecture's query-based detectors, in which visual features merely serve as queryable items, with information being decoded through the iterative refinement of object queries. Our core idea is to manipulate the content and positional information embedded in the output queries. Specifically, we induce false negatives by disrupting benign alignments between content representations and class prototypes, and generate false positives by constructing malicious alignments. Furthermore, by preserving the positional representations, we enable location-controllable false positives, thereby getting adversarial results that conform to the prior distribution. The proposed method exhibits strong generality, applying to both unimodal detectors and the text-guided multimodal open-set detector. It exhibits excellent orthogonality and can be effectively combined with existing strategies to enhance cross-model transferability of perturbations. Experimental results confirm that our approach successfully achieves the intended adversarial detection results.

Abstract:
Unified large multimodal models, capable of both understanding and generation, face significant challenges during continued training without access to original pre-training data. This post-pre-training task, critical for enhancing core foundational capabilities, is plagued by catastrophic forgetting and compounded by task interference inherent in unified architectures. To address this, we propose MEANING, a novel model that unifies sparse adaptation with targeted knowledge distillation. We introduce a Dual Layer-Level LoRA Routing mechanism, where modality-specific routers dynamically activate a sparse, input-dependent subset of LoRA adapters to mitigate task interference. Crucially, this sparse adaptation is coupled with a novel Layer-Routed Attention Distillation loss, which aligns internal reasoning patterns exclusively within these active layers. This integrated approach simultaneously teaches the model where to adapt and what to learn. Experiments show MEANING prevents the catastrophic forgetting seen in naive continued training (\eg, an average 60.1% performance drop across understanding benchmarks) and successfully enhances both faculties, which has been verified across multiple tasks spanning understanding, generation, and editing.

Abstract:
Despite progress in Embodied AI, Vision-and-Language Navigation (VLN) systems remain vulnerable to adversarial visual disturbances. Most existing methods rely on white-box access to target model gradients, which is often unrealistic for real-world deployed systems and computationally exhaustive due to recursive backpropagation for optimization, limiting their applicability. While previous black-box methods predominantly target single-step, instantaneous decision tasks, they struggle to handle the task complexities and temporal dependencies of VLN. This highlights the need for a gradient-free attack method that can effectively disrupt the multi-step sequential perceptionaction loop using only observable inputs and outputs. Therefore, we propose AdvNav, a behavior-guided black-box adversarial attack framework that disturbs an agents first-person views during navigation. To construct an informative surrogate objective for effective optimization guidance in gradient-free search under the black-box setting, we design a dual-granularity behavior-based feedback, aggregating a trajectory-level performance score representing overall navigation degradation, an action-level reward score considering the potential decision risk, and a deviation indicator, all of which are extracted from the agent's self-output behaviors. This feedback guides a hybrid optimization strategy that (i) heuristically tunes perturbation strength via adaptive updates and (ii) evolves noise spatial structure genetically, to iteratively discover the most disruptive noise configuration. Evaluated against Transformer-based model HAMT and LLM-based model MapGPT with two types of backbones on R2R dataset, AdvNav achieves 49.70%, 65.96%, and 87.30% Attack Success Rate, respectively. The result demonstrates the effectiveness and generality of AdvNav, reveals critical perception vulnerabilities and offers insights for the design of future resilient VLN models.

Abstract:
Embodied agents operating in immersive environments often rely on incomplete egocentric observations, where the current view is insufficient to infer the underlying scene dynamics and make reliable decisions. World models learn compact latent representations to predict future dynamics and rewards, enabling sample efficient policy optimization. However, in partially observable embodied environments, the same observation-action pair may correspond to different hidden states, leading to ambiguous transition and reward dynamics from the agents perspective. We present BRAVE (Belief Representation for pArtially obserVable Environments), which constructs latent belief representations by integrating the current observation with informative visual cues retrieved from historical memory. Specifically, BRAVE first selects representative key frames from past experience to preserve both relevance and temporal diversity, and then fuses them to construct a more informative belief state for dynamics prediction and downstream decision making. By explicitly modeling belief construction as a retrieval-and-fusion process over egocentric visual memory, BRAVE improves long-horizon prediction and policy learning in partially observable settings. We evaluate BRAVE on diverse embodied decision-making benchmarks, including immersive navigation, open-ended interaction, and multi-agent coordination tasks. Experimental results demonstrate that BRAVE consistently outperforms strong baselines in both sample efficiency and final performance.

Abstract:
Large vision-language models (LVLMs) exhibit remarkable capabilities in cross-modal tasks but face significant safety challenges, which undermine their reliability in real-world applications. Efforts have been made to build LVLM safety evaluation benchmarks to uncover their vulnerability. However, existing benchmarks are hindered by their labor-intensive construction process, static complexity, and limited discriminative power. Thus, they may fail to keep pace with rapidly evolving models and emerging risks. To address these limitations, we propose VLSafetyBencher, the first automated system for LVLM safety benchmarking. VLSafetyBencher introduces four collaborative agents: Data Preprocessing, Generation, Augmentation, and Selection agents to construct and select high-quality samples. Experiments validate that VLSafetyBencher can construct safety benchmarks within one week at a minimal cost. The generated benchmark effectively distinguishes safety, with a safety rate disparity of 70% between the most and least safe models.

Abstract:
In multimedia synthesis, high-fidelity 3D human motion editing with local control remains difficult. Current diffusion-based methods often suffer from both temporal and spatial artifacts due to complex nonlinear dependencies. We propose Temporally-Aware Noise Modulation (TANM), a novel framework decoupling spatial boundary optimization from temporal noise scheduling. Specifically, we introduce the Velocity-Adaptive Gaussian Soft Mask (VAGSM) mechanism, which leverages velocity dynamics in the spatial dimension to calibrate the bandwidth of transition regions, thereby ensuring the smoothness of editing boundaries at the dynamic level. Complementing this, and derived from first principles, we formulate a Dynamic Noise Attenuation (DNA) mechanism based on the instantaneous Signal-to-Noise Ratio (SNR); this mechanism aims to adaptively decouple generative degrees of freedom from structural preservation throughout the entire diffusion process. TANM's fine-grained spatiotemporal modulation reduces distribution shifts. Evaluations on HumanML3D and KIT-ML datasets demonstrate its performance across multiple core metrics.

Abstract:
Class-incremental learning (CIL) aims to adapt to new classes sequentially while avoiding forgetting of previously learned knowledge. Due to their strong generalization ability and rich prior knowledge, pre-trained models (PTMs) have been applied to CIL. Recent PTM-based methods often disentangle features with task-shared and task-specific modules to enable effective continual adaptation. However, our study reveals that such dichotomous disentanglement overlooks features that cannot be captured by either module alone, i.e., task-oriented features, which limits the models ability to fully exploit the representational capacity of PTMs. Additionally, the task-specific modules end up acquiring only dominant patterns that are sufficient to make decisions on seen classes, overlooking non-dominant counterparts crucial for discriminating between seen and forthcoming classes. To address the above issues, we propose the two-level COmplementary REpresentations (CORE) for PTM-based CIL. At the task level, CORE learns task-oriented features as complementary representations between task-shared and task-specific modules, thereby bridging the gap between them. At the class level, we introduce Class-Specific Steering Vectors to complement class-discriminative representations. Extensive experiments demonstrate that the proposed CORE outperforms state-of-the-art PTM-based CIL methods, achieving higher accuracy while effectively mitigating forgetting.

Abstract:
RGB-IR multi-object tracking (MOT) typically assumes spatially aligned visible and thermal observations. In practice, distinct imaging geometries, imperfect calibration, and synchronization errors introduce cross-modal misalignment. Using M3OT as a testbed, we show that under large and temporally unstable disparity, conventional feature-level fusion, decision-level fusion, and tracker-side score modulation strategies fail to yield consistent tracking gains; detector-side interaction does not reliably improve temporal tracking and can degrade it instead. To address this, we propose DispTrack (Disparity-aware Tracking), which moves cross-modal interaction from spatial detection to temporal association. DispTrack employs a Dual-Decoder RGB-IR Detector, a Relative Position Matcher (RPMatcher) for cross-modal correspondence under spatial offset, and a two-stage TID-assisted association strategy for online association refinement and residual offline consolidation. On M3OT, DispTrack improves the standard IR-only references and remains beneficial under tracker-matched comparisons. These results support a narrower, empirical, and regime-specific conclusion: on M3OT-like large and temporally unstable disparity settings, association-stage integration is more reliable than the detector-side fusion interfaces studied here. \endabstract

Abstract:
Communication topology plays a central role in the effectiveness of large language model (LLM)-based multi-agent systems. Recent hypergraph-based methods improve group coordination by connecting multiple agents within unified hyperedges, enabling direct one-step synchronization. However, they largely assume that stronger synchronization is uniformly beneficial. This assumption overlooks a key reliability dilemma: the same mechanism that accelerates useful insight sharing can also accelerate the spread of erroneous outputs. To address this issue, we conduct a causal analysis of hypergraph communication and show that enlarging hyperedges amplifies both insight propagation and error propagation, but the latter grows faster, causing performance to peak at intermediate collaboration intensity. Motivated by this observation, we propose \ourmethod, a dual-view hypergraph topology learning framework that treats topology design as collaboration-intensity allocation. Specifically, \ourmethod learns how strongly each task and agent pair should favor cautious communication for error containment or intensive collaboration for insight sharing, and instantiates this balance as a task-specific hypergraph topology. Experiments on six benchmarks show that \ourmethod is: (1) high-performing, achieving 91.46% average accuracy and outperforming the strongest topology baseline by 3.56%; (2) communication-efficient, substantially reducing communication overhead through more economical collaboration structures; and (3) broadly effective, yielding consistent improvements across reasoning, mathematical reasoning, real-world agentic tasks, and different language models.

Abstract:
Reconstructing dynamic 3D scenes from monocular video remains fundamentally challenging due to the need to jointly infer motion, structure, and appearance from limited observations. Existing dynamic scene reconstruction methods based on Gaussian Splatting often entangle static and dynamic elements in a shared representation, leading to motion leakage, geometric distortions, and temporal flickering. We identify that the root cause lies in the coupled modeling of geometry and appearance across time, which hampers both stability and interpretability. To address this, we propose SplitGaussian, a novel framework that explicitly decomposes scene representations into static and dynamic components. By decoupling motion modeling from background geometry and allowing only the dynamic branch to deform over time, our method prevents motion artifacts in static regions while supporting view- and time-dependent appearance refinement. This disentangled design not only enhances temporal consistency and reconstruction fidelity but also accelerates convergence. Extensive experiments demonstrate that our approach outperforms prior state-of-the-art methods in rendering quality, geometric stability, and motion separation.

Abstract:
Building interactive simulators and scalable robot-learning environments requires a large number of articulated assets. However, most existing 3D assets used in simulation are rigid, and manually converting them into articulated objects with executable kinematic structures is highly labor-intensive, time-consuming, and expensive. This limitation significantly restricts the scalability of simulator construction and downstream embodied AI research. To address this challenge, we present ArtiWorld, a scene-aware pipeline that automatically localizes rigid objects in 3D environments and converts them into interactive URDF models. At the core of this pipeline is Arti4URDF, a framework that integrates 3D point cloud geometry, the prior knowledge of a large language model (LLM), and a URDF-oriented prompt design to infer articulated structures and generate executable articulated assets while preserving the original 3D shape. By explicitly modeling object parts, geometric relations, and articulation priors, our method enables direct conversion from rigid 3D assets to robot-ready articulated objects. We evaluate our method on simulated articulated objects, full-scene assets, and real-world scanned data. Experimental results show that ArtiWorld achieves state-of-the-art performance and consistently outperforms existing methods in articulation prediction quality and asset usability. Furthermore, the generated assets can be directly deployed in simulation environments and even can be generalized to real-world. Code and data will be released.

Abstract:
Mixture-of-Experts (MoE) architectures provide an effective way to scale model capacity while keeping inference costs manageable. However, training MoE models from scratch typically requires substantial computational resources and often leads to unstable training dynamics. In this work, we present Upcycling CLIP with Shared and Fine-grained Experts, a lightweight framework that upgrades a pretrained CLIP model into a MoE architecture with only a small amount of additional training. Specifically, we introduce a Shared Expert to retain the global representation ability of the original CLIP encoder, while a set of Fine-grained Experts are learned to model diverse visualtextual patterns. We also propose a group-wise expert formulation that ensures structured and balanced activation of fine-grained experts. Extensive experiments on multiple datasets show that our method achieves state-of-the-art average performance on both zero-shot imagetext retrieval and zero-shot image classification tasks. Moreover, when applied as an encoder in downstream models such as LLaVA and Stable Diffusion, the upcycled CLIP improves the performance of these systems. Our results highlight the effectiveness of our method as a practical strategy for enhancing pretrained multimodal models.

Abstract:
Fairness in medical vision is typically assessed through demographic disparities in model performance, but this framing is insufficient for mobile tongue-image diagnosis, where the measurement process itself can distort the visual signal before classification. Illumination, color cast, blur, framing, and partial visibility can interact with demographic variation at input time, making fairness a property of the full acquisition-to-model pipeline rather than of the classifier alone. We present FAME-Bench (Fairness Assessment under Measurement Environments), a measurement-aware fairness benchmark for mobile tongue-image diagnosis. FAME-Bench evaluates intervention families across four complementary regimes spanning in-domain subject-wise evaluation, capture-stress testing, standardized clinical acquisition, and external multimodal scope checking. It further formalizes demographic and style axes, dual checkpoint-selection rules, and shared evaluator outputs, enabling direct audit of ranking changes rather than reliance on a single fixed pipeline. Experiments show that the benchmark does not collapse to a universal winner: deterministic correction, fairness-only optimization, and measurement-aware normalization become preferable under different regimes, metrics, and selection rules. These results establish that, in acquisition-sensitive medical vision, fairness conclusions are inherently regime-dependent.

Abstract:
Machine unlearning methods have been proposed to delete undesired knowledge from Text-to-Image diffusion models (DMs), yet, unlearned models remain vulnerable to adversarial prompts that regenerate erased content. While existing noise-prediction optimization attacks have shown promising results, they generally rely on the intuitive premise that similar noise predictions between the original and unlearned models will naturally yield similar generated images. In this paper, we observe that this noise-level alignment may not fully guarantee image-level consistency, primarily due to cumulative deviations between noise prediction and image reconstruction during the iterative denoising process. Building on this diagnosis, we introduce a latent consistency loss and integrate it into existing optimization-based attack frameworks to bridge this gap. We further examine the vulnerability of unlearned DMs under more strict settings, and propose a gradient-free attacker (ZOOAtk) via zeroth-order stochastic variance reduction to enable effective prompt optimization under limited access. Experimental results demonstrate that latent consistency strengthens attacks in different settings, while ZOOAtk remains effective under restricted access, indicating that current unlearned DMs remain vulnerable even when internal model information is unavailable.

Abstract:
Continuous visual synthesis and cross-domain interpolation demand seamless visual transitions, yet conventional visual transformation approaches often rely on intermediate noise mapping, complex feature manipulation, and cumbersome parameter adjustments. In practical multimedia applications, these limitations frequently lead to compromised fidelity and contextual inconsistencies. To address these system-level challenges, we present a novel, highly efficient framework that formulates cross-source visual connection as direct trajectory modeling between two input distributions. Optimized for real-world deployment, this approach features three core advantages: no intermediate noise inversion, no model fine-tuning, and broad compatibility with pre-trained architectures. To tackle the inherent spatial misalignments in complex visual data, we develop a lightweight, DiT-guided feature alignment strategy to establish spatial consistency without relying on external optical flow networks. Furthermore, to bridge theoretical modeling and practical application, a timestep-adaptive generation strategy and dual-path evolution mechanism effectively mitigate contextual mismatches and suppress visual artifacts. Extensive experiments demonstrate that the proposed approach outperforms state-of-the-art methods across diverse visual tasks, delivering outputs with enhanced perceptual smoothness, consistent semantic preservation, and significantly improved computational efficiency. Our framework retains simplicity while avoiding the pitfalls of traditional methods, providing a robust and accessible tool for practical visual synthesis.

Abstract:
Spatial intelligence, which refers to the ability to reason about geometric and physical structure from visual observations, remains a core challenge for multimodal large language models. Despite promising performance, recent multimodal large language models (MLLMs) often exhibit fragile reasoning traces in spatial intelligence tasks that involve consistent spatial state recognition. We argue that these failures stem from a mismatch between the spatial recognition mechanism and the text-only reasoning behavior of these MLLMs. Effective spatial reasoning requires low-level geometric structure to be faithfully preserved and updated throughout the reasoning process, whereas textual representations tend to abstract away precisely these critical details. To address this issue, we propose SpatialImaginer, a unified multimodal generation framework that integrates textual reasoning with visual imagination. Our framework adopts a divide-and-conquer strategy, using text chain-of-thought for high-level semantic planning and the visual imagination for geometry-sensitive state transformation and consistency preservation. To support this capability, we further introduce a difficulty-aware data engine with closed-loop verification to train the model to invoke visual imagination selectively when stable spatial state tracking is required. Extensive experiments on diverse spatial intelligence benchmarks show that SpatialImaginer achieves state-of-the-art performance and substantially improves robustness on complex multi-step spatial reasoning tasks.

Abstract:
Advancement of Large Vision-Language Models (LVLMs) has enabled comprehension of long and high-resolution multi-modal contexts, but this capability introduces a significant Key-Value (KV) cache memory bottleneck during autoregressive generation. Existing KV cache compression methods typically rely on greedy top-k eviction driven by attention-based importance scores. However, this approach tends to redundantly sample highly similar visual tokens, thereby potentially overlooking the long-tail semantic information. To address this issue, we propose RankKV, an approach that reformulates KV cache compression as a Constrained Column Subset Selection Problem. Rather than heuristically scoring isolated tokens, our approach identifies a compact, mutually complementary subset of original physical tokens that better spans the feature space. RankKV first applies CountSketch-based randomized sketching together with observation-guided attention signals to build a candidate set, and then uses deterministic Pivoted QR factorization to remove redundancy and retain informative tokens. In addition, RankKV dynamically allocates cache budgets across attention heads according to their information capacity. Experiments on six LVLMs across seven benchmarks covering general image understanding, multimodal long-context reasoning, and video understanding show that RankKV achieves competitive performance on long-context tasks and maintains robust multimodal reasoning even under an extreme 1% token retention budget, improving generation quality while maintaining comparable efficiency.

Abstract:
Text-to-audio-video (T2AV) generation has advanced rapidly, but its evaluation still underestimates the audio modality. Existing benchmarks either treat audio as an auxiliary component of video quality or assess it in isolation from audiovisual grounding, making it difficult to diagnose where current systems truly succeed or fail in audio generation. We present PRISM-Bench, the first audio-centric diagnostic benchmark for T2AV generation. Built from a rigorously curated dataset of 900 human-verified samples, PRISM-Bench factorizes audio evaluation along two orthogonal axes: audio type (Speech, Music, and Sound) and sound-source visibility (On-screen vs. Off-screen). It evaluates generated content across four perceptual dimensions (Audio-Visual Coherence, Audio Quality, Audio Expressiveness, and Prompt Following) with 35 fine-grained criteria. To ensure reliable assessment, we adopt an enhanced MLLM-as-a-Judge protocol based on blind, side-by-side comparison against ground-truth references, demonstrating strong alignment (over 70% mean agreement) with human raters. Our evaluation of recent T2AV systems highlights a significant performance gap between frontier and open-source models. Furthermore, we demonstrate that current generation paradigms overfit to perceptual fidelity while struggling with complex grounding and control tasks, particularly in generating music and synchronized On-screen audio.

Abstract:
AI video chat, powered by Multimodal Large Language Models (MLLMs), has emerged as a promising paradigm for real-time human-AI interaction on AIoT devices. However, the prevailing paradigm requires client devices to continuously stream audio and video to cloud-hosted MLLMs, imposing substantial power consumption and bandwidth overhead on the client side while incurring significant inference costs on the server side. We propose Rems, a Resource-Efficient Multimodal Streaming framework that enables intent-aware on-demand video streaming. Rems uses lightweight audio to infer intent and streams video only when visual information is necessary. To address the trade-off between response accuracy and energy consumption, we propose an Energy-aware Intent-based Decision module that adapts its policy to the battery level. To eliminate camera startup latency, we propose a Keywords-triggered Proactive Camera Initialization module that overlaps initialization with MLLM inference. To support heterogeneous AIoT devices, we propose an MCP-based Universal Function-Calling mechanism that enables plug-and-play hardware integration. Experiments on an ESP32-S3 hardware prototype demonstrate that Rems achieves a 37.24% to 63.55% reduction in power consumption, while retaining 97.28% of the baseline response accuracy and reducing end-to-end system latency by 312~ms.

Abstract:
Learning discriminative representations from multi-view data remains challenging because different views exhibit heterogeneous structural patterns across scales. Many existing approaches model local and global relations in a single geometric space, which may limit their ability to jointly characterize fine-grained neighborhoods and non-uniform global organization. To address this limitation, we propose the Euclidean-hyperbolic correlational network, termed ECHO, a cross-scale collaborative learning framework. ECHO comprises a Poincar\'e-informed global perception module that performs non-local structural routing through a compact set of supernodes and a Euclidean local perception module for neighborhood aggregation. Within each view, a learnable scalar gate balances neighborhood information from the Euclidean branch with non-local structural context from the Poincar\'e-informed branch. Consensus-guided attention subsequently refines the resulting view-specific representations, with the mean representation across views serving as a consensus query to selectively integrate complementary information while preserving view-specific contributions. Across eight benchmark datasets, ECHO outperforms the compared methods in semi-supervised node classification in terms of both classification accuracy and Macro-F1. Qualitative analyses suggest that the learned representations exhibit more compact intra-class structure and clearer inter-class separation, supporting ECHO as an effective approach to semi-supervised multi-view node classification.

Abstract:
Humanmachine co-driving is emerging as a critical paradigm in advanced driver-assistance systems (ADAS), where understanding drivers behavioral tendencies is essential for reliable collaboration. Existing studies either model drivers at a coarse-grained level or remain limited to generic scene-level descriptions, lacking the ability to reason about human behaviors from multimodal evidence in a semantically grounded manner. To bridge this gap, we introduce behavior-aware humanmachine co-driving understanding, a new task that uses multimodal observations from driving videos, vehicle dynamics, and driver gaze to generate structured, behavior-aware explanations and actionable decisions. To address this task, we introduce CoDriveMind, a dual-pathway framework inspired by fastslow reasoning. The Fast Pathway predicts co-driving behavioral tendencies through lightweight multimodal fusion and a behavior-anchored latent alignment mechanism, where expert-defined semantic descriptions guide multimodal representations toward semantically grounded behavior concepts. The Slow Pathway performs structured multi-stage reasoning through hypothesis instantiation, cross-modal evidence verification, and evidence-grounded synthesis, suppressing unsupported interpretations and enforcing cross-modal consistency. We evaluate CoDriveMind on a newly constructed multimodal co-driving dataset with expert-annotated behavioral tendencies and structured reports, designed to support both tendency prediction and behavior-aware reasoning. Experimental results show that CoDriveMind consistently outperforms strong baselines on both tendency prediction and behavior-aware reasoning, yielding more accurate decisions and more coherent, evidence-grounded explanations.

Abstract:
Differential diagnosis of mental disorders remains challenging in real-world clinical practice due to substantial symptom overlap across disorders. However, existing public datasets are largely developed for single-disorder settings and rely on limited elicitation paradigms, restricting their ability to capture disorder-specific behavioral patterns. To address this limitation, we introduce a psychology-inspired multimodal elicitation paradigm for differential mental disorder detection and construct a large-scale multimodal mental health dataset (MMH) covering depression, anxiety, and schizophrenia, with all diagnostic labels clinically verified by licensed psychiatrists. We further propose a paradigm-aware multimodal learning framework that leverages prompt-guided semantic descriptions of inter-disorder differences to model task-specific affective and interaction contexts for multimodal representation learning. Extensive experiments demonstrate that the proposed framework consistently outperforms strong baselines, highlighting the effectiveness of psychology-inspired elicitation for differential mental disorder detection.

Abstract:
The rapid progress of 3D generative models has made 3D asset creation increasingly accessible. However, achieving intuitive, precise, and interactive user control, such as point-based drag editing, remains challenging. Existing 3D drag methods either rely on slow 2D diffusion priors, such as Score Distillation Sampling, or directly warp geometry, often causing multi-view inconsistency and topological artifacts. In this paper, we present 3DRAG, a training-free framework that executes precise drag editing natively within the explicit 3D latent space of large-scale generative models, without SDS-based optimization or iterative point tracking. To ensure geometrically plausible deformations, 3DRAG explicitly extends the drag displacement field into 3D space, incorporating an Occupancy-Aware Winner-Takes-All (OA-WTA) strategy to robustly handle sparse geometries. Furthermore, we introduce a generative editing pipeline that explicitly transforms latent representations via the 3D displacement field. By fully leveraging the model's generative capabilities, it achieves smooth geometric deformations while maintaining the object's identity. To further transfer original geometry and texture with high fidelity, we propose Logits-Level Token Fusion, which accurately injects source information into the destination region. To evaluate 3D point-based editing, we construct a novel dataset featuring diverse 3D assets and complex multi-point drag instructions. Quantitative and qualitative experiments demonstrate that 3DRAG achieves state-of-the-art performance, surpassing existing baselines in drag accuracy, geometric fidelity, and unedited region preservation.

Abstract:
Editing videos with data is essential for exploring multimodal datasets, yet existing systems are too slow for interactive applications. "Show me a compilation of every time X happens", "Draw a box over each Y", or "Overlay Z telemetry data on this video". Each task creates a new video by joining source videos with data. Scripts are the only tool expressive enough for these tasks, but their imperative control flow precludes the optimizations that make video editors fast. Scripts take minutes or hours to complete, significantly disrupting interactive workflows. We present Vidformer, which reconciles the expressivity of visualization scripts with the responsiveness of a video editor. Vidformer (i) transparently lifts existing code into a declarative representation, (ii) optimizes and parallelizes rendering, and (iii) instantly serves videos through a Video on Demand protocol with just-in-time segment rendering. This enables sub-second playback (0.250.5 s) for data-edited videos, decoupling playback latency from video length and task complexity, a 400 speedup. When operated by an LLM agent, Vidformer lets users explore multimodal datasets through conversational video queries in real time.

Abstract:
Live immersive video streaming, particularly 360-degree video, is increasingly adopted in applications such as virtual events, sports broadcasting, and remote education. Existing approaches struggle to support high-bitrate immersive streaming for large numbers of concurrent users, with coarse-grained delivery limiting responsiveness and insufficient support for coordinating concurrent tile streams. Media over QUIC (MoQ) has recently emerged as a promising solution for large-scale media delivery, yet it lacks robustness under bandwidth-constrained conditions, often resulting in playback stalls. To address these challenges, we present MultiMoQ, a multi-access tile streaming framework built on MoQ that redesigns its delivery mechanism for robust high-bitrate streaming across multiple access paths while supporting flexible tile scheduling and seamless access switching. We implement a fully functional prototype of MultiMoQ and evaluate it in network emulation under heterogeneous real-world network conditions, comparing against Dynamic Adaptive Streaming over HTTP (DASH) and standard MoQ. Results show that MultiMoQ increases goodput for enhancement tiles and base video and reduces enhancement-tile tail end-to-end latency relative to DASH, while preserving audio continuity and avoiding the persistent stalls of standard MoQ. These transport gains translate into smoother viewport playback, and the ablation results further confirm the contribution of multi-access control to playback continuity.

Abstract:
Multimodal Sewing pattern generation aims to infer executable sewing patterns from design cues such as sketches and textual descriptions. As an interpretable and simulation-compatible representation, sewing patterns are particularly valuable for digital garment creation. However, existing methods often model garment specifications as flat long sequences, which entangles garment structure with detailed parameters and leads to redundant components, inaccurate local details, and poor simulation compatibility. In this paper, we present GarmentWeaver, a schema-aware framework for multimodal Sewing pattern generation. GarmentWeaver constructs compact hierarchical targets by activating garment-relevant structural branches and predicts executable Sewing patterns in a structured manner. Specifically, we introduce a schema-aware target construction strategy, build the generator on top of a pretrained vision-language model for multimodal garment understanding, and impose feasibility-aware regularization to encourage structurally valid and simulation-compatible outputs. Extensive experiments show that GarmentWeaver produces more accurate and more executable sewing patterns than strong baselines, while also yielding better simulation results. These findings demonstrate the effectiveness of schema-aware structured

Abstract:
As large foundation models are increasingly deployed in distributed settings, transmitting intermediate representations has become a bottleneck, making feature coding important for reducing communication and storage overheads. However, intermediate representations from large models often exhibit significant heterogeneity and peaky, heavy-tailed distributions, posing challenges for end-to-end optimization and effective entropy modeling. Existing methods typically rely on handcrafted preprocessing pipelines for feature compression, which limits generalizability and may introduce structural and semantic distortions. Although recent studies introduce learned distribution transformations to mitigate feature heterogeneity, their decoupled design does not support end-to-end joint optimization. To address these issues, we propose Dual-Stream Companding for Structure-Preserving Feature Coding (DSC-SPFC), an end-to-end framework that operates directly on tensor structures. DSC-SPFC decomposes features into a bounded main stream and a complementary residual stream, and maps both into a stable latent space via differentiable companding for joint entropy coding. This design improves robustness to outliers while preserving reversibility and end-to-end learnability. Experiments on LLaMA3, DINOv2, and SD3 across common sense reasoning, semantic segmentation, and text-to-image synthesis show that DSC-SPFC provides improved rate--task trade-offs under the BPFP metric over packing baselines and recent decoupled distribution-transformation methods.

Abstract:
Vision-Language Models (VLMs) inherit significant social biases from their training data, notably in gender representation. Current fairness interventions often adopt a ``difference-unaware'' perspective that enforces uniform treatment across demographic groups. These approaches, however, fail to distinguish between contexts where neutrality is required and those where group-specific attributes are legitimate and must be preserved. Building upon recent advances in difference-aware fairness for text-only models, we extend this concept to the multimodal domain and formalize the problem of difference-aware gender fairness for image captioning and text-to-image generation. We advocate for selective debiasing, which aims to mitigate unwanted bias in neutral contexts while preserving valid distinctions in explicit ones. To achieve this, we propose BioPro (Bias Orthogonal Projection), an entirely training-free framework. BioPro identifies a low-dimensional gender-variation subspace through counterfactual embeddings and applies projection to selectively neutralize gender-related information. Experiments show that BioPro effectively reduces gender bias in neutral cases while maintaining gender faithfulness in explicit ones, thus providing a promising direction toward achieving selective fairness in VLMs. Beyond gender bias, we further demonstrate that BioPro can effectively generalize to continuous bias variables, such as scene brightness, highlighting its broader applicability.

Abstract:
Synthesizing plausible 3D scenes and layouts remains a fundamental challenge in computer graphics. Autoregressive scene generation typically models scenes in an object-wise manner, which suffers from inherent locality and limited global coordination. In this paper, we reformulate autoregressive generation on indoor scene as a top-down, coarse-to-fine autoregressive paradigm and introduce a GPT2-style autoregressive generative model producing plausible and diverse interior scenes with next-scale prediction. We pretrain a multi-codebook residual quantnizer that represents interior scenes as discretized residual tokens of different scales. Given conditional inputs (e.g., room type or textual descriptions), our transformer decoder causally predicts the discrete tokens of the next scale and synthesizes the final scene. The presented model predicts various attributes of 3D instances, including class, translation, orientation, size, and shape, retrieving the most compatible 3D asset from the database. Extensive experiments demonstrate that our method outperforms prior autoregressive approaches to scene generation and achieves superior performance compared to diffusion-based methods.

Abstract:
While diffusion models have made significant progress in text-to-image tasks, they still exhibit limitations when directly optimizing downstream objectives. Although Reinforcement Learning (RL) enables targeted optimization, existing methods are generally constrained by low-efficiency fine-tuning and sparse rewards. To address these challenges, we propose PAST, which provides differentiated rewards while adaptively regulating training episode length by jointly perceiving denoising progress and prompt difficulty. Specifically, we design an intrinsic reward paradigm to compensate for sparse extrinsic rewards and guide the model to explore paths that diverge more efficiently from noise patterns. We further provide theoretical justification for intrinsic rewards. Then, PAST dynamically monitors denoising completion and semantic alignment between image structures and prompt semantics. When both metrics satisfy generation requirements, the system adaptively terminates training. This enables appropriate allocation of episode lengths based on prompt difficulty and the current generation process. Finally, based on the predicted residual noise level, we establish a dual adaptive coordination mechanism. Specifically, it not only balances the extrinsic and intrinsic rewards but also balances the exploration and convergence. Experimental results demonstrate that PAST enhances computational efficiency of existing RL fine-tuning methods by up to 66.7%, while improving preference optimization quality by up to 29.5% through its dual adaptive regulation mechanism.

Abstract:
Domain Generalization Semantic Segmentation (DGSS) focuses on generalizing knowledge from labeled source domains to unseen target domains where data is unavailable during the training phase. While conventional methods utilize style randomization or feature normalization to mitigate domain shifts, they often impair feature integrity. Specifically, style randomization distorts the underlying feature manifold due to its coarse-grained nature, while feature normalization suppresses discriminative, domain-sensitive semantic details owing to its rigid design. To address these limitations, we propose the Language-and-Source-Anchored Alignment (LASA) framework, which comprises three synergistic components: Text-and-Source-Guided Style Transfer (TSGST), Domain-Aware Query Adapter (DAQA), and Domain-Aware Decoder Optimizer (DADO). Concretely, the TSGST module addresses manifold distortion by utilizing source features as structural anchors and vision-language model (VLM) priors as fine-grained guidance. To restore suppressed discriminative and domain-sensitive details, the DAQA module recalibrates object queries via categorical guidance and domain-aware signatures, while the DADO module aligns the resulting query distributions with a shared classifier to ensure consistent categorical responses across domains. Extensive experiments on challenging benchmarks demonstrate that our method significantly outperforms state-of-the-art approaches.

Abstract:
Inspired by the remarkable success of autoregressive models in language modeling, this paradigm has been widely adopted in visual generation. However, the sequential token-by-token decoding mechanism inherent in traditional autoregressive models leads to low inference efficiency. In this paper, we propose RadAR, an efficient and parallelizable framework designed to accelerate autoregressive visual generation while preserving its representational capacity. Our approach is motivated by the observation that visual tokens exhibit strong local dependencies and spatial correlations with their neighborsa property not fully exploited in standard raster-scan decoding orders. Specifically, we organize the generation process around a radial topology: an initial token is selected as the starting point, and all other tokens are systematically grouped into multiple concentric rings according to their spatial distances from this center. Generation then proceeds in a ring-wise manner, from inner to outer regions, enabling the parallel prediction of all tokens within the same ring. This design not only preserves the structural locality and spatial coherence of visual scenes but also substantially increases parallelizability. Furthermore, to address the risk of inconsistent predictions arising from simultaneous token generation with limited context, we introduce a nested attention mechanism. This mechanism dynamically refines implausible outputs during the forward pass, thereby mitigating error accumulation and preventing model collapse. By integrating radial parallel prediction with dynamic output correction, RadAR reduces the number of inference steps from 256 to only 13, achieving up to a 5.6 speedup on the ImageNet dataset and significantly improving generation efficiency.

Abstract:
Action recognition in complex scenes often involves multiple concurrent fine-grained actions, making it challenging to model internal action structures. Most existing methods rely on holistic representations, which are insufficient for capturing subtle interactions and fine-grained semantics. While recent prompt-based approaches introduce disentanglement, they lack explicit semantic guidance, and methods based solely on visual or structured cues remain coarse-grained. In this paper, we propose Knowledge-guided Disentanglement with Atomic Actions (KDA), which leverages fine-grained semantic knowledge to enhance action representations and enable more precise disentanglement. Specifically, we use Large Language Models (LLMs) to decompose action labels into atomic actions, providing explicit spatial-temporal semantics. A Knowledge Injection Module (KIM) first integrates atomic action knowledge into video features. Based on this enhanced representation, a Knowledge Disentanglement Module (KDM) further disentangles atomic action knowledge to produce more precise semantic guidance for action disentanglement. A Knowledge Disentanglement Loss (KD Loss) is introduced to encourage clearer disentanglement of knowledge components within KDM. Extensive experiments demonstrate that KDA improves feature discriminability and achieves state-of-the-art performance on multi-label action recognition benchmarks. Moreover, KIM and KDM can be readily integrated into other methods, demonstrating strong generality.

Abstract:
\beginabstract In real-world non-stationary environments, deep models face performance degradation due to sudden or gradual distribution shifts in unlabeled data streams, motivating continual domain shift learning (CDSL). Existing methods focus solely on model architecture modifications and pseudo-labeling strategies, yielding limited improvements. Mimicking the iterative polishing manner through peer review and feedback in article writing, we propose a novel peer-review self-correction framework that utilizes a Multimodal Large Language Model (MLLM) as an expert reviewer and designs a reinforcement learning (RL) policy for automatic correction. Our framework alternates between two phases: initialization adaptation and self-correction refinement. The first phase enables effective initial adaptation by separating shift-sensitive and shift-agnostic learning, allowing the model to quickly capture domain changes while mitigating catastrophic forgetting. The second phase formulates error correction as a RL process, which enables to iteratively refine predictions guided by MLLM-generated rewards that convert recognition capabilities into optimization signals. Our paradigm achieves effective adaptation without noisy pseudo-labeling and domain ID requirement while maintaining low parameter overhead. Extensive experiments across three benchmarks demonstrate that our method consistently outperforms state-of-the-art approaches, with improvements driven by the proposed peer-review self-correction paradigm rather than raw MLLM recognition ability, underscoring its effectiveness and robustness. \endabstract

Abstract:
Building a generalized affordance grounding model is vital for real-world applications. Existing methods to train the model can be divided into weakly and fully supervised ways. However, the former method requires a complex training framework design and can not infer new actions without an auxiliary prior. While the latter often struggle with limited annotated data and components trained from scratch despite being simpler. This study focuses on fully supervised affordance grounding and overcomes its limitations by proposing AffordanceSAM, which extends SAM's generalization capacity in segmentation to affordance grounding. Specifically, we design an affordance-adaption module and curate a coarse-to-fine annotated dataset called C2F-Aff to thoroughly transfer SAM's robust performance to affordance in a three-stage training manner. Experimental results confirm that AffordanceSAM achieves state-of-the-art (SOTA) performance on the AGD20K benchmark and exhibits strong generalized capacity in other scenarios.

Abstract:
Gait recognition methods typically operate within a stationary domain, predetermined by a single dataset. However, in real-world contexts, those methods are ineffective because the environment is continually changing, in which case incremental learning over multiple domains is essentially required. In this work, we explore a new and challenging gait recognition task, namely Lifelong Gait Recognition (LGR), which requires the model to learn continuously across multiple domains and generalize to both seen and unseen domains. To tackle this challenge, we design a Prototype-guided Low-Rank Adaptive Distillation (PLoRA) framework that consists of two crucial modules: the domain-specific LoRA module and the prototype-guided topology distillation module. Our method effectively alleviates catastrophic forgetting in seen domains and enhances the generalization across unseen domains. In addition, we provide a new and large-scale benchmark for LGR. Extensive experiments demonstrate that our method achieves state-of-the-art performance, outperforming the joint-training baseline in terms of Mean Rank-1 accuracy on CCPG as well as Rank-1 accuracy under the NM condition on CASIA-B.

Abstract:
Knowledge distillation for image restoration typically aligns intermediate features or relation matrices between teacher and student networks as static targets, ignoring the dynamic structure of the knowledge transfer process. In this paper, we propose Flow-Map Distillation on Relation Manifolds (FoRM), which reformulates relation-based knowledge transfer as a continuous flow mapping problem on the relation manifold. Rather than regressing a constant velocity field between student and teacher relation states, FoRM learns a flow map operator F_theta(z, t, s) that directly predicts the relation state at any target time s given the current state at time t, enabling richer trajectory-level supervision. To ensure global self-consistency of the learned flow map, we introduce a safe semigroup consistency constraint that enforces compositional agreement using ground-truth bridge states, eliminating phantom-state error accumulation. An endpoint anchoring loss further prevents the operator from drifting away from the teacher target. Extensive experiments on five image restoration tasks, including super-resolution, deraining, denoising, deblurring, and low-light enhancement, demonstrate consistent gains over state-of-the-art distillation baselines across multiple backbone architectures, reducing training variance by approximately 50% compared to naive flow matching distillation while achieving superior restoration quality.

Abstract:
Detecting video deepfakes has become increasingly urgent in recent years. Given the audio-visual information in videos, existing methods typically expose deepfakes by modeling cross-modal correspondence using specifically designed architectures with publicly available datasets. While they have shown promising results, their effectiveness often degrades in real-world scenarios, as the limited diversity of training datasets naturally restricts generalizability to unseen cases. To address this, we propose a simple yet effective method, called AVPF, which can notably enhance model generalizability by training with self-generated Audio-Visual Pseudo-Fakes. The key idea of AVPF is to create pseudo-fake training samples that contain diverse audio-visual correspondence patterns commonly observed in real-world deepfakes. We highlight that AVPF is generated solely from authentic samples, and training relies only on authentic data and AVPF, without requiring any real deepfakes. Extensive experiments on multiple standard datasets demonstrate the strong generalizability of the proposed method, achieving an average performance improvement of up to \bm7.4%.

Abstract:
Collaborative edge-cloud frameworks have emerged as the main- stream paradigm for mobile automation, mitigating the latency and privacy risks inherent to monolithic cloud agents. However, existing approaches centralize administration in the cloud while relegating the device to passive execution, inducing a cognitive lag regard- ing real-time UI dynamics. To tackle this, we introduce AdecPilot by applying the principle of administrative decentralization to the edge-cloud multi-agent framework, which redefines edge agency by decoupling high-level strategic designing from tactical grounding. AdecPilot integrates a UI-agnostic cloud designer generating ab- stract milestones with a bimodal edge team capable of autonomous tactical planning and self-correction without cloud intervention. Furthermore, AdecPilot employs a Hierarchical Implicit Termi- nation protocol to enforce deterministic stops and prevent post- completion hallucinations. Extensive experiments demonstrate pro- posed approach improves task success rate by 21.7% while reducing cloud token consumption by 37.5% against EcoAgent and decreas- ing end to end latency by 88.9% against CORE. The source code is available at https://anonymous.4open.science/r/Anonymous_code- B8AB.

Abstract:
Learning with long-tailed noisy labels presents a fundamental challenge for classification, as the combined effects of class imbalance and label corruption distort the feature space. These distortions lead to two forms of inseparability: i) semantic confusion, where mislabeled samples resemble semantically similar classes, and ii) structural degradation, where clean tail samples are misclassified as noise. Existing methods typically depend on internal visual features, which are themselves biased and corrupted, making reliable sample selection intrinsically difficult. In this paper, we propose a paradigm shift from feature-based heuristics to external knowledge-guided verification. Drawing inspiration from the human coarse-to-fine cognitive process, we introduce CoEG (Cognitively-Inspired External Guidance), a novel framework that synthesizes holistic perception from vision-language models and analytical disambiguation from large language models. This dual-system design yields robust cognitive confidence for noise verification, enabling accurate and interpretable sample selection. To further enhance representation quality, CoEG employs relational knowledge distillation to transfer structured topological relations from external models to correct the students distorted feature manifold. Extensive experiments on both synthetic and real-world long-tailed noisy datasets demonstrate the effectiveness of CoEG.

Abstract:
Dense video captioning aims to simultaneously localize and describe multiple events within untrimmed videos, presenting the inherent challenge of balancing overarching storylines with fine-grained local details. Existing autoregressive models generate repetitive or locally inconsistent descriptions without a global view of the video narrative, while retrieval-augmented methods rely on computationally expensive external databases and static features that fail to adapt to specific video content. To overcome these limitations, we propose HiPrompt, a novel multimodal framework that replaces rigid prefixes with dynamic, video-specific latent variables acting as narrative anchors. Specifically, our approach formulates prompt generation as a conditional variational process by extracting a global prompt from multimodal inputs for macro-level narrative planning and generating independent local prompts for uniform temporal windows. By interleaving these local prompts directly with their corresponding visual frames, HiPrompt structurally guides the decoder to strictly follow the chronological order of events and accurately predict their boundaries. Furthermore, we introduce a neighbor-masked Window-level Contrastive Alignment Loss to ensure these local prompts remain deeply grounded in their visual context, preventing them from degrading into meaningless parameters while maintaining smooth narrative transitions. Extensive experiments on the YouCook2 and ViTT datasets demonstrate that the proposed model achieves state-of-the-art performance, consistently improving caption quality and temporal localization without relying on massive external retrieval banks.

Abstract:
Universal multi-modal fusion often fails not because modalities are insufficiently aligned, but because the fusion process allocates information poorly: redundant evidence is repeatedly amplified, modality-unique evidence is over-smoothed during alignment, and task-dependent synergistic evidence is insufficiently activated. We present RADIANT (Redundancy-Aware Dynamic Information Allocation Network), a task-general framework that organizes fusion around three complementary roles: suppressing redundant cross-modal overlap, preserving modality-unique evidence, and composing task-conditioned synergy through cross-modal interaction. Specifically, stable matrix-exponential projections are used to reduce interference among shared and modality-unique streams, while task memory with hard positive updates selects which synergistic interactions should be activated for the current fusion regime. Extensive experiments across six heterogeneous fusion tasks show consistent improvements over the compared methods. On FLIR, RADIANT improves downstream detection mAP@0.5 from 91.2% to 96.8%, a gain of 5.6 percentage points (6.1% relative) over GIF. RADIANT also exhibits strong direct cross-dataset transfer, with further gains from inference-time task-memory adaptation while keeping all network weights frozen. Additional ablations and sensitivity studies support the robustness of the allocation mechanism.

Abstract:
Chart understanding is an important capability of multimodal large language models (MLLMs) for scientific research and data analysis. However, severe occlusions in real-world charts remain challenging for MLLMs. Existing methods primarily address occlusion through dataset expansion or synthetic augmentation, while overlooking the local geometric ambiguity and unstable cross-patch instance associations caused by occlusion in patch-based visual representations. Inspired by the Gestalt Principle of Continuity, we propose OccluChart, an instance-centric framework that shifts chart understanding from patch-level perception to structured instance reasoning. Specifically, OccluChart performs ambiguity-aware instance reconstruction to recover structurally consistent chart elements from occluded observations, followed by a hyperspherical association mechanism that leverages angular consistency and semantic constraints to establish stable instance correspondences. Finally, structure-aware reasoning injection integrates the recovered instance representations into the reasoning process for improved occluded chart understanding. Extensive experiments on ChartQA, Chart-to-Table, and Chart Captioning show that OccluChart significantly outperforms existing state-of-the-art methods, with an average gain of 4.1%, particularly in real-world highly occluded chart scenarios.

Abstract:
Feed-forward 3D reconstruction models have achieved remarkable progress through large-scale training and strong learned priors. However, they remain fragile when applied to dynamic in-the-wild videos, where non-rigid motion, occlusions, and illumination changes are prevalent. A key limitation lies in their reliance on expensive depth and camera pose annotations, which hinders scaling to diverse real-world dynamic data and restricts generalization. To address this issue, we propose PURE, an unlabeled post-training refinement framework that improves depth and camera estimation without requiring ground-truth supervision. Instead of relying on annotations, PURE elegantly distills robust structural knowledge from state-of-the-art 2D priors into 4D feed-forward architectures. PURE leverages relative depth constraints, affine-invariant geometry distillation, and pose refinement based on feature correspondences to stabilize geometric predictions under challenging dynamic conditions. Extensive experiments on both synthetic and real-world dynamic benchmarks demonstrate that PURE delivers consistent improvements across multiple state-of-the-art architectures. Furthermore, our data-efficient paradigm effectively mitigates domain shifts, successfully unlocking the architectural limits of these base models in challenging, in-the-wild dynamic scenes.

Abstract:
Although dynamic scene reconstruction has long been a fundamental challenge in 3D vision, the recent emergence of 3D Gaussian Splatting (3DGS) offers a promising direction by enabling high-quality, real-time rendering through explicit Gaussian primitives. However, existing 3DGS-based methods for dynamic reconstruction often suffer from spatio-temporal incoherence during initialization, where canonical Gaussians constructed in the initialization stage are formed by aggregating observations from multiple frames without temporal distinction. This results in spatio-temporally entangled representations, which hinders the subsequent modeling of dynamic motion. To overcome this limitation, we propose STDR (Spatio-Temporal Decoupling for Real-time rendering), a plug-and-play module that learns spatio-temporal probability distributions for each Gaussian. STDR introduces a spatio-temporal mask, a separated deformation field, and a consistency regularization to jointly disentangle spatial and temporal patterns. Extensive experiments demonstrate that incorporating our module into existing 3DGS-based dynamic scene reconstruction frameworks leads to notable improvements in both reconstruction quality and spatio-temporal consistency across synthetic and real-world benchmarks.

Abstract:
Composed Image Retrieval (CIR) aims to retrieve target images using reference images and modification texts. Current CIR methods rely on fine-tuning vision-language pre-trained models. However, we find these approaches commonly suffer from severe overfitting, posing challenges for CIR with limited triplet data. To understand this issue, we present a systematic study of overfitting in VLP-based CIR, revealing a significant and previously overlooked generalization gap across models and datasets. Motivated by these findings, we introduce WRF4CIR, a Weight-Regularized Fine-tuning network for CIR. Specifically, during fine-tuning, we apply adversarial perturbations to the model weights for regularization, where perturbations are generated in the opposite direction of gradient descent. Intuitively, WRF4CIR increases the difficulty of fitting the training data, which helps mitigate overfitting in CIR under limited triplet supervision. Extensive experiments on benchmark datasets demonstrate that WRF4CIR significantly narrows the generalization gap and achieves substantial improvements over existing methods.

Abstract:
Reference-driven image generation has made rapid progress on identity preservation, but reliable viewpoint control across subjects remains poorly understood. The challenge is to infer the viewpoint of one subject and transfer it to another using only image evidence, without camera poses, depth, or ray-based conditions. In this setting, existing multi-reference generators often rely on spurious semantic correlations, which lead to viewpoint drift, part-level structural mismatches, and missing or unsupported target-specific content. We formulate this challenge as cross-subject viewpoint alignment and propose RAVA, a retrieval-augmented framework that supplies explicit geometric evidence before generation. RAVA first learns a cross-instance viewpoint embedding that retrieves target-subject images aligned with the anchor viewpoint, then applies a LogDet-based subset selection strategy to retain a compact reference set that is both view-consistent and structurally complementary. The selected references are finally consumed by a fine-tuned multi-reference image generator. Experiments show that generic semantic embeddings are nearly random for this task, while the proposed retriever substantially improves viewpoint retrieval quality and, on cross-subject generation, RAVA consistently outperforms zero-shot baselines and stronger retrieval alternatives under the same generation backbone, demonstrating the value of retrieval-augmented geometric grounding over end-to-end generation alone.

Abstract:
Unified multimodal models (UMMs) integrate understanding and generation within a shared representation space. While this unified modeling enables comprehensive multimodal capabilities, it also causes undesirable concepts, such as not-safe-for-work (NSFW) content and copyrighted identities, to be encoded across both pathways. Existing unlearning methods are primarily designed for pathway-isolated models, and thus are insufficient to consistently remove concepts entangled across the understanding and generation pathways in UMMs. To address this challenge, we propose UniForget, the first concept unlearning framework for unified multimodal models. UniForget formulates concept unlearning as a unified self-distillation process, in which concept-aligned multimodal supervision is derived from internalized knowledge of the model under a shared concept hierarchy. This supervision enables consistent unlearning across both understanding and generation pathways, achieving the complementary objectives of erasure, redirection, and retention. Furthermore, we introduce a unified benchmark for concept unlearning in UMMs. Extensive experiments show that UniForget effectively removes targeted concepts across both pathways while preserving overall model utility.

Abstract:
Recent advances in deep learning have driven significant progress in neural network-based phase hologram generation. Many existing approaches adopt network architectures and activation functions originally designed for natural image processing, overlooking the periodicity of phase values. This mismatch can cause three practical issues: blocked optimization near the 2\pi phase-wrapping boundary, a lack of control over output phase distribution which is often required for better reconstruction on real SLMs, and excessive spatial correlation that leads to optical artifacts such as high-diffraction orders. We show that these limitations can be addressed by simply replacing the final activation function with our proposed FlexMU (Flexible Modulo Unit). FlexMU combines modulo-based wrapping with two explicit control parameters: \alpha, which shapes the phase distribution toward a more favorable SLM operating region, and \beta, which controls spatial decorrelation of the generated phase map. Applied across three representative neural holography models, which are HoloNet, DPRC and NHVC, FlexMU consistently improves not only phase generation quality by up to 1.81~dB in PSNR and phase compression performance by 1.25~dB in BD-PSNR. Optical experiments further prove that FlexMU improves optical reconstruction quality and suppresses hardware-sensitive artifacts. These results suggest that FlexMU is a practical drop-in component for neural architectures dealing with periodic outputs, potentially extending beyond holography to wave-based applications such as metrology, SAR, and sonar.

Abstract:
Artificial Neural Networks (ANNs) have emerged as the dominant paradigm in gait recognition, achieving strong performance by effectively extracting complex spatiotemporal features from silhouette sequences. However, their dense-valued computation is inherently mismatched with the binary and sparse nature of gait silhouettes. In this paper, we present SpikingGait, a hierarchical backbone driven by Spiking Neural Networks (SNNs). It exploits the inherent sparsity of spike-based computation and formulates gait recognition as a progressive structural-to-volumetric representation problem. The architecture begins with the Spiking Structural Residual Block (S-SRB) to encode silhouette and short-range motion, then transitions to the Spiking Volumetric Residual Block (S-VRB) for long-range spatiotemporal aggregation. A Global Spiking Projector (GSP) bridges the spike-driven backbone and standard gait heads, preserving compatibility with mainstream pipelines while isolating the contribution of the spiking feature extractor. Across CASIA-B, CCPG, SUSTech1K, and Gait3D, SpikingGait achieves competitive performance, including state-of-the-art results on CASIA-B and Gait3D among silhouette-based methods; on Gait3D, it obtains 83.3 percent Rank-1 accuracy and 75.9 percent mAP. Furthermore, we quantitatively evaluate the energy efficiency of SpikingGait. Benefiting from high firing sparsity, our model achieves significantly lower computational energy consumption compared to ANN-based methods.

Abstract:
Mask-based recognition with multimodal large language models (MLLMs) offers a promising route to vocabulary-free segmentation. However, existing methods typically predict each candidate mask independently, limiting shared-context reasoning, while naive joint multi-mask generation suffers from cross-mask interference. We present VFSeg, a vocabulary-free segmentation framework for joint multi-mask recognition that predicts attribute-category outputs for all masks in a single pass. Specifically, Position-Anchored Embedding (PAE) is designed to establish stable positional anchors by binding observations of the same mask across crops to shared slots, thereby decoupling position from content. Building on this design, Attribute-Guided Disambiguation (AGD) further improves output discrimination by generating concise, visually grounded attributes to guide mask category prediction, thereby reducing confusion among similar targets. To support VFSeg training, we build VGAMask, a large-scale supervision corpus of mask-category-attribute tuples. Experiments show that VFSeg substantially outperforms prior vocabulary-free baselines and demonstrates stronger open-ended semantic capability.

Abstract:
Text-to-image person re-identification (TIReID) aims to retrieve pedestrian images from a gallery according to a natural-language query. Recent multimodal large language model (MLLM)-based interactive TIReID frameworks improve retrieval flexibility through test-time interaction, but their effectiveness heavily depends on anchor selection. Existing methods typically rely on first-positive acceptance, which suffers from two coupled issues. First, the MLLM judge is commonly adapted with coarse response-level supervision, so a positive decision may be weakly grounded in query-relevant visual evidence. Second, positive compatibility alone does not guarantee anchor suitability: among positively matched candidates, some may still conflict with the query, provide weaker support for query-critical attributes, or expose only a fragmented portion of the query-mentioned evidence. To address these issues, we reformulate anchor selection in interactive TIReID as a reliable anchoring problem and propose Evidence-Grounded Reliable Anchoring (EGRA), a two-stage framework that improves anchor selection from both judgment grounding and anchor qualification perspectives. Specifically, EGRA combines Evidence-Grounded Anchor Learning (EAL), which augments MLLM adaptation with structured human-centric evidence supervision to produce more grounded positive judgments, and Interaction-Ready Anchor Qualification (IAQ), which replaces first-positive acceptance with staged non-compensatory qualification over a tiny positive shortlist based on contradiction, explicit query support, and visual completeness. Extensive experiments on CUHK-PEDES, ICFG-PEDES, and RSTPReid demonstrate that EGRA consistently outperforms strong interactive baselines under both standard and ReID-domain pre-training settings. Additional ablation, efficiency, and anchor-selection analyses further confirm the effectiveness and practicality of the proposed framework.

Abstract:
Out-of-context (OOC) misinformation, where real images are paired with deceptive captions, remains a core challenge for multimodal fact-checking. Existing methods still rely on similarity-based or holistic caption-level verification. Many are also shaped by synthetic benchmarks, where similarity can be overly influential, reducing their effectiveness on real-world OOC cases that are globally plausible but locally inconsistent. To bridge this gap, we propose GUIDE, a training-free, divide-and-conquer framework that reformulates OOC detection as question-guided verification. Instead of verifying the caption as a whole, GUIDE decomposes it into targeted verification questions over decision-critical factual attributes, making localized factual conflicts explicit. To verify these localized questions, GUIDE gathers structured and complementary evidence, including visual descriptions, structured context references, retrieved web captions, and a similarity score. To avoid reverting to a coarse holistic judgment, GUIDE then performs comprehensive verification over these localized questions, aggregating local judgments into a final prediction. Experiments demonstrate that GUIDE achieves significant results compared with SOTAs on benchmarks, improving accuracy on 5Pils-OOC from 58.2% to 67.2% and achieving the best VERITE result of 62.7%. These results indicate that realistic OOC detection benefits from question-guided verification over similarity-based or holistic caption-level verification.

Abstract:
Multi-view clustering aims to exploit complementary information from multiple modalities to learn discriminative representations without supervision. However, existing methods often suffer from feature entanglement, where view-shared semantics are mixed with view-specific variations. This issue leads to less discriminative representations and degrades clustering performance. To address this issue, we propose DvMF (Disentangled Multi-View Fusion with von Mises-Fisher Distribution), a unified framework that combines representation disentanglement with uncertainty-aware feature fusion. Specifically, we introduce a tripartite mutual information objective to control dependencies among latent variables. It enhances cross-view consistency while reducing redundancy between shared and private representations, thus promoting the separation of view-shared semantics and view-specific factors. To robustly integrate disentangled representations, we project them onto a hyperspherical latent space modeled by the von Mises-Fisher (vMF) distribution. Fusion is then performed via a Product-of-Experts mechanism. The concentration parameter in the vMF distribution encodes representation reliability, enabling uncertainty-aware aggregation that down-weights less reliable features and mitigates the influence of noisy view-specific components in the fused embedding. Extensive experiments on multiple benchmark datasets show that DvMF consistently outperforms recent state-of-the-art multi-view clustering methods.

Abstract:
Creating presentation slides from scientific papers is not simply a matter of summarizing paragraphs. A presenter is required to decide what story to tell, which figures and equations to highlight, and how to arrange them into pages that are visually clear rather than crowded or repetitive. The need to jointly reason over long contexts and layout-sensitive design makes paper-to-slide generation a uniquely challenging multimodal task. Most existing approaches, however, focus mainly on textual content selection, producing slides that often lack visual balance, narrative flow, or coherent integration of multimodal evidence. In this work, we introduce SlideGen, a collaborative vision-language multi-agent framework that coordinates narrative planning, multimodal grounding, and layout composition. SlideGen assigns specialized agents to outline the presentation structure, align supporting figures and tables with key claims, generate speaker notes, and compose editable PPTX slides through a diverse layout library. By refining layouts at the deck level, the system produces slide decks that are both faithful to the source paper and effective as presentations. To evaluate slide generation beyond text fidelity, we propose geometry-aware density (GAD), a metric that captures visual clutter, sparsity, and fragmentation, and shows strong agreement with human judgments. Evaluated across four complementary dimensions on our 200-paper benchmark, SlideGen consistently and significantly improves layout balance, content coverage, and text coherence, outperforming competitive baselines in paper-to-slide generation. Our findings suggest that effective slide generation requires multimodal design reasoning, and that agent collaboration offers a principled bridge between document understanding and scientific communication.

Abstract:
Recent advances in diffusion models have significantly improved the quality of image synthesis. At the same time, the misuse of such high-fidelity generated images raises growing concerns about intellectual property, privacy, reputation, and other security-related issues. Image watermarking offers a promising solution, and recent studies have increasingly explored Noise-as-Watermark (NaW) methods that embed watermarks directly into the diffusion process. However, existing NaW approaches remain insufficiently robust for real-world deployment. We argue that this limitation stems from the prevailing embedding paradigm, which directly maps watermark bits to individual sampled values. Instead, we propose encoding watermark bits into the structural relationships among groups of values, so that the encoded information can survive as long as the underlying structure is preserved. Furthermore, to better maintain generation diversity, we introduce a dedicated randomization mechanism that reshuffles noise element positions, preventing watermark embedding from inducing fixed spatial patterns or locations. Extensive experiments show that our method achieves state-of-the-art robustness while preserving high generation quality across a wide range of lossy scenarios.

Abstract:
Embedding-as-a-Service (EaaS) has become an important semantic infrastructure for natural language and multimedia applications, but it is highly vulnerable to model stealing and copyright infringement. Existing EaaS watermarking methods face a fundamental robustness--utility--verifiability tension: trigger-based methods are fragile to paraphrasing, transformation-based methods are sensitive to dimensional perturbation, and region-based methods may incur false positives due to coincidental geometric affinity. To address this problem, we propose GeoMark, a geometry-aware localized watermarking framework for EaaS copyright protection. GeoMark uses a natural in-manifold embedding as a shared watermark target, constructs geometry-separated anchors with explicit target--anchor margins, and activates watermark injection only within adaptive local neighborhoods. This design decouples where watermarking is triggered from what ownership is attributed to, achieving localized triggering and centralized attribution. Experiments on four benchmark datasets show that GeoMark preserves downstream utility and geometric fidelity while maintaining robust copyright verification under paraphrasing, dimensional perturbation, and CSE (Clustering, Selection, Elimination) attacks, with improved verification stability and low false-positive risk.

Abstract:
Vector sketch animation offers tremendous advantages for multimedia and creative design through concise line expressions and flexible editing. Learning-based generation methods of sketch animation have made significant progress in the last decade, but still suffer from limited style diversity and controllability. This paper presents a controllable video sketching method that automatically converts videos into multi-style vector sketch animations. A continuous style hyperspace is constructed by multi-dimensional sketch styles (fidelity, simplicity, text guidance strength) and the timeline. With this hyperspace, stroke control points are parameterized as 4-variable Bernstein polynomials, ensuring smooth and differentiable style transitions. A multi-task, multi-stage optimization framework is designed to learn stroke hyperparameters accurately and efficiently. We further developed a web-based interactive interface that allows real-time style manipulation via editable curves. Experiments show the style controllability, high-quality, and user-friendliness of our method, which outperforms SOTA methods.

Abstract:
Recent advancements in text-guided image editing have demonstrated remarkable capabilities in synthesizing realistic textures and following complex instructions. Nevertheless, most existing models are trained primarily on static images and inherently lack the ability to reason about physical causality. This limitation often leads to ``physical hallucinations'', where generated content violates physical laws (e.g., inconsistent lighting, implausible gravity, or multi-step causal effects). In contrast, video generation models implicitly encode rich physical priors, but they are ill-suited for high-fidelity image editing due to task misalignment and limited spatial resolution. To address this gap, we present SVIE, a novel framework that seamlessly transplants the physical reasoning capabilities of video generation models into image editing models. Technically, SVIE achieves this by stitching a frozen video generation backbone into the image editing pipeline via a specialized Stitching Adapter. This lightweight adapter acts as a semantic bridge, projecting causality-rich video latents directly into the image model's feature space. These aligned priors are then injected into the denoising process to encourage more physically consistent outputs. Extensive experiments show that SVIE substantially improves physical plausibility in challenging scenarios while preserving instruction adherence and detail fidelity. Notably, SVIE enhances physical reasoning accuracy, achieving an absolute gain of 5.2 percentage points in overall accuracy on RISEBench and consistent gains on PICABench. Crucially, it secures these gains while maintaining competitive performance on general editing benchmarks like ImgEdit.

Abstract:
The deployment of Vision Transformers (ViTs) on resource constrained platforms is increasingly hindered by significant storage and transmission overheads. Existing compression methods typically treat model weights as static tensors, overlooking their functional heterogeneity and the potential for progressive recovery. In this paper, we propose DNA-ViT, a biologically inspired compression framework that reformulates ViT storage as a structured, progressive expression process. DNA-ViT first maps weights into a high-density domain via statistical normalization and a 3D Haar transform for multiresolution energy compaction. It then employs Hessian-aware sensitivity analysis to prioritize task-critical functional units, mirroring the coding-region logic of genomic DNA. Finally, the compressed representation is organized into a progressive code-stream, enabling coarse-to-fine model reconstruction to accommodate heterogeneous deployment budgets. Extensive experiments across multiple ViT backbones and low-bit configurations demonstrate that DNA-ViT consistently achieves superior accuracy--rate trade-offs over state-of-the-art baselines. Notably, it outperforms state-of-the-art baselines by up to 11.96% in 4-bit Top-1 accuracy while reducing inference latency by 28.2%. These results suggest that structured archiving and progressive recovery offer a promising paradigm for efficient Vision Transformer compression.

Abstract:
Open-vocabulary multi-label recognition (OV-MLR) aims to predict multiple labels beyond seen training categories by transferring knowledge from vision-language models. Existing methods typically rely on either raw patch-level matching or a single global token, which makes them vulnerable to seen-category bias under seen-only supervision. In particular, patch features tend to collapse onto a few competitive seen categories, causing predictions to be dominated by seen-category responses, and limiting generalization to unseen labels. To address this issue, we propose PAIR (Prototype-Aware Inertia Rectification), a unified framework that reduces seen-category bias from both local and global perspectives. First, PURE (Prototype Recovery with Uncertainty-aware Region Enhancement) reorganizes patch tokens into teacher-guided latent prototypes, and refines them via transport-based prototype recovery, yielding more transferable local representations. Second, AIR (Alignment Inertia Rectification) builds a prototype-conditioned global representation, and rectifies global alignment through hard-negative shaping and local-to-global consistency regularization. By jointly rectifying local prototype learning and global representation alignment, PAIR provides a more balanced basis for open-vocabulary prediction. Extensive experiments on MS-COCO and NUS-WIDE show that PAIR consistently outperforms the strong OV-MLR baseline, demonstrating its effectiveness across both benchmarks.

Abstract:
Embodied planners interact with the physical world by reasoning over scene observations, a process increasingly augmented by retrieving external cross-modal memories to tackle long-horizon tasks. However, this retrieval mechanism creates a novel attack surface, where poisoned procedural memory can be activated by scene-native cues and ultimately translated into unsafe physical actions. In this paper, we present TOXRobo, a low-budget procedural memory poisoning attack that contaminates only a few retrievable records and couples them with scene-native object or object-composition cues. Each poisoned record is derived from a semantically matched clean memory and edited mainly through its planning rationale, with minimal trigger-consistent changes to adjacent fields, so that it remains retrievable under joint scene-instruction cues while still biasing downstream plan generation after retrieval. Across LIBERO, CALVIN, and a myCobot deployment, TOXRobo consistently induces targeted misbehavior while preserving clean-task utility on non-triggered tasks. Ablations show that compositional triggers and rationale-centered edits are important to attack effectiveness and selectivity. These results identify procedural memory as a concrete multimodal vulnerability in embodied AI and show that securing retrieval-augmented embodied planners requires defenses beyond prompt- or retrieval-only safeguards.

Abstract:
Beyond semantic content, camera parameters play a pivotal role in dictating the geometric perspective and appearance of any given image. While recent image editing models excel at semantic and stylistic manipulation, they struggle with explicit camera parameter control. When handling large perspective shifts, instruction-driven models face a dilemma: they either suffer from structural tearing or generate conservative outputs that ignore geometric instructions. To address this, we introduce CameraEditor, a framework that reformulates camera-controlled editing from a spatial problem into a temporal sequence prediction task. By leveraging the temporal coherence of video diffusion models, our approach integrates an explicit geometric perception module with a dynamic reference routing mechanism. This allows us to construct geometrically rigorous visual reference pairs via dynamic panorama cropping, overcoming the ambiguity of text-based instructions. Furthermore, CameraEditor strategically inserts intermediate transition frames to decompose large perspective shifts, providing a robust temporal buffer that preserves content identity and spatial coherence. To support this paradigm, we construct a training dataset of 5,760 instances and introduce CamEditor-Bench, a rigorous evaluation suite of 462 test cases. Extensive experiments demonstrate that CameraEditor achieves state-of-the-art camera control precision and source identity preservation, outperforming existing methods.

Abstract:
Collision anticipation in autonomous driving requires not only accurate early warnings but also interpretable reasoning about what risk factors are being tracked and how risk evolves over time. Existing methods fall short in this regard: feature-driven models are opaque, post-hoc explanations often lack fidelity, and concept-based methods are mostly designed for static recognition rather than dynamic driving scenes. We propose CARA (Concept-Aware Risk Attention), an intrinsically interpretable spatio-temporal framework for collision anticipation. CARA derives domain-grounded risk concepts from accident narratives, aligns them with video frames via vision-language similarity, and organizes them into evolving concept trajectories. These trajectories provide explicit risk evidence that guides spatial attention, temporal attention, and anticipation, allowing semantic concepts to directly influence both where the model attends and how it predicts risk over time. By treating semantic risk factors as dynamic intermediate evidence rather than auxiliary post-hoc explanations, CARA tightly couples interpretability with the predictive process. Extensive experiments on three benchmarks show that CARA consistently improves anticipation accuracy and warning earliness over strong baselines, while providing sparse and semantically grounded concept evidence.

Abstract:
Deep multi-view clustering has achieved remarkable progress but remains vulnerable to complex noise in real-world applications. Existing noisy robust methods predominantly rely on a simplified binary assumption, treating data as either perfectly clean or completely corrupted. This overlooks the prevalent existence of heterogeneous observation noise, where contamination intensity varies continuously across data. To bridge this gap, we propose a novel framework termed Quality-Aware Robust Multi-View Clustering (QARMVC). Specifically, QARMVC employs an information bottleneck mechanism to extract intrinsic semantics for view reconstruction. Leveraging the insight that noise disrupts semantic integrity and impedes reconstruction, we utilize the resulting reconstruction discrepancy to precisely quantify fine-grained contamination intensity and derive instance-level quality scores. These scores are integrated into a hierarchical learning strategy: at the feature level, a quality-weighted contrastive objective is designed to adaptively suppress the propagation of noise; at the fusion level, a high-quality global consensus is constructed via quality-weighted aggregation, which is subsequently utilized to align and rectify local views via mutual information maximization. Extensive experiments on five benchmark datasets demonstrate that QARMVC consistently outperforms state-of-the-art baselines, particularly in scenarios with heterogeneous noise intensities.

Abstract:
Vision-Language Models (VLMs) have demonstrated remarkable multimodal capabilities but suffer from prohibitive computational overhead due to the massive number of visual tokens. While token pruning is a promising solution, existing methods still suffer from two major drawbacks: 1) unreliable pruning criteria that fail to accurately localize task-specific regions due to inherent attention biases or query-agnostic designs, and 2) suboptimal pruning strategies that employ rigid, non-adaptive schedules, overlooking the dynamic evolution of token redundancy across different model depths. To address these challenges, we propose LearnPrunerV2, which extends the previous two-stage heuristic framework (LearnPruner) into a fully learnable architecture. Specifically, we integrate lightweight learnable modules into every layer of the VLM backbone, which estimate token importance by fusing intrinsic visual saliency with extrinsic query relevance captured via a shadow token. By uncovering a consistent staircase redundancy pattern across VLM layers, we further introduce Budget-Deterministic Pruning (BDP) to establish a hardware-friendly schedule with fixed per-layer budgets. Extensive experimental results across various benchmarks and VLMs demonstrate that LearnPrunerV2 consistently outperform previous state-of-the-art methods. Notably, LearnPrunerV2 preserves 99.8% of the original performance when pruning 77.8% of visual tokens on LLaVA-v1.5, achieving a superior accuracy-efficiency trade-off.

Abstract:
Co-speech gesture generation in dyadic conversations for creating immersive virtual avatars remains highly challenging. Existing models often struggle to balance kinematic stability with rhythmic responsiveness and suffer from semantic entanglement, leading to either generic mean-pose regression or unnatural micro-jitters. To address these challenges, we propose SynerGest, a high-fidelity dyadic gesture generation framework grounded in Contrastive Flow Matching (CFM). Specifically, our CFM objective explicitly penalizes mismatched semantic pairs during training, discouraging spatial shortcuts and generic, over-smoothed predictions while improving conditional semantic alignment. Furthermore, to overcome architectural bottlenecks, we introduce the Impulse-Trend Synergistic Attention (ITSA) module. Grounded in the key insight of decoupling "what moves" from "when it moves," ITSA models sharp, speech-driven rhythmic impulses separately from continuous kinematic trends, supporting precise audio-motion synchronization and global structural stability. Extensive qualitative and quantitative evaluations on two public datasets demonstrate that SynerGest achieves state-of-the-art performance, delivering contextually responsive and kinematically fluid dyadic interactions.

Abstract:
Underwater video enhancement (UVE) is essential for marine exploration, yet non-rigid refractive distortions and frequency-dependent attenuation often compromise the efficacy of standard restoration models. Although State Space Models (SSMs) like Mamba provide computational efficiency, their traditional direct-scan paradigm fails to accommodate the stochastic feature shifts and turbulent motion inherent in aquatic scenes. In this paper, we present UVEMamba, a framework specifically engineered to bridge the gap between linear-complexity sequence modeling and complex aquatic degradation. To address temporal instability, we propose the Motion-Rectified Spatiotemporal Mamba (MR-STM). By integrating a Sparse Selective Perception Module, this module explicitly rectifies non-rigid motion offsets before state evolution, ensuring that Mambas hidden states capture stable trajectories rather than corrupted noise. To further decouple complex degradations, we develop the Spatial-Frequency Mamba Block (SFMB), which alternates between spatial geometric correction and wavelet-based frequency aggregation. This isolates global chromatic casts from high-frequency scattering, allowing for targeted artifact suppression and detail preservation. Complementing this architecture, our Unified Spatio-Temporal Color Loss (USTC-Loss) employs motion-aware contrastive learning to enforce long-term consistency, mitigating the common trade-off between structural fidelity and visual perception. Extensive experiments on four datasets demonstrate that our approach generalizes well, delivering high-fidelity and temporally stable restoration for complex underwater videos.

Abstract:
Multimodal Emotion Recognition (MER) integrates audio, visual, and textual signals to recognize emotions. Most existing MER systems rely on a single fixed fusion pipeline for all samples. However, modality reliability, signal quality, and informative cue patterns vary considerably across samples, making a one-size-fits-all pipeline inherently suboptimal. Existing adaptive methods often rely on opaque latent routing or generic competence estimation, treating expert diversity as a byproduct rather than an explicit design goal. We propose MATE, an interpretable condition-space-guided expert teaming framework. MATE introduces a human-interpretable meta-prior condition space that organizes samples into operating-condition subdomains, profiles expert competence across those subdomains, and guides team selection over a pool spanning complementary architectural biases and frozen representations. The selected compact team of four complementary experts is aggregated using a parameter-free uniform Product-of-Experts (PoE) consensus. Extensive experiments on IEMOCAP, RAVDESS, and CREMA-D demonstrate that this carefully constructed team with frozen encoders achieves performance comparable to state-of-the-art methods that rely on large-scale pre-training or encoder fine-tuning, suggesting that deliberate expert composition, rather than more complex fusion, is a highly effective yet underexplored design axis for MER.

Abstract:
Adapting Image Large Language Models (Image-LLMs) to long-form video is hindered by the large number of redundant visual tokens. Existing video adapters reduce this cost, but their unconstrained projections may generate representations that deviate from the input distribution expected by a frozen Image-LLM, degrading its pre-aligned visual--language semantics. We introduce the Alignment-Preserving video Tokenizer (APT), a FLOP-efficient interface for non-destructive Image-to-Video adaptation. APT first identifies informative spatio-temporal anchors with a Spatio-Temporal Visual Token Refiner, and then condenses them using a Text-Conditioned Token Aggregator. The final value aggregation is constrained to a parameter-free convex combination of the selected anchors, providing a structural inductive bias that conditionally limits representational drift without requiring external shot detection or additional image training data. APT generalizes across diverse Image-LLM backbones and achieves leading performance on short-, long-, and open-ended video understanding benchmarks while substantially reducing adapter FLOPs and inference latency. Evaluations on static-image benchmarks further show statistically insignificant performance variations, indicating that APT mitigates adapter-induced distribution shift while preserving zero-shot compatibility with the original Image-LLM.

Abstract:
Uncertainty Quantification (UQ) aims to measure the reliability of model predictions, serving as a critical safeguard for deploying Vision-Language Models (VLMs) in safety-critical scenarios. Post-hoc approaches are widely adopted due to their lightweight nature, mapping the outputs of VLMs to uncertainty measures through learnable modules or inductive summarization. However, Post-hoc approaches remain inherently confined to fitting the failure patterns of the source domain, ignoring the dynamic nature of test distributions. To address this challenge, we propose a Dynamic Distribution-Aware Uncertainty Quantification framework (DDA-UQ) that shifts the paradigm from static mapping to a dynamic distribution-aware process. During training, we leverage a Gaussian Mixture Model to model the VVLMs'embedding space and extract distributional evidence, thereby dynamically deriving uncertainty estimates. During inference, the design dynamically responds to changes in the data distribution. Extensive experiments demonstrate that our approach significantly outperforms state-of-the-art methods.

Abstract:
Vision Transformers exhibit strong global modeling capabilities in crowd localization, yet their efficiency-oriented self-attention designs introduce a low-frequency bias. This suppresses high-frequency details, leading to over-smoothed responses and degraded instance separability. To address this issue, a Frequency-aware Attention Network (FANet) is proposed to jointly model global low-frequency semantics and local high-frequency structures. At its core, a frequency-aware dual-branch transformer module is designed, which is composed of three key components. First, a local-enhanced attention branch is designed alongside the global branch to recover fine-grained high-frequency details at the original resolution. Furthermore, to enable effective integration of complementary frequency components, an adaptive frequency gating mechanism is developed to perform spatially selective and adaptive modulation of high-frequency features. In addition, a distribution-constrained learning strategy is introduced to regularize the gating responses, ensuring stable and discriminative spatial representations. Extensive experiments demonstrate that FANet achieves state-of-the-art performance, significantly improving localization accuracy and robustness, particularly in extremely dense scenarios and under strict evaluation criteria.

Abstract:
Zero-Shot Multimodal Anomaly Detection (ZSMAD) aims to identify both appearance and geometric defects in novel categories without relying on target-specific training data, making it especially valuable in real-world scenarios constrained by data scarcity, privacy concerns, or high annotation costs. While recent approaches successfully adapt vision-language models like CLIP to achieve this open-vocabulary generalization, they still face significant challenges: (i) severe geometric information loss during 3D-to-2D point cloud rendering; (ii) spatially noisy 2D feature maps produced by raw CLIP patch embeddings, compounded by shallow late-stage modality fusion; and (iii) severe cross-modal semantic misalignment due to decoupled processing streams. To address these limitations, we propose AggCLIP, a novel framework that adapts frozen CLIP representations for ZSMAD by explicitly incorporating local geometric structure and visual spatial context. Specifically, to accurately localize fine-grained defects, we design a Geometry-Aware Spatial Aggregation (GSA) module that refines back-projected 3D point features via density-adaptive neighborhood weighting, alongside a Context-Aware Spatial Aggregation (CSA) module that smooths multi-scale 2D patch embeddings using a lightweight trainable adapter. Furthermore, to prevent semantic drift between modalities, we introduce a Dual-Stream Prompt Contrastive (DSPC) loss, which aligns the abstract concepts of normality and abnormality across the point cloud and RGB streams. Extensive experiments demonstrate that the proposed AggCLIP framework achieves state-of-the-art performance on the MVTec 3D-AD and Eyecandies benchmarks under both one-vs-rest and cross-dataset zero-shot settings, notably surpassing the strongest competitive methods by up to 2.9% in pixel-level AUPRO.

Abstract:
Motion deblurring has achieved remarkable progress with deep learning, yet pre-trained deblurring models often suffer from performance degradation in real-world scenarios due to the domain shift between training and testing distributions. To remedy this, we propose EvBS, an event-guided blur synthesis framework that generates diverse training pairs for calibrating pre-trained models to the target domain. While existing methods are constrained by the inherent entanglement between motion and visual content, our method leverages the high temporal resolution of event cameras to effectively decouple them. This enables us to utilize not only the intrinsic motion that is inherent to the given content but also extrinsic motion transferred from different sources within the target domain, thereby facilitating effective adaptation via fine-tuning. Specifically, EvBS comprises two complementary strategies: Intrinsic-Blur Synthesis, which blurs sharp contents with their own motion patterns, and Extrinsic-Blur Synthesis, which transfers motion from blurry patches to distinct sharp content. This approach generates a diverse set of training pairs that break the inherent constraints of naturally coupled motion and content, resulting in enhanced domain-adaptive deblurring performance. Extensive experiments on multiple benchmarks demonstrate that EvBS effectively enhances the robustness of existing deblurring models on unseen testing datasets.

Abstract:
Multimodal agents operating in long-horizon environments must build and continually update multimedia memories to support entity-consistent, temporally grounded reasoning. However, existing agentic memory approaches often discard fine-grained identity cues under aggressive compression and segment-wise processing. They also rely heavily on vector-similarity retrieval, which can surface semantically related yet identity-mismatched evidence, leading to entity confusion, error propagation, and hallucinated answers. We propose ViSAGE, a multimodal agentic memory framework that constructs self-correcting, entity-centric memories. Specifically, ViSAGE anchors entity identity via cross-modal binding over long temporal ranges. It then applies bidirectional memory refinement to propagate delayed identity evidence, retroactively unifying historical records and improving future reasoning. We also introduce multi-agent cross-verification to assess retrieved evidence under an identity--evidence alignment constraint, enabling abstention instead of unsupported answers when evidence is missing. Extensive results demonstrate that ViSAGE consistently outperforms the strongest baseline, achieving 5.9% higher accuracy.

Abstract:
Image colorization is an inherently underconstrained task requiring precise semantic understanding and strict structural alignment. Existing colorization methods typically model high-frequency signals and low-frequency semantics jointly, overlooking their distinct and complementary characteristics. This frequency coupling frequently leads to color bleeding across semantic boundaries. To address these limitations, we introduce ColorFlow, a novel generative model that decouples frequency representations. The proposed architecture clearly separates high-frequency signals from low-frequency semantics and then ultimately integrates them. To effectively fuse this information, the model employs a fully differentiable Fast Guided Filter as its core module. This filter directly uses low-frequency semantics to spatially guide high-frequency signals, ensuring a coherent color distribution. To prevent blurring and desaturation, we further incorporate two loss functions: LPIPS loss preserves fine details, while DINO loss enforces global semantic consistency. Extensive evaluations demonstrate that ColorFlow achieves highly competitive overall performance, successfully synthesizing perceptually vibrant images with precise semantic alignment.

Abstract:
Long-horizon embodied navigation requires agents to visit multiple semantic targets in cluttered, partially observed 3D environments. While neural perception and reinforcement learning are effective for local control, selecting which goal to visit next and optimizing the overall visiting order remain combinatorial bottlenecks. We present Q-Nav, a quantum-classical hybrid framework that formulates high-level multi-goal route planning as a quadratic unconstrained binary optimization (QUBO) problem and solves it with quantum optimizers such as QAOA and quantum annealing, while retaining classical modules for visual perception, mapping, and low-level control. This decomposition enables explicit global route optimization without sacrificing robust embodied execution. We evaluate Q-Nav on MultiON and GOAT-Bench under matched compute budgets against RL-only, classical planning, and hybrid ILP baselines. Q-Nav achieves the best overall performance, improving success rate by up to 7.5% over the strongest baseline while reducing planning cost by roughly 30-40% on the evaluated benchmarks. These results suggest that quantum-assisted combinatorial optimization is a promising high-level planning primitive for long-horizon multi-goal embodied navigation.

Abstract:
Incomplete multi-view clustering (IMVC) is a critical task in multimedia analysis, where heterogeneous modalities frequently suffer from both missing views and complex noise. Existing deep IMVC methods typically employ generative imputation or instance-level alignment, prioritizing low-level and noise-sensitive local features to maximize performance. However, this tendency causes models to reconstruct noise rather than underlying structures, leading to semantic drift and significant performance degradation under high missing rates. To address the above issues, we propose Globally Robust Incomplete Multi-View Clustering (GRIMC) under complex noise. Specifically, GRIMC employs a gradient-guided adversarial masking mechanism to adaptively mask highly discriminative but noise-sensitive feature regions, compelling the network to bypass local shortcuts and infer globally consistent semantics. Subsequently, a prototype contrastive module employs dynamically maintained global prototypes to guide the semantic reconstruction and soft probability assignment of missing views, stabilizing the clustering structure in a noise-robust latent space. Experiments on various noisy datasets demonstrate that GRIMC effectively prevents the performance degradation seen in existing baselines, especially maintaining relatively stable performance under high missing rates.

Abstract:
Internet memes are a pervasive form of multimodal online communication; however, such communication often involves users from diverse linguistic and cultural backgrounds. Therefore, adapting memes across cultures and languages is a central challenge for enabling mutual understanding in online communication. Unlike ordinary translation or standalone text rewriting, cross-cultural meme transcreation must jointly preserve communicative intent, adapt culture-dependent meaning for the target audience, and maintain coherence between text and image. In this work, we first provide an explicit task analysis of cross-cultural meme transcreation and identify three core challenges: Culture-specific knowledge understanding, Intent and Tone Preservation, and Multimodal Consistency. Based on this analysis, we propose a multi-agent framework with specialized agents that are coordinated to address these challenges through cultural adaptation, target text rewriting, revision, and conditional visual adjustment. The framework strengthens target text adaptation with coordinated feedback to handle difficult cases that require deeper cultural or visual intervention. We evaluate the framework on bidirectional Chinese-English meme transcreation using both human evaluation and LLM-as-a-Judge. Our method consistently outperforms all baselines across both human evaluation and LLM-as-a-Judge settings. In human evaluation, it achieves the best performance on all four dimensions and delivers a 33.1% average improvement over the strongest baseline, while in LLM-as-a-Judge, it attains the highest Top-1 ranking rate (60 % vs. 26% for the second-best baseline). Further analysis indicates that each component contributes to the performance, highlighting the effectiveness of the proposed architecture, and our error analysis suggests that the remaining bottlenecks lie in humor reconstruction and image-text alignment rather than simple cultural knowledge gaps, pointing to future work on humor transfer.

Abstract:
Dynamic Facial Expression Recognition (DFER) in the wild remains challenging due to subtle motion cues, severe temporal redundancy, temporal misalignment, and long-tailed class distributions. While masked video modeling has shown strong potential for video representation learning, its conventional pixel-level reconstruction objective is suboptimal for DFER, as it emphasizes low-level appearance statistics rather than motion-sensitive semantics. To address this, we propose Motion-Phase Collaborative Learning (MPCL), a unified two-stage framework for DFER. In the pre-training stage, we introduce Motion-Aware Dual-Region Masking (MDRM), which decomposes motion-augmented videos into facial and trajectory regions and applies region-specific masking via Temporal-Consistent Tube Masking and Trajectory-Guided Sparse Masking to enhance motion-sensitive representation learning. Building on this, we further propose CLIP-Guided Semantic Reconstruction (CGSR), which replaces pixel reconstruction with high-level semantic feature reconstruction under a frozen CLIP teacher, thereby providing semantic targets for motion-sensitive representation learning. In the fine-tuning stage, we develop Phase-Aware Progressive Fine-Tuning (PAPF), which progressively models spatiotemporal dependencies and incorporates a phase-aware supervised contrastive objective to mitigate temporal misalignment and improve robustness under long-tailed distributions. Extensive experiments on FERV39k, MAFW, and DFEW demonstrate that MPCL achieves state-of-the-art performance, with consistent improvements in both weighted and unweighted average recall.

Abstract:
Domain generalization person re-identification (DG-ReID) remains challenging due to severe domain shifts caused by camera variations, illumination changes, and background clutter. Recent CLIP-based methods improve generalization by leveraging vision-language priors, yet they struggle to disentangle domain-invariant identity cues from domain-specific noise, as holistic alignment often suppresses fine-grained discriminative details. We propose a blue novel Frequency-Aware Multi-Granularity Alignment (FAMG) framework that unifies dual-frequency visual purification with multi-granularity vision-language alignment. In the visual branch, we employ the wavelet transform to decompose features into low- and high-frequency components while preserving spatial structure. The high-frequency components are further refined via Fourier transform, where phase information is preserved to retain identity-related structures, while amplitude filtering suppresses domain-specific variations. In the text branch, we design a cross-modal granularity alignment strategy that explicitly associates low-frequency representations with class-level prompts and purified high-frequency features with instance-level prompts. These coarse-to-fine semantic anchors enable more precise cross-modal supervision, preserving both global semantics and subtle identity cues. Extensive experiments on standard DG-ReID benchmarks demonstrate that FAMG consistently outperforms state-of-the-art methods, with notable improvements in cross-domain generalization.

Abstract:
Multi-modal surface understanding requires jointly modeling appearance and geometry, yet reliable cross-modal perception remains challenging on reflective free-form surfaces due to specular artifacts, spatially varying curvature, and non-uniform local geometric support. Existing approaches typically separate semantic segmentation from downstream roughness estimation, which limits both robust 2D--3D fusion and geometry-sensitive surface analysis. In this paper, we propose a unified multi-modal framework for reflective industrial components that links semantic segmentation with roughness estimation. We first introduce a geometry-guided visual attention mechanism that exploits a geometry-informed reliability prior, combining projected geometric responses with visibility and local photometric stability, to suppress unreliable visual responses and facilitate asymmetric 2D--3D feature fusion. We further develop a semantic-aware surface decoupling strategy that transfers semantic priors to downstream roughness estimation, enabling adaptive separation of microscopic roughness from macroscopic form. Experiments on a real-world aero-engine blade dataset demonstrate the effectiveness of the proposed framework. Compared with representative multi-modal baselines, our method improves mIoU by 2.25%. Moreover, the proposed roughness estimation strategy reduces the roughness estimation error from 0.6\mum to 0.28\mum relative to fixed-scale measurement, yielding more stable roughness evaluation in regions with varying curvature.

Abstract:
Generalized animal pose estimation remains a challenging task due to substantial variations in anatomy and appearance across diverse species. Existing methods predominantly rely on visual cues, which limits their generalization to unseen species. Recent studies have explored cross-modal learning paradigms, showing that incorporating textual priors can effectively alleviate these limitations. However, these approaches typically adopt coarse-grained prompts, such as keypoint names or object-level descriptions, which are inherently ambiguous and fail to establish strong semantic alignment with pose representations. To address this issue, we propose HPP (Hierarchical Pose Prompting), a novel framework that leverages hierarchical textual priors to enable explicit and fine-grained cross-modal alignment for generalized animal pose estimation. Specifically, we introduce pose-aware prompts to provide global semantic guidance and adaptive keypoint prompts to capture fine-grained spatial semantics. Furthermore, we incorporate pose activation to highlight pose regions and spatial-aware keypoint adaptation to refine local representations, thereby facilitating cross-modal alignment at both pose and keypoint levels. Furthermore, we design a unified objective to enforce global pose semantic alignment and keypoint fine-grained spatial supervision for cross-modal alignment. Extensive experiments on multi-species animal pose benchmarks demonstrate that HPP consistently outperforms state-of-the-art methods under fully supervised, few-shot, and zero-shot settings.

Abstract:
Cross-Domain Few-Shot Segmentation (CD-FSS) transfers segmentation knowledge learned from large-scale source domains to target domains with only a few annotated samples. However, most previous methods primarily focus on adapting or generalizing source-trained features to target domains. We argue that focusing solely on feature generalization is insufficient; an important but overlooked factor lies in cross-domain support-query correspondence, specifically in its granularity selection. From this perspective, we revisit CD-FSS and identify two critical limitations: (i) modeling correspondence at the coarse image-level granularity is prone to source-domain bias, and (ii) establishing correspondence at the fine-grained fragment-level introduces semantic ambiguity and local noise due to over-segmentation. Motivated by these insights, we advocate a shift toward semantic-level correspondence modeling, where semantically consistent regions provide more stable and accurate cross-domain correspondence than alternative granularities. Building on this paradigm, we propose a Semantic-aware Aggregation and Refinement Network (SARNet) with two components: a Semantic-level Mask Aggregation Module (SMAM) that aggregates consistent and discriminative semantic cues into domain-general semantic masks, and a Semantic-level Feature Refinement Module (SFRM) that captures semantic global context and inter-region relationships for domain robustness. By establishing more generalizable and stable semantic-level correspondences, SARNet achieves state-of-the-art performance on four cross-domain benchmarks.

Abstract:
Learning reliable correspondences between images and point clouds is fundamental for 2D3D matching. Despite recent progress in detection-free methods, existing approaches primarily optimize matching within a single model, and often struggle to maintain reliable correspondences under challenging conditions such as noisy inputs, low overlap, and ambiguous structures. In this work, we propose TeaMatch, a novel framework that introduces teachability as a criterion for cross-modal representation learning. We define teachability as the ability of a representation to be effectively recovered by weak learners under degraded inputs, reflecting its structural consistency and robustness. To this end, we construct a set of task-specific weak students that simulate common failure modes, and train them to imitate the teacher on a training split while evaluating their recoverability on a disjoint meta split. The teacher is then optimized to improve the students ability to recover reliable correspondences, guided by correspondence-level and geometry-aware constraints. Our framework can be seamlessly integrated into existing coarse-to-fine matching pipelines without additional inference cost.Extensive experiments demonstrate that TeaMatch improves matching robustness and achieves state-of-the-art performance on challenging 2D3D matching benchmarks.

Abstract:
Open-vocabulary perception has transformed multimedia understanding by allowing models to operate beyond fixed label spaces, yet olfaction remains stuck in a closed-set regime. Indeed, existing structureodor models can predict only pre-defined categories, making them fundamentally incapable of handling realistic odor queries expressed in natural language. In this work, we address this limitation by introducing zero-shot odor-to-molecule prediction: given a free-form odor description, the goal is to identify molecules that match the queried olfactory concept, even when that concept never appears in training. We propose OdorZero, which aligns molecular representations with language-based odor concepts, and uses a label-semantic graph to transfer predictive knowledge from seen to unseen queries. To balance closed-set discrimination and open-vocabulary generalization, we further design a two-stage training strategy with stochastic path masking. On two benchmarks, OdorZero matches foundation models pretrained on over one billion molecules for standard prediction without requiring large-scale pretraining, while improving over prior graph-based methods by more than 5%. Moreover, OdorZero enables zero-shot prediction on entirely unseen odor categories that existing closed-set methods cannot handle by design, opening up molecular-language alignment as a promising direction for cross-modal research.

Abstract:
Vision-Language Models (VLMs) are highly effective in retrieving semantically relevant images. However, in practice, relevance alone is often insufficient. Systems must also achieve Result Diversification (RD) across composite attributes such as geography and time, a task for which precise control remains challenging. Current re-ranking methods, such as Multi-Source Determinantal Point Processes (MS-DPP), address this using manifold-based repulsion over similarity representations. Although this strategy is effective for broad exploration, it exposes a key limitation in manifold-based models: when subjected to diversity-decrease tasks on discrete metadata, they suffer substantial degradation in early-rank recall. To bridge this gap, we introduce MASCOT (Model-Aware Submodular Coverage for Composite-Attribute Text-to-Image Retrieval). Instead of relying on manifold repulsion, MASCOT formulates multi-attribute diversity as a resource allocation problem, projecting attributes into a soft-binning space weighted by query-driven importance. Averaged across the three PixelProse diversity-decrease tasks, MASCOT preserves an early-rank recall (R@10) of 88.58%, while MS-DPP retains 67.63%. The margin widens under composite constraints: on PP_geo_hour, where temporal and geographic diversity must be suppressed simultaneously, MS-DPP's recall collapses from 0.9737 to 0.4931 and its top-ranked result degrades to R@1 = 0.23, while MASCOT holds R@10 = 0.9410 and R@1 = 0.7202 at a diversity metric above the unconstrained baseline. We do not claim uniform superiority: on aggregate diversity-relevance scores our own simpler ablations attain higher harmonic means on all three decrease tasks, and MASCOT's advantage is specific to recall beyond rank 1 under composite constraints.

Abstract:
Similes provide a compact and expressive way to describe visual characteristics in text prompts. Recent text-to-image models (t2i models) can produce visually compelling outputs from simile prompts, yet even frontier models frequently misinterpret the metaphorical vehicle and confuse it with the object. These systematic failures reveal a gap between figurative language and object-level visual grounding in t2i models. To investigate this issue, we propose a scalable evaluation framework for simile understanding. Our framework includes (1) a controlled simile dataset in which metaphorical vehicles are drawn from a predefined set of object-detectable categories and combined with diverse templates, (2) automatic grounding metrics based on YOLO (You Only Look Once) detection, and (3) text encoder layer analysis using Diffusion Lens to track how metaphorical vehicles emerge during generation. Experiments across architecturally diverse t2i models reveal consistent literalization failure patterns. We further discuss potential mitigation strategies for improving simile grounding in t2i models.

Abstract:
Low-light image enhancement remains challenging due to the spatially heterogeneous nature of degradation, as different regions within the same image often require substantially different degrees of illumination correction and detail restoration. However, existing methods typically adopt spatially uniform processing, which is often inadequate for such diverse restoration demands. In this paper, we propose a frequency-aware and content-prioritized framework for low-light image enhancement. Specifically, the input image is decomposed into low- and high-frequency components via the discrete wavelet transform, enabling illumination-dominant and detail-dominant information to be restored separately. Building upon this decomposition, we introduce a Complexity-Brightness Dual-Aware State Space Group (CBDA-SSG), which performs selective patch-wise enhancement in a coarse-to-fine manner by prioritizing low-frequency patches with low brightness and high-frequency patches with high structural complexity. Furthermore, to mitigate the inconsistency between the two independently enhanced frequency branches, we propose an Inter-Frequency Attention Enhancement Module (IFAEM) that enables bidirectional cross-frequency refinement. By integrating content-prioritized restoration and cross-frequency interaction, the proposed framework achieves more targeted illumination adjustment and detail recovery. Extensive experiments on multiple benchmark datasets demonstrate that our method achieves superior visual quality and competitive quantitative performance compared with existing state-of-the-art approaches.

Abstract:
Accurately predicting the popularity of micro-videos is a critical but challenging task, characterized by volatile, `rollercoaster-like' engagement dynamics. Existing methods often fail to capture these complex temporal patterns, leading to inaccurate long-term forecasts. This failure stems from two fundamental limitations: \ding172 a superficial understanding of user feedback dynamics, which overlooks the mutually exciting and decaying nature of interactions such as likes, comments, and shares; and~\ding173 retrieval mechanisms that rely solely on static content similarity, ignoring the crucial patterns of how a video's popularity evolves over time. To address these limitations, we propose M^3TR, a Temporal Retrieval enhanced Multi-Modal framework that uniquely synergizes fine-grained temporal modeling with a novel temporal-aware retrieval process for Micro-video popularity prediction. At its core, M^3TR introduces a Mamba-Hawkes Process (MHP) module to explicitly model user feedback as a sequence of self-exciting events, capturing the intricate, long-range dependencies within user interactions (for limitation \ding172). This rich temporal representation then powers a temporal-aware retrieval engine that identifies historically relevant videos based on a combined similarity of both their multi-modal content (visual, audio, text) and their popularity trajectories (for limitation \ding173). By augmenting the target video's features with this retrieved knowledge, M^3TR achieves a comprehensive understanding of prediction. Extensive experiments on two real-world datasets demonstrate the superiority of our framework. M^3TR achieves state-of-the-art performance, outperforming previous methods by up to 19.3% in nMSE and showing significant gains in addressing long-term prediction challenges.

Abstract:
\beginabstract Semi-supervised semantic segmentation (SSSS) aims to achieve accurate pixel-level prediction with limited labeled data and abundant unlabeled data. Most existing approaches rely on pseudo-label generation and consistency regularization purely within the spatial RGB domain. However, the inherent entanglement of low- and high-frequency components in spatial features makes it difficult to distinguish stable semantic structures from noisy boundaries, often leading to pseudo-label degradation near object edges. To address these issues, we propose SG-FASS, a unified framework that introduces hierarchical frequency-domain modeling throughout both feature representation and pseudo-label learning. SG-FASS employs a Frequency Decomposition (FD) module to explicitly disentangle low-frequency semantics and high-frequency boundaries for decoupled feature representation. A Spatial-Adaptive Frequency Selection (SAFS) module further adjusts frequency responses dynamically according to spatial context, enhancing intra-region semantic consistency and suppressing boundary errors. Additionally, a Global Frequency Calibration (GFC) module is incorporated to align spectral statistics across labeled and unlabeled domains, improving global stability and pseudo-label reliability. Extensive experiments across multiple benchmarks demonstrate that SG-FASS consistently achieves substantial improvements over state-of-the-art methods under various labeling ratios, validating its effectiveness. \endabstract

Abstract:
Image restoration aims to recover clean, high-fidelity images from degraded inputs. However, existing vision-only methods lack explicit semantic context, often struggling to distinguish degradation artifacts from intrinsic image textures. While recent multimodal approaches introduce language guidance, they typically rely on coarse holistic descriptions and positive-only supervision, which limits fine-grained semantic alignment under severe degradations. To overcome these limitations, we introduce the Unified Multimodal Restoration 1-Million (UMR-1M) dataset, a large-scale multimodal dataset tailored for low-level image restoration. UMR-1M goes beyond standard positive captions by incorporating hallucinated hard negatives and fine-grained region grounding, enabling more discriminative cross-modal alignment. Built on this foundation, we propose Language-Guided Restoration Alignment Pre-training (LG-RAP), which distills grounded multimodal priors into the visual backbone. Furthermore, we present Semantic-Informed Grounded Multimodal Alignment (SIGMA), a CNN-ViT hybrid restoration framework built upon this alignment paradigm for high-fidelity pixel-wise reconstruction. Extensive experiments on deraining, desnowing, denoising, and deblurring benchmarks demonstrate that SIGMA achieves competitive performance across multiple restoration tasks.

Abstract:
We present DynActiveGS, a dynamic-aware active reconstruction framework based on 3D Gaussian Splatting (3DGS) for autonomous exploration in dynamic environments. The framework incrementally reconstructs a 3D Gaussian scene representation while suppressing motion-corrupted observations through online uncertainty prediction and uncertainty-weighted Gaussian optimization. A key component of DynActiveGS is the explicit decomposition of uncertainty into structural uncertainty and motion-induced uncertainty, which enables the system to distinguish under-reconstructed static regions from dynamically unreliable areas. Based on these uncertainty fields, DynActiveGS performs dynamic-aware viewpoint selection and dynamic-constrained path planning to favor informative yet stable observations during exploration. The resulting system forms a unified closed-loop pipeline for robust active reconstruction in dynamic scenes. Extensive experiments on challenging dynamic benchmarks demonstrate consistent improvements over existing active reconstruction baselines in reconstruction accuracy, completeness, rendering quality, and exploration efficiency.

Abstract:
The human hand is our primary means of interacting with the environment, making photorealistic and personalized 3D hand modeling essential for understanding human interaction from images. The recent 3D Gaussian Splatting method combines the advantages of primitive-based and volumetric representations, achieving state-of-the-art rendering quality and speed. However, existing methods applying 3D Gaussians to hand modeling overlook the hands underlying structure, limiting adaptability to complex articulations. In this paper, we introduce DAGS, a novel framework that employs control vertices to distribute local 3D Gaussians, where each vertex carries weight features to guide non-rigid deformations. In particular, we introduce a Learnable Skinning Correction (LSC) module that predicts corrective weights, enabling personalized and anatomically consistent deformations. Benefiting from its explicit construction, the DAGS-based hand model can be seamlessly integrated into existing hand pose estimation pipelines, thereby improving the accuracy of pose estimation. Extensive experiments validate the effectiveness of the proposed method in generating high-fidelity 3D hand representations and consistently enhancing the performance of state-of-the-art pose estimation approaches.

Abstract:
Dynamic 3D Gaussian Splatting (3DGS) is becoming an important representation for dynamic scene rendering and reusable 3D content. Yet most existing dynamic-scene resources are released as videos or synchronized multi-view observations rather than as native Gaussian assets. This limits their utility for asset-level studies, where researchers operate directly on reconstructed dynamic Gaussian primitives and would otherwise need to introduce an additional reconstruction and export stage before any asset-level study can begin. We present DyGS-Dataset, a native dynamic 3DGS dataset in which each sequence is represented as a temporally organized collection of explicit per-frame Gaussian assets. DyGS-Dataset is built through a two-stage pipeline: we first reconstruct a temporally coherent anchor-based teacher sequence and then convert each frame into standard native 3DGS primitives. This design preserves strong visual quality while improving the temporal stability of the released assets. In addition to the Gaussian assets, DyGS-Dataset includes aligned camera metadata, semantic augmentation, and proxy geometry, together with quantitative and qualitative validations on rendering fidelity, temporal behavior, and representative asset-level studies. These results indicate that DyGS-Dataset is a useful native dynamic Gaussian resource for asset-level studies of rendering, semantic access, interaction, and compression.

Abstract:
Cross-modal retrieval has made remarkable strides with embedding-based methods that map different modalities into a unified space. However, these approaches fail on compositional queries that require understanding multiple entities and their relationships. The root cause is the flat feature paradigm, which compresses structured semantics into single vectors, irrevocably losing relational information. We present GraphMatch, a neural subgraph alignment framework that explicitly preserves and matches compositional structure. Our key innovations are threefold: (i) a dual-channel architecture that dynamically routes simple queries to efficient embedding matching and complex queries to structured subgraph alignment, (ii) a query-to-graph parser that transforms natural language into executable query subgraphs with entities, attributes, and relations, and (iii) a differentiable subgraph matching algorithm that learns to align query graphs with subgraphs through joint optimization of node semantics and structural consistency. This enables retrieval to understand not just what entities are present, but how they relate. Extensive experiments show GraphMatch achieves state-of-the-art on standard tasks while delivering a 19.2% absolute improvement on complex compositional queries, closing the gap between simple embedding matching and true relational understanding. Our work establishes neural subgraph alignment as a new paradigm for fine-grained cross-modal retrieval where structure matters.

Abstract:
Language-guided 3D affordance grounding is crucial for embodied multimedia applications. However, real-world 3D point cloud acquisition is often affected by severe structural corruptions, such as sensor jitter, physical occlusions (local dropout), and scanning sparsity (global dropout). Existing methods tightly couple language prompts with local 3D geometry, making them highly vulnerable to semantic degradation when the underlying structures are corrupted. To address this limitation, we propose DeGSA, a robust framework built upon a novel disentangle-and-route paradigm. Specifically, instead of indiscriminately fusing multimodal features, DeGSA explicitly decomposes the language prompt into geometry-sensitive descriptions and topology-agnostic action intents. To effectively leverage this disentangled representation, we further introduce a Confidence-Aware Gated Fusion mechanism. Under clean conditions, it adaptively integrates both semantic branches for accurate affordance grounding; under severe structural corruption, it suppresses unreliable geometric cues and routes prediction primarily through the more resilient action prior. Extensive experiments on both standard and corrupted benchmarks (LASO-C and PIAD-C) show that DeGSA consistently achieves state-of-the-art performance and exhibits strong robustness against extreme real-world 3D structural noise.

Abstract:
\beginabstract Although existing generative models such as GAN and Diffusion have made significant progress in image illumination harmonization task, they still struggle to generate harmonious foreground appearance with accurate geometric cast shadow and the intensity matching illumination. To this end, we propose a novel physics-guided illumination diffusion model named PGIDiff. Our PGIDiff separately constructs accurate physical priors of casting shadows and illumination to effectively guide the diffusion model to generate realistic illumination-shadow effects for foreground objects. Specifically, we first use image depth information to infer scene illumination position, which is further combined with the foreground geometry to calculate the direction and scale of the foreground cast shadow through perspective theory, thereby accurately estimating foreground cast shadow space information. Meanwhile, we also propose an illumination field module that learns the illumination pattern from the background for predicting reasonable illumination intensity. Then, with the foreground cast shadow space and illumination intensity information as prior conditions, we inject them into the diffusion model for effectively guiding the generation of illumination consistent foreground and reasonable cast shadows. Extensive experiments on public image harmonization datasets DESOBAv2 and iHarmony4 and real illumination harmonization dataset IH-SG show that our PGIDiff achieves superior results. \endabstract

Abstract:
Generating scenes with reasonable layouts and semantic coherence is essential for embodied AI, virtual reality, and augmented reality. This is particularly critical in language-guided 3D scene editing, where modifying one object can alter the spatial organization of the scene. However, existing methods mainly focus on the edited object and neglect overall scene adjustment, often leading to unreasonable layouts and degraded semantic coherence. To address this issue, we propose CausalScene, a unified framework that models the triggerpassive causal structure through anchor-guided diffusion, enabling coherent scene edits without manual intervention. Specifically, CausalScene conditions both a scene graph diffusion model and a layout diffusion model on an anchor encoding the edited object and edit type, enabling coherent passive adjustment for consistent target scene generation and extending to real-world object insertion without manual annotation. To overcome the lack of training data for passive adjustment in 3D scene editing, we develop an automated data generation pipeline that synthesizes physically plausible and semantically coherent editing pairs and introduce CausalScene-DB, a large-scale dataset of 93k editing pairs for learning coherent passive responses. Experiments show that CausalScene consistently outperforms existing baselines across all metrics, producing more accurate and coherent language-guided 3D scene edits.

Abstract:
The rapid evolution of Multimodal Large Language Models (MLLMs) is bottlenecked by the saturation of high-quality public data, while vast amounts of heterogeneous multimodal data remain inaccessible in privacy-sensitive silos. Federated Learning (FL) offers a promising solution to leverage these distributed resources while preserving privacy, but existing research focuses predominantly on fine-tuning, leaving the foundational pre-training phase largely unexplored. We argue that the core of MLLM pre-training lies in establishing effective cross-modal alignment, which bridges visual and textual representations into a unified semantic space. Motivated by this insight, we formally introduce the Federated Post-Alignment (Fed-PA) task, which studies cross-modal alignment under federated MLLM pre-training by collaboratively training a shared cross-modal projector. We identify two critical challenges in this setting: (i) parameter interference in aggregating local projectors; and (ii) gradient oscillations in one-pass collaborative SGD. To address these challenges, we propose FedPivot, a pioneering framework for federated post-alignment. FedPivot employs Canonical Reliability-Aware Aggregation (CRA), which constructs a canonical space to decompose client projectors into a shared alignment basis and client-specific coefficients, then performs reliability-weighted fusion to suppress parameter interference. Furthermore, FedPivot introduces Orthogonality-Preserved Momentum (OPM), which applies momentum to the shared alignment basis via orthogonal projection, accumulating historical optimization directions while preserving geometric structure. We construct four federated multimodal pre-training scenarios based on public datasets, and extensive experiments validate that FedPivot significantly outperforms existing baselines.

Abstract:
Audio-visual speech separation (AVSS) is trained to reconstruct the target speech from a speech mixture, yet its intermediate audio representation remains incomplete and interference-contaminated. Under this condition, repeatedly enforcing cross-modal interaction can become counterproductive: the current audio state progressively filters the visual stream, causing target-related visual cues that have not yet established correspondence with the audio branch to be prematurely weakened. This reduces the complementary role of the visual modality and limits the benefit of audio-visual fusion. To address this overlooked issue, we propose a \emphControllable Consistency Repair (CCR) module for AVSS. Rather than enforcing unconditional interaction, CCR monitors intermediate cross-modal consistency with a shift-tolerant score derived from the leading singular values of the channel correlation matrix between audio and visual features, and activates repair only when the score falls below a running reference. Once triggered, CCR performs lightweight repair through cross-modal gating, bounded residual correction, and accept-if-improved verification, turning cross-modal correction into a measurable, triggerable, and verifiable process. We integrate CCR into the audio-visual encoder of a time-frequency AVSS framework without introducing direct supervision for audio-visual matching. Experiments on standard AVSS benchmarks show strong overall performance and improved robustness under temporal misalignment, strong interference, and degraded visual conditions.

Abstract:
Modern online shopping platforms increasingly embed product descriptions and specifications within images, making their text content inaccessible to conventional keyword-based search engines. This leads to degraded search engine optimization (SEO) and limited visibility in generative engine optimization (GEO) environments. To address this problem, we introduce a novel, style- and alignment-aware framework that reconstructs structured web layouts from product page images, enabling machine-readable access to text content otherwise locked inside images. Our framework integrates (1) hierarchical structure-aware text extraction, (2) vision-language model (VLM)-based text clustering informed by visual attributes, and (3) segmentation-guided text removal and inpainting for clean background restoration. We also construct a dedicated ECTL dataset, a style-annotated product page dataset with word-level labels for style clustering, alignment clustering, and text filtering. Experimental results demonstrate accurate color and font recognition (99.7% and 99.3% Top-1 accuracy), clean text removal (FID\,=\,0.954 on the Easy subset), and visually faithful reconstruction (LPIPS\,=\,0.0399, SSIM\,=\,0.9394). QA-based evaluations across three foundation models show that reconstructed layouts improve factual retrievability (e.g., BERT-Score 0.883 \to 0.928 on Gemma-3-27B) while reducing input tokens by 11--14% relative to OCR-only inputs.

Abstract:
The rapid evolution of video generation technologies poses a significant challenge to media forensics, as conventional detection methods often fail to generalize beyond their training distributions. To address this, we propose DVAR (Debate-based Video Authenticity Reasoning), a training-free framework that reformulates video detection as a structured multi-agent forensic reasoning process. Moving beyond the paradigm of pattern matching, DVAR orchestrates a competition between a Generative Hypothesis Agent and a Natural Mechanism Agent. Through iterative rounds of cross-examination, these agents defend their respective explanations against abnormal evidence, driving a logical convergence where the truth emerges from rigorous stress-testing. To adjudicate these conflicting claims, we apply Occam's Razor through the Minimum Description Length (MDL) framework, defining an Explanatory Cost to quantify the "logical burden" of each reasoning path. Furthermore, we integrate GenVideoKB, a dynamic knowledge repository that provides high-level reasoning heuristics on generative boundaries and failure modes. Extensive experiments demonstrate that DVAR achieves competitive performance against supervised state-of-the-art methods while exhibiting superior generalization to unseen generative architectures. By transforming detection into a transparent debate, DVAR provides explicit, interpretable reasoning traces for robust video authenticity assessment.

Abstract:
Streaming video understanding requires models to process unbounded visual streams while preserving rich visual semantics across vast temporal horizons, posing a fundamental challenge for memory modeling. Existing approaches primarily focus on increasing memory capacity, either by compressing historical information into fixed-size representations or by extending storage beyond GPU memory. However, these methods largely rely on global or coarse-grained representations, inevitably losing fine-grained visual information. In this work, we argue that streaming video memory should explicitly encode structured and semantically meaningful representations, particularly at the entity level. To this end, we propose MEMO, a novel framework that models streaming video through multi-level, entity-aware structured memory. MEMO performs multi-level perception to jointly capture global semantics, entity dynamics, and spatial structures, partitioning streaming video into semantically coherent chunks. Each chunk is organized into a structured memory, where lightweight global and entity-level representations serve as retrieval indices, while the corresponding high-resolution visual content is retained separately for on-demand access. At inference time, MEMO performs query-specific retrieval over the structured memory and selectively recalls relevant visual evidence for downstream reasoning. Notably, MEMO is training-free and plug-and-play with existing multimodal large language models. Extensive experiments on StreamingBench and OVO-Bench demonstrate that MEMO consistently improves multiple base models and achieves state-of-the-art performance.

Abstract:
Joint reconstruction of human-object interaction (HOI) provides crucial cues for understanding the intrinsic relationships between humans and objects. Existing HOI reconstruction methods typically rely on explicit or implicit contact modeling to characterize human-object interactions; however, their contact-driven reconstruction paradigm limits generalization to non-contact interactions. To extend interaction generalization, we introduce LTF-HOI, a framework that learns a tightness field for human-object interaction reconstruction, elevating interaction semantics from contact prediction to tightness modeling. Specifically, we propose an interaction tightness field, in which a tightness field predictor estimates a direction-magnitude-semantics-confidence quadruple for each human and object vertex, thus enabling accurate reconstruction of both contact and non-contact interactions. Although the tightness field provides fine-grained vertex-level information, its dense representation is not conducive to global geometric reasoning. To address this limitation, we further introduce a marker-aggregation refinement transformer, which aggregates the tightness field into a set of semantically consistent markers and leverages these key markers to guide the refinement of interaction-driven reconstruction. Extensive experiments on BEHAVE and InterCap datasets validate the effectiveness of our method.

Abstract:
Recently, multimodal large language models (MLLMs) have exhibited strong potential in the automatic analysis of Optical Coherence Tomography (OCT) images, especially in predicting the progression of Age-related Macular Degeneration (AMD). However, despite their advanced visual-linguistic capabilities, they struggle with interpreting fine-grained OCT images that require specialized reasoning steps. To address this, we propose Macula-CoT, a macular subregion-aware chain-of-thought model for AMD prognostic prediction. By learning reasoning chains that align with the diagnostic process of ophthalmologist, it allows an accurate, transparent, and credible clinical reasoning. Specifically, we first introduce a multi-stage reasoning strategy that guides MLLMs to sequentially identify key macular anatomical subregions, thereby aggregating structured clues with fine-grained representations for disease prediction. Additionally, to ensure the understanding of detailed OCT macular regions, we propose a Patch-to-Region (P2R) module for visual encoding. It simplifies the encoding of irregular, complex macular patterns with explicit patch-level markers, enabling an effective perception and discrimination of local regions. In this way, various macular regions can be seamlessly modeled and captured in a structured manner, providing enriched contexts for disease prediction and interpretable reasoning chains for clinical reference. We evaluate Macula-CoT on two AMD prognosis datasets, achieving SOTA performance in both accuracy and interpretability for subretinal fluid absorption and visual acuity improvement prediction. The code will be publicly available.

Abstract:
With the rapid development of generative models and multimodal content editing technologies, the key challenge in synthetic image detection (SID) lies in cross-distribution generalization to unknown generation sources. Visual foundation models (VFM), which acquire rich visual priors through large scale image-text alignment pretraining, have become a promising approach to improving SID generalization. However, existing VFM-based methods remain coarse-grained in their adaptation strategies. They typically either directly use the final layer representations of VFM or simply fuse multi layer features, lacking explicit modeling of the optimal representational hierarchy for transferable forgery cues. Although fine-tuning VFM can enhance task adaptation, it may also damage the cross-modal pretrained structure that supports open-set generalization. To address this task specific tension, we reformulate VFM adaptation for SID as a joint optimization problem: it is necessary both to identify the critical representational layer that is more suitable for carrying forgery discriminative information and to constrain the disturbance caused by task knowledge injection to the pretrained structure. Based on this, we propose I2P, an SID framework centered on intrinsic importance perception. I2P first adaptively identifies the critical layer representations that are most discriminative for SID, and then constrains task-driven parameter updates within a low sensitivity parameter subspace, thereby improving task specificity while preserving the transferable structure of pretrained representations as much as possible. We conduct systematic experiments and ablation analyses on multiple cross-domain benchmarks, and the results show that I2P achieves stable and competitive generalization performance.

Abstract:
Visual motifs such as ornaments, figures, symbols, and scene elements are fundamental units of artistic creation, digital curation, and cultural media reuse, yet building reusable motif-level archives from motif-rich images remains underexplored. Existing generative methods are poorly suited to this goal: full-model tuning and LoRA-based approaches are costly and difficult to organize into lightweight archives, while token-learning methods become unstable when multiple motifs are learned from the same image, often causing spatial leakage, semantic overexpression, and inter-motif semantic conflict. We present MotifArchive, a framework for constructing reusable visual motif token archives for artistic creation. Starting from a source image, users specify motifs of interest through simple point-based interaction, after which motif masks and semantic anchors are automatically obtained. Our method then learns a dedicated token for each motif while keeping the generative backbone frozen, and stores the learned tokens as a lightweight local archive that can be reloaded for later creative use. We introduce Pyramidal Attention Alignment (PAA) for coarse-to-fine motif-region correspondence, Anchor-guided Norm-and-Conflict Regularization (ANCR) for suppressing semantic overexpression and improving semantic decoupling, and Manifold Decomposition Steering (MDS) for anchor-guided test-time adjustment that corrects semantic drift while preserving motif-specific identity. MotifArchive enables archive-based creative authoring, motif reuse, and interpolation-driven artistic exploration from lightweight local motif token archives.

Abstract:
Rapidly evolving Generative AI enables sophisticated visual text manipulations that increasingly evade current forensic detectors. Existing discriminative models often overfit specific forgery patterns, limiting their generalization to unseen, open-set attacks. To address this challenge, we propose a generative detector that localizes tampering by estimating the local restoration cost required to align a query image with authentic visual-text statistics, rather than by learning forgery-specific decision boundaries. Specifically, we introduce Sparse-Constraint Rectified Flow (SC-RF), a detector-oriented adaptation of Flow Matching for spatially sparse anomaly localization. We further mitigate data scarcity via self-supervised Artifact Injection and preserve high-frequency forensic traces using a pixel-space Forensic-DiT. Extensive experiments on three benchmarks show that our method achieves state-of-the-art performance, surpassing the runner-up by 3.2 and 4.8 percentage points in F1 and IoU, respectively. In particular, the proposed detector demonstrates strong zero-shot performance on challenging unseen text editing patterns. We further provide an auxiliary stress-test analysis showing that local harmonization produced by our model can weaken the statistical cues relied upon by existing detectors, offering a complementary vulnerability-analysis perspective.

Abstract:
Creating a compelling social media video from raw clips requires decisions at multiple levels, from story structure and pacing to clip selection and local editing effects. Existing language-based editing systems typically respond to explicit edit commands, retrieve clips, or generate videos directly, but they do not produce an inspectable plan that can be optimized for engagement. We present LoopEdit, a hierarchical multi-LLM framework that converts a user directive and a set of input assets into an editable short-video plan and a rendered video. LoopEdit first generates a high-level plan that defines the video arc, scene structure, pacing, and style, and then generates a low-level plan that maps scenes to specific clips, durations, effects, text, and narration. To train this system, we reformulate existing short-videos into asset-plan pairs through reverse-engineering. We also learn evaluator models from human annotations to predict how well a candidate video plan will perform, capturing both overall engagement and fine-grained attributes that drive it. These learned evaluators also serve as reward models to align the hierarchical planners with human preferences through reinforcement learning (GRPO). In our experiments, the learned evaluators correlate strongly with human judgments, and the optimized planners outperform strong zero-shot Multimodal LLM baseline in automatic evaluation and user studies. These results show that data-driven hierarchical planning with preference alignment can produce more engaging and controllable short-form videos.

Abstract:
Recent Fake News Detection (FND) pipelines increasingly leverage Large Language Models (LLMs) and Vision-Language Models (VLMs) for reasoning-based fake news detection. However, without ground-truth Chain-of-Thought (CoT) supervision, how to iteratively improve reasoning quality through self-reflection, how such improved CoT can benefit downstream model finetuning, and how to leverage large models to better connect single-sample fraudulent pattern discovery with cross-sample verification remain open problems. We propose SRM-FND, a self-reflective multi-modal reasoning framework for short-video fake news detection. SRM-FND constructs self-reflective and quality-improved reasoning through contrastive deliberation with iterative root-cause diagnosis and corrective prompt refinement, where a Blind Analyst, a Counter-Conclusion Reasoner, and a Self-Consistency Arbiter collaboratively retain discriminative rationales.It further adopts dual-phase topic-adaptive VLM finetuning for multi-modal grounding and lightweight topic specialization, and performs confidence-driven cross-sample review by retrieving co-event credible and suspicious witnesses for uncertain cases. Experiments on FakeSV and FakeTT demonstrate that SRM-FND outperforms strong baselines while yielding more reliable, interpretable predictions, and also provide noticeable improvement in cross-dataset performance.

Abstract:
The proliferation of AIGC-driven face manipulation and deepfakes poses severe threats to media provenance, integrity, and copyright protection. Existing versatile watermarking systems typically rely on embedding explicit localization payloads, which introduces a fidelity--functionality trade-off: larger localization signals degrade visual quality and often reduce decoding robustness under strong generative edits. Moreover, these methods rarely support content recovery, limiting their forensic value when original evidence must be reconstructed. To address these challenges, we present VeriFi, a versatile watermarking framework that unifies copyright protection, pixel-level manipulation localization, and high-fidelity face content recovery. VeriFi makes three key contributions: (1) it embeds a compact semantic latent watermark that serves as a content-preserving prior, enabling faithful restoration even after severe manipulations; (2) it achieves fine-grained localization without dedicated payloads by correlating image features with decoded provenance signals; and (3) it introduces an AIGC attack simulator that combines latent-space mixing with seamless blending to enhance robustness against realistic deepfake pipelines. Extensive experiments on CelebA-HQ and FFHQ demonstrate that VeriFi consistently outperforms state-of-the-art baselines in watermark robustness, localization accuracy, and recovery quality, providing a practical and verifiable defense for deepfake forensics.

Abstract:
Deep Unfolding Networks (DUNs) have emerged as a promising framework for compressive sensing (CS), effectively combining iterative optimization principles with deep learning techniques. Nevertheless, existing DUN methods still suffer from the following two challenges: i) The majority of CS algorithms restrict their inference to a single solution space, overlooking the multi-solution property caused by the ill-posed nature of CS inverse problem. ii) Current CS approaches do not adequately leverage the underlying relationships among multiple potential solutions, limiting their ability to efficiently explore informative knowledge. To address these limitations, a novel Multi-Expert Collaborative Deep Unfolding Network (MEC-DUN) is proposed. By unrolling the classical Proximal Gradient Descent (PGD) algorithm, the proposed MEC-DUN introduces multiple expert networks that collaboratively construct and aggregate multiple candidate solutions across multi-scale spaces. Specifically, regarding the gradient descent process, a Multi-Expert Collaborative Gradient Descent Module (MEC-GDM) is proposed, which incorporates a well-designed step size generation network to collaboratively perform dynamic gradient updates within the multi-solution space constructed by multiple experts. On the other hand, corresponding to the proximal operation, a Multi-Expert Collaborative Proximal Mapping Module (MEC-PMM) is developed, which hierarchically integrates multi-scale candidate solutions via an elaborated pixel-wise gating mechanism, balancing structural consistency and complementary diversity across different solutions and scales. Extensive experiments on multiple datasets demonstrate that MEC-DUN achieves superior reconstruction quality and generalization performance compared to existing image CS methods. The code will be made publicly available on MEC-DUN.

Abstract:
Parameter-efficient fine-tuning like low-rank adapters is widely used to adapt multimodal large language models to downstream tasks, yet we identify an unaddressed failure mode called Non-Textual Norm Collapse (NTNC). NTNC causes non-textual modality representations to shrink to near-zero after low-rank projection, leading to lost feature discriminability, vanishing non-textual gradients, and text-dominated optimization that impairs cross-modal adaptation. To resolve this, we propose a geometry-aware framework optimizing adapter updates directly on the fixed-rank matrix manifold, with two key mechanisms: Riemannian Gradient Projection strictly preserves the rank of modality-specific adapters to avoid non-textual feature collapse, and Modality-Aware Momentum Balancing maintains independent Riemannian momentum buffers for modalities and adaptively rebalances their contributions to break text-dominated optimization. Extensive experiments on visual-text, audio-visual-text and speech-text benchmarks show our method consistently outperforms state-of-the-art multimodal PEFT and Riemannian optimization approaches.

Abstract:
Unsupervised Visible-Infrared person Re-Identification (US-VI-ReID) aims to learn robust cross-modal representations without identity labels, reducing reliance on costly manual annotation. However, existing methods rely solely on visual cues, which are hampered by the immense modality gap, causing the learned representations to lack fine-grained, modality-invariant details. To address this challenge, we propose an Attribute-Guided Attention Regularization Network (AGAR-Net), which incorporates the text modality as a semantic bridge to facilitate cross-modal association and learning. Specifically, we first propose an LLM-based Semantic Parser (LSP) that processes auto-generated BLIP captions into a structured attribute space. Building on this foundation, we propose a Cluster-level Attribute Matching (CAM) that fuses visual and attribute similarities to improve pseudo-label quality during cross-modal matching. Then, we propose a BERT-enhanced Attribute Attention (BAA) Module that utilizes these attributes as semantic queries to ground the model's attention onto specific visual regions. This forces the model to capture semantically-aligned local details, enhancing the extraction of fine-grained features for robust recognition. Finally, extensive experiments on the SYSU-MM01, RegDB and LLCM datasets validate the superiority of our AGAR-Net over several other state-of-the-art methods.

Abstract:
Mixture-of-Experts (MoE) Vision-Language Models (VLMs) are powerful but extremely large and expensive. Post-Training Quantization (PTQ) offers a training-free compression pathway. However, two critical and interrelated problems undermine its effectiveness on MoE VLMs. The first is quantization-induced routing instability. Small weight perturbations alter router logits near narrow selection boundaries. This redirects tokens to suboptimal experts and cascades errors across layers. The second is modality-mismatched calibration. Uniform quantization strategies fail to account for the distributional discrepancy between vision and language tokens. They also neglect non-uniform expert activation patterns. We propose Routing-Stabilized Modality-Aware Quantization (RSMAQ), a unified PTQ framework that addresses both challenges. It does so through three synergistic components. Routing Margin Preservation (RMP) enforces logit distribution alignment via KL-divergence minimization. It also uses expert-set and boundary-expert consistency metrics to protect discrete routing topology. Modality-Specific Expert Calibration (MSEC) constructs a modality-aware Hessian matrix for second-order quantization. It jointly encodes token-expert affinity and per-modality gradient sensitivity, directing the reconstruction budget toward cross-modal critical weights. Global Expert Consensus Smoothing (GECS) derives a shared activation-smoothing matrix by aggregating quantization error across all experts. It weights this aggregation by routing probability and applies an outlier-robust multi-level quantile search. We conduct extensive experiments on Kimi-VL-Instruct, Qwen3-VL-30B-A3B-Instruct, and a LLaVA-MoE-3.6B variant. We evaluate across nine multimodal benchmarks. RSMAQ consistently outperforms state-of-the-art baselines.

Abstract:
Cross-View Geo-Localization (CVGL) supports a wide range of geography applications by establishing precise correspondences between low-altitude platforms and satellite imagery. Despite recent performance gains with advances in visual representation learning, current research remains confined to static settings, neglecting real-world demands for streaming data and dynamic environmental evolution (e.g., urban renewal). To address this, we introduce a pioneering yet practical task termed Lifelong Cross-View Geo-Localization (L-CVGL), which requires continuous adaptation to new domains without forgetting prior knowledge. Specifically, compared to standard lifelong learning, L-CVGL faces two unique challenges in balancing stability and plasticity. First, maintaining stability is severely hindered by inherent visual ambiguity, which creates a densely populated feature space where naive adaptation inevitably induces feature aliasing and destroys historical integrity. Second, promoting plasticity is inherently constrained by the underlying retrieval mechanism, as the strict necessity to preserve relational consistency renders conventional rigid anti-forgetting paradigms ineffective for modeling new domains. To address aforementioned challenges, we propose an innovative Omni-view Subspace-Topological Knowledge Transfer (OST-KT) framework tailored to L-CVGL. Specifically, OST-KT integrates Omni-View Subspace Stability Transfer to preserve historical integrity via imposing constraints in the principal subspace, and On-the-Fly Topological Plasticity Transfer to facilitate flexible adaptation through relational topology alignment. To harmonize these modules, we introduce Dynamic Reliability Calibration, which employs a dual-gating strategy to mitigate negative transfer and ensure reliable supervision. Extensive experiments demonstrate that OST-KT significantly mitigates catastrophic forgetting while effectively adapting to new domains, consistently outperforming state-of-the-art methods.

Abstract:
Multi-task dense prediction has recently achieved significant performance gains but remains challenging in learning fine-grained and discriminative task-specific representations. In this paper, we propose a novel decoder-focused Mixture of Unit-Rank Experts (MURE) for multi-task dense prediction. Inspired by the subspace mixing property of LoRA, we introduce unit-rank experts with a LoRA-style structure into a mixture-of-experts (MoE) framework to facilitate adaptive fine-grained modulation of feature representations. Building upon this design, we develop spatial and channel modulation units in the decoder, which progressively refine task-specific features by capturing structural patterns and semantic relationships, respectively. Furthermore, to mitigate the interference between cross-task generic knowledge and task-specific information in multi-task learning, we employ a hybrid MoE design. Specifically, MURE incorporates both task-shared and task-private expert groups within each modulation unit to enable effective simultaneous modeling of inter-task general- and intra-task specialized representations, thereby promoting more discriminative task-specific features. Extensive experiments on the PASCAL-Context and NYUD-v2 benchmarks demonstrate that our MURE achieves superior overall performance compared to previous state-of-the-art methods.

Abstract:
Coreset selection has emerged as a compelling strategy for the efficient fine-tuning of large foundation models by identifying compact yet informative data subsets. However, we observe that visual fine-tuning exhibits pronounced difficulty heterogeneity across classes. Consequently, existing class-agnostic selection methods disproportionately prune samples from overly easy or hard classes. Furthermore, naive class-uniform strategies fail to address these critical variations, ultimately undermining the representational quality of the coreset. To overcome these limitations, we propose Non-Uniform Class-Wise Coreset Selection (NUCS), a framework that bridges class-level budget allocation with intra-class sample selection. Specifically, we quantify the global difficulty of each class using a winsorized average of per-sample scores. Based on this quantification, NUCS allocates non-uniform selection budgets by assigning higher proportions to more challenging classes, while adaptively select the most informative samples within optimal difficulty ranges for each class. Extensive experiments across a wide range of visual classification tasks, including both standard balanced and challenging long-tailed datasets, demonstrate that our method consistently outperforms state-of-the-art baselines. Notably, our non-uniform budget allocation strategy not only achieves superior accuracy and computational efficiency across diverse pre-trained models (e.g., CNNs and ViTs), but also effectively mitigates classification bias, highlighting its broad potential for the efficient fine-tuning of foundation models.

Abstract:
Recent Large Vision-Language Models (LVLMs) demonstrate remarkable capabilities in image understanding and natural language generation. However, current approaches focus predominantly on global image understanding, struggling to simulate human visual attention trajectories and explain associations between descriptions and specific regions. Inspired by how humans shift gaze to understand a scene, we propose TraceVision, a vision-language model that leverages explicit attention trajectories to guide both language generation and region localization, mirroring human gaze-driven perception. TraceVision employs a Trajectory-aware Visual Perception (TVP) module for bidirectional fusion of visual features and trajectory information. We design geometric simplification method to extract semantic keypoints from raw trajectories and propose a three-stage training pipeline where trajectories guide description generation and region localization. We further extend TraceVision to trajectory-guided segmentation and video scene understanding, enabling tracking and temporal attention analysis. We construct the Reasoning-based Interactive Localized Narratives (RILN) dataset to enhance reasoning and interpretability. Extensive experiments on trajectory-guided captioning, text-guided trajectory prediction, understanding, and segmentation demonstrate that TraceVision achieves state-of-the-art performance, establishing a foundation for interpretable visual understanding.

Abstract:
Deep learning-based character vision studies, e.g., text recognition, character image denoising, and historical text completion, are offering new solutions for learning, managing, and utilizing character resources. However, the performance of these studies peaks with large and balanced datasets, which is a rarity with real-world character datasets, especially for logographic character languages, e.g., Chinese. The imbalance in data distribution of logographic characters is an issue due to differences in character usage frequency and new characters being continuously created. In this paper, we propose a novel method for logographic character recognition, which introduces a multi-modal learning approach using visual semantics and contextual semantics of characters. A novel pre-training strategy is designed to enhance deep visual representations, especially for datasets suffering from issues of imbalanced and rare instances, by extracting the contextual semantics of each character from the corresponding language models. We conduct experiments across various datasets to evaluate our character recognition method and validate the contrastive pre-training strategy through several downstream tasks. Experimental results demonstrate the superiority of our method compared to state-of-the-art methods.

Abstract:
Efficient Transformers for PDE solving typically aggregate mesh points into compact latent tokens to achieve linear complexity. We show that this aggregation acts as a spatial low-pass filter, systematically erasing high-frequency boundary details---a phenomenon we term geometric aliasing. Through diagnostic experiments on standard benchmarks, we confirm that prediction errors of existing methods concentrate near complex geometric boundaries, validating this analysis. To address this, we build upon the slice-based attention framework and introduce two targeted modules. First, Spectrum-Preserving Geometric Attention injects multi-scale geometric encodings into both the slicing assignment and feature reconstruction stages, recovering boundary information while preserving O(N) complexity. Second, a Taylor-Decomposed Feed-Forward Network routes features through linear and non-linear expert paths based on local geometric context, adapting computational capacity to the smoothness of the physical field. Experiments on four standard benchmarks and three large-scale industrial simulations (including a ～300K-node 3D mesh) demonstrate consistent improvements, with up to 12.7% error reduction on standard tasks and 81.3% improvement on surface field prediction, with promising applications in industrial aerodynamic design and digital twin simulation.

Abstract:
Remote virtual reality (VR) streaming offloads rendering to remote servers, enabling high-fidelity experiences on standalone headsets but introduces significant bandwidth and latency challenges. Foveation reduces these costs by prioritizing quality in the gaze-contingent foveal region. However, existing dynamic foveation methods typically treat latency as an external constraint rather than a design parameter. We present a latency-driven optimization (LDO) framework that combines gaze prediction, hysteresis-based stabilization, and adaptive foveal region sizing to reduce gaze-to-render misalignment and temporal instability in remote VR. Evaluated across diverse content and latency conditions, the proposed framework reduces mean gaze error by 24.04% and P95 error by 28.17%, cutting outside-fovea misses from 7.27% to 0.21%. Perceptual quality improves by 14 dB in EWPSNR and 0.20.8 in FovVideoVDP over baseline, with gains growing at higher latencies. A subjective user study confirms statistically significant improvements in perceived visual quality and overall quality over the previous state of the art.

Abstract:
In the field of multi-view learning, incomplete multi-view clustering (IMVC), where some views suffer from missing data, has become an important yet challenging problem. Numerous approaches have been proposed to address this issue, and most of these methods involve imputing the missing view information. However, most existing methods primarily focus on completing the representations of missing samples, while overlooking the structural information associated with those samples. Moreover, many contrastive learning-based approaches treat the consistency constraints among all view pairs as equally reliable, although different views may exhibit different semantic and distributional gaps. Aligning views across a large view gap can propagate unreliable information and weaken the representation learned from more informative views. To address this issue, we propose a Neighbor-Induced Consistency Enhancement for Incomplete Multi-View Clustering termed NICE. During training, NICE estimates the semantic similarity among views and selects reliable neighbor views with smaller view gaps for each target view. The selected neighbor not only guides missing data imputation through its intrinsic local structure, but also serves as the primary counterpart for contrastive and clustering alignment, thereby enhancing cross-view consistency without forcing all views to be over-aligned. Extensive experimental results demonstrate that our approach significantly outperforms state-of-the-art methods on incomplete multi-view learning tasks.

Abstract:
Long video understanding poses a fundamental challenge for large video-language models (LVLMs) due to the overwhelming number of frames and the risk of losing essential context through naive downsampling. Inspired by the way humans watch videos on mobile phones, constantly zooming in on frames of interest, we propose ZoomV, a query-aware temporal zoom-in framework designed for efficient and accurate long video understanding. Specifically, ZoomV operates in three stages: (1) Temporal interests grounding: guided by the query, ZoomV retrieves relevant events and their associated temporal windows as candidates. (2) Event interests spotlighting: within pools of candidate windows, each window is scored through the model itself reflection and filtered accordingly, where higher-confidence windows are more representative. (3) Compact representation: the selected events are encoded and temporally downsampled to preserve critical semantics while significantly reducing redundancy. Extensive experiments demonstrate that ZoomV substantially outperforms prior video-agentstyle approaches. On temporal grounding, ZoomV unlocks the latent capability of LVLMs, achieving an 11.8% mIoU gain on Charades-STA. Remarkably, ZoomV further boosts accuracy on LVBench by 9.7%, underscoring its effectiveness on long-video benchmarks.

Abstract:
Large multimodal models with CoT reasoning improve video question answering, yet the think-once-answer-twice paradigm incurs prohibitive KV cache pressure and decoding overhead. We present DualAnchor, a weight-frozen framework that exploits three structural opportunities of the dual-answer architecture, two strictly unique to it. (1) CAVKC (CoT-Aware Video KV Compression) tiers tokens by saliency, quantizes moderate-saliency values to INT8, and fires a one-shot eviction at low EMA visual-attention ratio, reducing the steady-state visual KV footprint by 42.9%. (2) CACVO (Cross-Answer Consistency Value Cache Optimization) performs five Adam steps on a compact residual at medium-confidence CoT tails, re-anchoring the value cache toward the initial first answer with weights frozen. (3) APASD (Answer-Anchored Phase-Adaptive Speculative Decoding) reuses the first answer as a draft-model-free proposal for high-confidence queries and applies a confidence-adaptive tree topology during CoT generation otherwise. By synergizing APASD with CAVKC, we unlock a distribution alignment where target KV eviction and draft visual sparsification converge to similar active contexts---bridging the perceptual gap and delivering a synergistic 2.12× throughput increase during CoT generation and a 45.5% decoding-phase latency reduction across all queries. Evaluated on five video QA benchmarks with no additional training, DualAnchor achieves 68.1% on VideoMME (+0.8%), 59.7% on Video-MMMU (+1.1%), and 40.2% on MVP (+0.8%), with consistent improvements across MVBench and NExT-GQA.

Abstract:
With the rapid progress of video editing technologies, video object insertion has attracted increasing attention. Existing approaches typically rely on either text instructions, which often lack precise spatial control, or dense masks, which are costly and impractical to annotate manually. To address these limitations, we present Click2Insert, a video object insertion framework that enables precise insertion using only a few point annotations on keyframes. Our framework is also compatible with mask inputs, enabling flexible control and supporting mask-based editing across different scenarios. To train the model, we further construct a large-scale paired insertion dataset containing 1.2M samples through an automated pipeline. Furthermore, we improve sparse-guided video object insertion from three perspectives: condition compatibility, spatial accuracy, and temporal coherence. Specifically, we introduce Task Aware Modulation to support both point- and mask-based conditions within a single model, Spatial-Aware Loss to improve positional accuracy under sparse point supervision, and Intra-Chunk Temporal Attention to enhance temporal consistency across adjacent frames. Extensive experiments demonstrate that our method achieves superior editing performance, enabling precise object insertion in a more user-friendly and flexible manner.

Abstract:
Despite rapid progress in text-driven motion retrieval and generation, effective evaluation remains a critical bottleneck. Existing benchmarks rely on small-scale, in-domain test sets and legacy metrics that fail to assess zero-shot performance or the emergent generalization capabilities unlocked by scaling. To resolve this, we introduce MBench, a comprehensive benchmark for zero-shot compositional motion understanding. Supported by a low-cost markerless capture pipeline, MBench provides 20K diverse motions and a curated human-preference set of 373 motion samples. Using this dataset, we expose the limitations of legacy metrics in zero-shot generalization and develop a human-aligned evaluation protocol. To overcome the vulnerabilities diagnosed by MBench, we introduce a systematic scaling strategysynchronously expanding training data diversity (up to 200 hours) and model capacity across architectures like TMR and MoPa. Consequently, our 300M-parameter ScaTMR establishes a new state-of-the-art for zero-shot generalizationdominating both MBench R-Precision and human preference hit rateswhile our scaled MoPa variant sets new in-domain records on HumanML3D. Ultimately, these results confirm that synergistically scaling data and capacity is indispensable for robust, human-aligned open-world generalization.

Authors: Shunxin Xiao (School of Computer and Information Engineering, Fujian Key Laboratory of Pattern Recognition and Image Understanding), Yanxin Xie (School of Computer and Information Engineering, Fujian Key Laboratory of Pattern Recognition and Image Understanding), Bao Zeng (School of Computer and Information Engineering, Fujian Key Laboratory of Pattern Recognition and Image Understanding), Zihan Fang (Fuzhou University), Jianmin Li (School of Computer and Information Engineering, Fujian Key Laboratory of Pattern Recognition and Image Understanding), Da-Han Wang (School of Computer and Information Engineering, Fujian Key Laboratory of Pattern Recognition and Image Understanding)

Abstract:
Multi-view semi-supervised classification aims to improve classification performance by leveraging complementary information across multiple views when labeled data are limited. Existing contrastive learning-based methods commonly align representations across views, but most of them focus only on the final embedding layer and overlook the hierarchical semantic structures encoded in intermediate layers. Consequently, cross-view consistency remains insufficiently modeled under limited supervision. To address this issue, we propose Joint Local--Global Contrastive Calibration (JLGCC), which employs view-specific encoders to learn hierarchical representations and performs layer-wise local affinity-guided contrastive calibration by aligning cross-view feature-similarity distributions with pseudo-label affinity distributions, enabling fine-grained semantic alignment at multiple depths. In addition, a global contrastive objective aligns view-specific pseudo-labels with fused consensus pseudo-labels at the semantic level, mitigating the adverse effects of noisy pseudo-labels. Extensive experiments on six benchmark datasets show that JLGCC achieves superior or competitive performance, demonstrating its effectiveness under limited supervision.

Abstract:
Insider attackers increasingly leverage LLM- and RAG-driven analytics to compromise facial privacy through external knowledge linking, posing a growing threat to facial recognition systems. Current Privacy-Preserving Face Recognition (PPFR) methods remain vulnerable as they still expose raw facial data to insiders during matching, or allow for the reconstruction of original identities. To prevent such insider betrayals, we propose Plug-and-Play Identity (PPID), a generative PPFR framework that explores a \emphlearnable new identity as a privacy-preserving plug-in for off-the-shelf face recognizers. PPID operates in three stages: (i) identity embeddings are extracted from a frozen face recognizer; (ii) a lightweight identity perturbator learns structured pseudo-identities under geometric constraints; and (iii) a diffusion generator synthesizes anonymized face images conditioned on the learned identities. This formulation suppresses visual identity while preserving feature separability in a protected domain, enabling direct recognition on anonymized images without retraining face recognizer. To stabilize generation and reduce real--synthetic domain shift, PPID employs multi-objective optimization to constrain embedding geometry and a projection module to align conditioning spaces. Empirical evaluation, theoretical analysis, and a real-world UAV case study consistently indicate that PPID achieves a favorable trade-off among privacy, efficiency, and recognition accuracy.

Abstract:
Multimodal large language models (MLLMs) have achieved strong performance on diverse multimodal tasks, yet current post-training methods remain largely answer-centric and do not directly optimize behavioral reliability. In practice, a trustworthy MLLM should remain consistent when the same evidence is presented in different forms, become cautious when critical evidence is missing, and explicitly recognize contradictions when different evidence sources conflict. To address this gap, we propose RAP (Relation-Aware Post-training), a simple and unified framework that post-trains MLLMs via behavior alignment over transformed evidence views. RAP organizes training data into relation pairs that encode four fundamental evidence relations: equivalence, irrelevant perturbation, causal deletion, and evidence conflict. The model is trained to produce a structured output consisting of an answer, supporting evidence, and a decision mode, and is then optimized with a single relation-aware objective that enforces the desired behavior across paired views. RAP is model-agnostic, lightweight, and reframes multimodal post-training from answer optimization to relation-aware behavior alignment, providing a practical direction for building MLLMs that are more grounded and more reliable under evidence variation, incompleteness, and conflict.

Abstract:
Person re-identification (ReID) remains challenging due to significant domain shifts across camera views. Recent CLIP-based methods introduce textual prompts to guide representation learning, yet most approaches rely on a single learnable token, which tends to focus on dominant semantics while overlooking fine-grained identity cues. Moreover, such compact prompt representations are prone to entangling identity information with domain-specific variations, leading to degraded generalization. In this paper, we propose a novel Memory-Guided Multi-Token Semantic Decomposition framework for text-guided ReID. Instead of compressing identity semantics into a single token, we introduce multiple orthogonal textual tokens to explicitly decompose pedestrian identity into diverse and complementary semantic components. These tokens interact with visual patch features via spatially constrained cross-attention, capturing distinct local discriminative evidence. Furthermore, to prevent these tokens from absorbing transient domain biases during cross-modal alignment, we propose a Token-wise Memory-Guided Relational Distribution Distillation strategy. We construct a token-aware memory bank that maintains dynamically updated feature centers for each identity. Based on this memory, we derive stable relation distributions across identities that serve as robust relational priors. By aligning the relation distributions of current token features with these temporally smoothed memory priors via KL divergence, the model effectively filters out domain-induced noise. Extensive experiments demonstrate that our method significantly improves cross-domain generalization.

Abstract:
Recent advances in foundation segmentation models have improved generalization, yet their direct transfer to fine-grained tasks remains limited by imprecise boundaries, incomplete regions, and weak structural adaptability. These limitations are particularly pronounced in structure-sensitive scenarios, where precise structural awareness is essential. In this work, we propose SAM3-Adapter, a parameter-efficient framework that hierarchically learns task-adaptive prompts to guide Segment Anything Model 3 (SAM3) for specialized fine-grained segmentation. Instead of fine-tuning the entire backbone, we keep the SAM3 encoder frozen and introduce a lightweight adaptation module that progressively learns and injects task-adaptive prompts across multiple representation levels. Specifically, our approach leverages this module to learn task-adaptive prompts by jointly modeling two complementary cues: an explicit structural cue that captures low-level, structure-sensitive patterns such as boundaries and shapes, and an implicit SAM3 embedding cue derived from intermediate SAM3 representations. These learned prompts are then hierarchically injected into SAM3 to progressively guide its representations across layers, enabling precise and task-specialized dense predictions with minimal trainable parameters. Extensive experiments on diverse fine-grained segmentation benchmarks demonstrate that SAM3-Adapter consistently outperforms existing SAM- and SAM2-based adaptation methods, achieving new state-of-the-art results.

Abstract:
Large audio-language models (LALMs) have advanced music understanding and driven progress in music information retrieval (MIR), which requires integrating low-level acoustic perception and high-level semantic interpretation. These two levels impose markedly different reasoning demands: over-deliberation on perceptual tasks can induce hallucinations, whereas under-reasoning on semantic tasks leads to shallow predictions. Adaptive reasoning is therefore crucial for MIR, yet remains largely unexplored. Pervasive annotation noise in standard benchmarks further compromises training and evaluation. Moreover, existing approaches typically treat the allocation of reasoning effort as a black-box optimization problem, providing little insight into how models assess difficulty internally. To address these limitations, we propose DART, a difficulty-aware adaptive thinking framework for music tagging that further shows promising results when transferred zero-shot to other MIR tasks. First, we construct CLEAR-MusicTags, an LLM-assisted consensus-based re-annotation of existing benchmarks that covers both training and evaluation splits, along with an expert-annotated test subset (ExpertEval-MusicTags) for gold-standard evaluation. Second, we introduce a Rubric-guided Adaptive Thinking mechanism via Group Relative Policy Optimization (GRPO). Leveraging a teacher-aligned rubric, our approach trains the model to dynamically calibrate whether to invoke reasoning. Third, we further identify internal difficulty directions within model activations and show that steering along these vectors at inference time enables fine-grained control over reasoning behavior without retraining. Extensive experiments show that DART achieves state-of-the-art performance on refined benchmarks while providing an interpretable approach for complex music understanding.

Abstract:
Addressing the User-Oriented Fairness (UOF) issue is a critical challenge in recommender systems (e.g., short-video platforms), as models tend to generate lower-fidelity embeddings for users with sparse signals. Many existing approaches simplify this issue using a binary advantaged/disadvantaged dichotomy. In terms of evaluation, such group-level metrics obscure intra-group unfairness. Methodologically, overlooking the continuous spectrum of embedding degradation homogenizes disadvantaged users, which discards the residual signals essential for global interest modeling. Consequently, existing knowledge transfer methods forcibly align their degraded embeddings with limited advantaged prototypes, ultimately causing interest distribution collapse. To address these problems, we first introduce a fine-grained fairness evaluation metric based on Optimal Transport to capture the distributional unfairness across the entire population. Furthermore, we propose FairDiff, a novel framework that employs Curriculum-Scheduled Ambient Diffusion to adaptively restore degraded embeddings by modeling high-fidelity interest distribution. Instead of indiscriminately discarding information from degraded embeddings, we incorporate the entire spectrum of user data into training by explicitly modeling their varying degradation levels. Consequently, this enriches the learned interest distribution without sacrificing its high fidelity, ultimately enhancing both overall utility and fairness. Extensive evaluations in both ID-based and multimodal scenarios, coupled with online A/B testing on a real-world short-video platform, demonstrate the effectiveness of our method.

Abstract:
Multimodal Large Language Models (MLLMs) have shown strong capabilities in complex reasoning, but their effectiveness often depends on Reinforcement Learning (RL) to reduce hallucinations caused by inherent vision-language semantic gaps. However, existing RL paradigms, including those that use process-supervised reward models, rely on retrospective supervision and do not actively intervene during the reasoning process. This lack of intermediate oversight makes models highly susceptible to error propagation, as early visual misperceptions often lead to logically consistent but factually incorrect outcomes. To address this limitation, we propose the Guided Verifier framework, a closed-loop, dual-agent reinforcement learning approach that applies real-time directional constraints to dynamically correct reasoning trajectories. To support this framework, we introduce CoRe (Correct and Refine), a specialized dataset of 3,000 annotated reasoning trajectories designed for training guided verifiers. Our architecture relies on a lightweight verifier trained on CoRe to detect visual-semantic inconsistencies and provide step-wise textual guidance during the rollout phase. The policy model is then optimized via Guided-GRPO using a composite reward function that balances final accuracy, format compliance, and hallucination penalties to prevent error propagation while preserving model autonomy. Experiments on benchmarks including MathVista, MathVerse, and MMMU demonstrate the effectiveness of our dynamic verification mechanism.

Abstract:
Zero-shot composed image retrieval (ZS-CIR) aims to retrieve a target image from a large gallery that satisfies a user-specified modification intent, given only a reference image and a textual instruction, without any task-specific training. Existing methods suffer from a language-dominant bias: by converting the reference image into textual surrogates such as pseudo-tokens or generated captions, they often discard fine-grained visual details and fail to preserve consistency in unmodified regions, especially under local edits. To address this limitation, we propose VCRD (Visual-Centric Region Decoupling), a visual-centric ZS-CIR framework that explicitly decouples editable and preserved regions at the feature level. The editing intent is first parsed by a multimodal model and grounded to produce a spatial binary mask. With all encoders frozen, lightweight test-time optimization is applied exclusively to the editable visual features, while the preserved features remain unchanged. In parallel, a text-guided global visual query is optimized to balance target semantics with the reference image's global visual representation. To support fine-grained retrieval, we further devise an adaptive multi-path scoring mechanism that jointly integrates preserved-content consistency, editable-region similarity, and global visual alignment, with final matching performed entirely in the visual embedding space. Extensive experiments on three prominent ZS-CIR benchmarks show that VCRD achieves strong performance among zero-shot methods, with particularly notable gains on local attribute modifications and complex compositional scenarios while improving visual consistency in preserved regions.

Abstract:
Recent years have witnessed the success of diffusion models in image customization tasks. However, existing mask-guided human erasing methods still struggle in complex scenarios such as \emphhuman--human occlusion, \emphhuman--object entanglement, and \emphhuman-background interference, mainly due to the lack of large-scale multi-instance datasets and effective spatial decoupling to separate foreground from background. To bridge these gaps, we curate the \emphMILD dataset capturing diverse poses, occlusions, and complex multi-instance interactions. We then define the \emphCross-Domain Attention Gap (CAG), an attention-gap metric to quantify semantic leakage. On top of these, we propose \emphMulti-Layer Diffusion (MILD), which decomposes the generation process into independent denoising pathways, enabling separate reconstruction of each foreground instance and the background. To enhance human-centric understanding, we introduce \emphHuman Morphology Guidance, a plug-and-play module that incorporates pose, parsing, and spatial relationships into the diffusion process to improve structural awareness and restoration quality. Additionally, we present \emphSpatially-Modulated Attention, an adaptive mechanism that leverages spatial mask priors to modulate attention across semantic regions, further widening the CAG to effectively minimize boundary artifacts and mitigate semantic leakage. Experiments show that MILD significantly outperforms existing methods.

Abstract:
Generative linguistic steganography (GLS) enables covert communication by embedding secret messages into the natural language generation process. In practical deployment, however, GLS is vulnerable to tokenization ambiguity: the same surface text may be re-tokenized into a different token sequence at the receiver, breaking the shared decoding state between the communicating parties so that a single local mismatch can propagate into complete extraction failure. Existing solutions either remove ambiguous tokens---distorting the generation distribution and compromising security---or preserve the distribution at the cost of substantially reduced embedding capacity or prohibitive runtime overhead. To address this issue, we propose ReTokSync (Re-Tokenization Synchronization), a self-synchronizing disambiguation framework that monitors the receiver-view tokenization during generation and triggers a corrective reset only when ambiguity actually occurs. By confining the effect of tokenization ambiguity to sparse residual bit errors rather than global desynchronization, ReTokSync leaves ambiguity-free positions entirely untouched and remains compatible with the underlying steganographic algorithm. Experiments on both English and Chinese settings show that ReTokSync preserves the observable visible-text distribution with zero KL divergence and remains close to the steganographic baseline in text quality, embedding capacity, and runtime, while achieving extraction accuracy of at least 99.7%. Building on this property, we further develop a two-channel covert communication mechanism in which ReTokSync serves as the primary channel and a reliable auxiliary channel corrects the remaining errors, achieving 100% end-to-end recovery with limited effective-capacity loss.

Abstract:
LiDAR Point Cloud Compression (LPCC) is critical in autonomous driving, where bandwidth and energy constraints limit practical deployment. The Just Recognizable Distortion (JRD) paradigm improves rateaccuracy trade-offs by preserving task-relevant information, but existing methods rely on energy-intensive artificial neural networks, limiting edge deployment. We propose a new spiking neural network framework for JRD-guided LPCC, namely SJRD-LPCC. Our method converts static point clouds into a coarse-to-fine spike stream via a new progressive point sampling, and performs efficient feature fusion using a new temporal bidirectional interaction module with Softmax-free spiking cross-attention. A specially-designed mixed time-step gating module further aggregates multi-step predictions for accurate quantization labeling. Experimental results show that SJRD-LPCC achieves competitive ratedistortion performance with state-of-the-art LPCC methods while reducing energy consumption by over 80%.

Abstract:
Embodied Foundation Models, encompassing Vision-Language-Action (VLA) models and world models show strong cross-modal reasoning, but both still face a core vision--action gap: mapping high-dimensional visual inputs to low-level action controls remains brittle and limits generalization. Existing fixes, including end-to-end fine-tuning and feature injection, often disrupt spatiotemporal consistency in dynamic scenes. To address this, we introduce World2Act, a world-to-action framework that bridges this vision--action gap by extracting spatial--temporal clues from latent states of a pretrained video world model. During action generation, World2Act performs Adaptive Dynamics Injection by combining these action-aware dynamics clues with proprioceptive state information, enabling more accurate modeling of physical dynamics while preserving predictive integrity for precise control; this leads to consistent gains on both simulated and real-world embodied tasks, especially in long-horizon and perturbation settings. In addition, World2Act employ world-model priors with lightweight adaptation, reducing training overhead while maintaining strong generalization under distribution shifts. Extensive evaluations on LIBERO, LIBERO-Plus, and VLABench show that World2Act achieves success rates of 97.4%, 86.6%, and 51.5%, respectively, surpassing prior SOTA methods by 2.2%, 6.8%, and 6.4%, while also improving parameter efficiency and data scalability.

Abstract:
Understanding dynamic multimedia content requires models to continually learn new visual concepts while preserving previously acquired scene knowledge. In continual panoptic segmentation (CPS), this challenge goes beyond category recognition, since robust performance also depends on maintaining coherent semantic relations, spatial layouts, and instance-level distinctions over time. Existing CPS methods mainly rely on evolving visual features as the reference for knowledge preservation, which becomes increasingly fragile under long incremental learning. We propose Language-Anchored Hierarchical Relational Distillation (LA-HRD), a framework for continual multimodal scene understanding. LA-HRD stabilizes scene representations at three levels: frozen language embeddings provide a semantic anchor for preserving global class topology; visual distillation maintains spatial-contextual structure; and prototype-guided contrastive alignment improves instance-level compactness and temporal consistency. By aligning different forms of knowledge in their appropriate reference spaces, LA-HRD enables more robust long-term scene parsing. Experiments on multiple ADE20K continual protocols and additional COCO evaluation show that LA-HRD consistently improves the stability-plasticity trade-off and produces more coherent scene parsing results over long incremental sequences.

Abstract:
Existing text-to-image person re-identification (TIReID) methods remain susceptible to corrupted supervision, facing both human-annotated noisy correspondence and MLLM-generated caption hallucinations. In this paper, we identify a unified geometric nature of human-annotated noisy correspondence and MLLM-generated caption hallucinations. To this end, we pioneer the Noise-as-Hallucination paradigm, shifting the robust learning philosophy from a ``data-inefficient detect-and-suppress'' practice to a principled ``detect-and-cure'' strategy. To operationalize this, we propose GCD-ReID, a Geodesic-Aware Contrastive Dehallucination framework grounded in Riemannian geometry to overcome the distortion of Euclidean metrics on the Contrastive Language-Image Pretraining (CLIP) hypersphere. Specifically, we project visual features into local Tangent Spaces via the Logarithmic Map to estimate Identity-Specific Dimensional Uncertainty. This geometric prior rigorously decouples benign view-induced variance from genuine semantic mismatches. Guided by this disentanglement, GCD-ReID integrates Instance-Aware Noise Detection (IND) to reliably isolate correspondence errors and Dimension-Aware Contrastive Dehallucination (DCD) to perform surgical rectification on spurious semantic conflicts. Extensive experiments on multiple standard benchmarks demonstrate that GCD-ReID exhibits superior robustness across the entire noise spectrum, maintaining high accuracy on clean datasets while outperforming competitors in high-noise regimes.

Authors: Qinwen Xu (State Key Laboratory of Multimedia Information Processing, School of Computer Science, Peking University), Jiaming Liu (State Key Laboratory of Multimedia Information Processing, School of Computer Science, Peking University), Rui Zhou (State Key Laboratory of Multimedia Information Processing, School of Computer Science, Peking University), Shaojun Shi (Peking University), Nuowei Han (Peking University), Zhuoyang Liu (Peking University), Chenyang Gu (Peking University), Shuo Gu (Simplexity Robotics), Yang Yue (Tsinghua University), Gao Huang (Tsinghua University), Wenzhao Zheng (Tsinghua University), Sirui Han (The Hong Kong University of Science and Technology), Peng Jia (Simplexity Robotics), Shanghang Zhang (Peking University)

Abstract:
Despite strong generalization capabilities, Vision-Language-Action (VLA) models remain constrained by the high cost of expert demonstrations and limited real-world interaction. While online reinforcement learning (RL) has shown promise, its application to real-world VLA manipulation is hindered by low exploration efficiency and restricted exploration coverage. Through systematic real-world experiments, we observe that the effective exploration space of online RL is largely constrained by the trajectory distribution induced during supervised fine-tuning (SFT). Motivated by this observation, we propose TwinRL, a digital twinreal-world collaborative post-training framework that expands and guides RL exploration for VLA models through three stages: SFT warm-up, twin RL warm-up, and real-world RL. TwinRL first reconstructs a high-fidelity digital twin from smartphone-captured scenes. During the SFT stage, we introduce an exploration space expansion strategy that expands the support of the trajectory distribution beyond real demonstrations, reshaping the exploration space for more effective RL. Rather than treating the twin as a data augmentation tool, we propose a twin RL warm-up strategy that enables it to act as an exploration guide for real-world RL. Specifically, TwinRL performs efficient parallel RL in the digital twin to generate interactive trajectories that populate the replay buffer and stabilize subsequent real-world RL learning. This process also identifies failure-prone yet informative configurations, enabling targeted human-in-the-loop rollouts to further improve on-robot efficiency. Across four tasks, TwinRL achieves near-100% success in both in-distribution and out-of-distribution regions, delivering over 30% faster convergence than prior real-world RL methods with only 20 minutes of on-robot interaction.

Abstract:
The rapid proliferation of Large Language Models (LLMs) and advanced diffusion models presents a major challenge for LLM-generated multimodal posts on social media. These models enable malicious actors to generate highly realistic text-image posts with flawless surface alignment, rendering conventional detection methods that rely on unimodal artifacts or cross-modal statistical correlations increasingly ineffective. In this work, we shift the focus from observational feature matching to the underlying structural causal models governing multimodal creation. We introduce Causal Generative Asymmetry: the inherent structural divergence where genuine human images and texts complement each other, driven by an unobserved latent intent. In contrast, LLM-generated modalities lack this latent reality and are deterministically bound to an explicit prompt, causing severe causal redundancy despite their flawless surface alignment. Motivated by these findings, we propose a novel framework named \Name to unmask LLM-generated multimodal social posts. It leverages a Vision-Language Model (VLM) to extract deeply contextualized feature matrices and introduces a subspace projection mechanism. It explicitly isolates the orthogonal human intent from visually redundant prompt translations, utilizing Von Neumann Spectral Entropy as a key structural signal to classify human and LLM-generated content. Experiments across multiple datasets show our approach outperforms SOTA methods by up to 7.74% in accuracy, validating Causal Generative Asymmetry as an effective marker for detecting synthetic multimodal content.

Abstract:
Recent Vision-Language-Action (VLA) models have shown impressive flexibility and generalization, yet their practical deployment remains limited by substantial computational overhead and inference latency. In this work, we present ActDistill, a general action-guided self-derived distillation framework that transfers the action prediction capability of a full-scale VLA model to a lightweight counterpart. Unlike previous efficiency approaches that primarily focus on vision-language correlations, ActDistill explicitly leverages action priors to guide knowledge transfer and model compression, achieving action-oriented efficiency for VLA models. Specifically, we employ a well-trained VLA model as the teacher and introduce a graph-structured encapsulation to model the hierarchical dependencies in action prediction. The student model, derived from the graph-encapsulated teacher, is further equipped with a dynamic router that adaptively selects computation paths conditioned on action requirements. The routing process is guided by hierarchical graph-informed supervision, enabling efficient and stable action prediction. During inference, all auxiliary components are removed, allowing the student to execute only dynamically routed layers and generate high-precision actions. Experiments on embodied benchmarks demonstrate that ActDistill achieves comparable or superior performance to full-scale VLA models while reducing computation by over 50% with up to 1.67 speedup, establishing a general paradigm for efficient embodied intelligence.

Abstract:
Vision-Language Navigation in Continuous Environments (VLN-CE) requires agents to follow natural-language instructions while navigating in real-world-like environments. Most VLN-CE approach\-es adopt a three-stage framework: a waypoint predictor proposes navigable waypoints, and a navigator selects the best waypoint, with a low-level controller executing the movement to it. However, this decoupled paradigm often leads to unreachable waypoints or inconsistencies between planning and control. In this work, instead of predicting isolated waypoints, we introduce a novel paradigm called Trajectory Waypoint, which grounds each candidate waypoint in an executable trajectory. To realize this, we design a Trajectory Waypoint Predictor formulated as a TSDF-guided diffusion policy, which steers trajectory generation away from obstacles, inherently ensuring the reachability of the predicted waypoints. We further propose a trajectory-enhanced navigator that injects the associated trajectory as additional information for planning, enabling strict consistency between high-level semantic decisions and low-level execution. Extensive experiments on the VLN-CE benchmark show that our Trajectory Waypoint paradigm achieves superior performance over the baselines.

Abstract:
Fundus images often suffer from complex and coupled degradations, which hinder both clinical diagnosis and automated retinal analysis. Existing fundus image enhancement methods formulate the problem as degradation-agnostic blind image restoration, forcing a single model to handle diverse degradation patterns and often leading to an averaging effect. While degradation-aware conditioning has shown promise in general all-in-one image restoration, it remains underexplored in fundus image enhancement, where degradations are often mixed, coupled, and continuously varying. More importantly, existing prior designs are inadequate for this setting: explicit priors are often too coarse or discrete to characterize degradation mixtures, whereas implicit priors are flexible but weakly grounded and semantically unclear. We argue that the key is to endow enhancement with a structured degradation representation. To this end, we propose Grounded Degradation Prompting (GroDePrompt), which models realistic fundus degradations as compositions of grounded latent degradation bases. Specifically, GroDePrompt constructs a latent degradation dictionary, predicts image-specific composition coefficients, and synthesizes a degradation-aware prompt to guide adaptive feature modulation in the enhancement network. This design yields a more structured and semantically grounded representation for mixed degradations than degradation-agnostic mappings or free-form prompts. Experiments across multiple datasets and degradation conditions validate the effectiveness of our method. Notably, on EyeQ, it outperforms the state-of-the-art method FD3 by 4.84 dB in PSNR.

Abstract:
Multi-view anomaly detection aims to exploit complementary observations from different viewpoints for improved normality modeling, while treating deviations as anomalies. Existing methods typically perform cross-view interaction at the patch level, which tends to align view-dependent local details rather than view-invariant normal patterns, and may even propagate abnormal cues, thereby weakening anomaly separability. To address this issue, we propose IACA, a multi-view anomaly detection framework based on global prototype aggregation and alignment. Specifically, IACA distills each view into a global prototype token through Intra-view Prototype Aggregation (IPA), forming a strict information bottleneck that captures its core normal representation, and aligns the resulting prototypes via Cross-view Prototype Alignment (CPA) to encourage consistent normal patterns across views. To enforce these aligned prototypes as normality constraints, we introduce Prototype Bottleneck Reconstruction (PBR), which reconstructs features under a prototype-induced low-rank bottleneck. In addition, we propose Multi-view Identity Recentering (MIR) to construct identity-consistent reconstruction targets, and a Hybrid Hard-mining Cosine Loss to emphasize hard regions during optimization. Extensive experiments on three multi-view anomaly detection benchmarks show that IACA establishes state-of-the-art performance with high efficiency. In particular, the lightweight variant IACA-S, with only 41.0M parameters and 41.6G FLOPs, achieves 95.7% sample-level and 92.6% image-level AUROC on the challenging Real-IAD Variety dataset, improving by +4.1% and +4.7%, respectively.

Abstract:
AI-generated video (AIGV) detection aims to distinguish real videos from AI-generated ones. In practice, detectors trained on existing data often fail to generalize to newly emerging generative models, making this task challenging. Therefore, continual learning (CL) is essential for improving the adaptability. However, CL frameworks for this task remain underexplored. To this end, we propose SphereVideo, a novel CL framework for AIGV detection built on two key observations. First, real videos exhibit a compact feature distribution. Based on this, we encourage real video features to cluster around a real prototype on a hypersphere while repelling AI-generated samples, thereby establishing a decision boundary. This prototype serves as a stable anchor for CL, regulating boundary evolution and mitigating catastrophic forgetting. Second, existing methods tend to rely solely on spatial artifacts as shortcuts. To enhance temporal modeling, we introduce a strategy that models the temporal dynamics of real data at both frame and clip levels. By strengthening real data modeling, this strategy further facilitates learning a real prototype and forming a stable decision boundary. Moreover, we construct a comprehensive and challenging benchmark. Extensive experiments demonstrate that SphereVideo achieves an improved plasticity-stability trade-off, outperforming prior methods by 3.08% on seen data and 4.00% on unseen AI-generated data.

Abstract:
Source-free domain adaptation (SFDA) provides a practical solution to cross-subject EEG decoding by adapting source-pretrained models to unlabeled target domains without accessing source data. However, existing SFDA methods rely solely on the limited internal knowledge of source-pretrained models, leading to inferior cross-domain generalization and unreliable pseudo-labels. Although EEG Foundation Models (FMs) pretrained on large-scale data exhibit strong generalizability, their potential in SFDA remains largely unexplored. To this end, we propose FUSED, a \underlineFoundation-g\underlineUided \underlineSource-free \underlineEEG \underlineDecoding framework that integrates a large-scale FM with a compact Specialist Model (SM) via dual-branch co-adaptation. Specifically, we introduce a Co-adaptation mechanism equipping both branches with linear and prototype views, enabling cross-branch pseudo-label generation. Additionally, we design a Consensus Filtering Mechanism that exploits the FM's inherent stability to identify high-quality samples, along with a Two-Stage Pseudo-Label Refinement scheme to suppress error accumulation through cross-branch arbitration. Finally, we calibrate the FM's decision boundaries via mutual information maximization with the SM, followed by knowledge distillation from FM to SM, forming a principled calibrate-then-distill pipeline. To our knowledge, FUSED is the first work to leverage EEG FMs within the SFDA framework for cross-subject EEG decoding. Extensive experiments across three EEG paradigms, including motor imagery, emotion recognition, and SSVEP, demonstrate consistent state-of-the-art performance, validating the effectiveness of foundation-guided synergy for robust and privacy-preserving EEG decoding.

Abstract:
Diffusion models have strong generative capabilities. However, their maximum likelihood training objective only focuses on reconstructing the data distribution, making it difficult to align with specific preferences. Reinforcement learning (RL) for preference alignment in diffusion models is promising but limited by reward sparsity. Since a single reward cannot support optimization, existing RL methods usually backpropagate the final reward to all previous steps. However, denoising is stage-wise, with distinct semantics and controllability. Repeating the final reward across all steps creates a temporal objective mismatch, encouraging reward shortcuts that lead to reward hacking. At the same time, due to reward backfilling, each time step receives the same reward, making it impossible to distinguish between actions, thereby weakening the optimization process. To resolve this issue, we propose Stage-Guided Per-step Optimization (SGPO) for diffusion models, which jointly leverages signal-to-noise ratio and semantic changes to identify generation stages and adaptively assign stage-specific objectives. Early denoising is chaotic and far from the final reward, resulting in weak rewardbehavior correlation. This stage should prioritize exiting the chaotic state. In the mid stage, the latent transitions to a stable structure, where the final reward better corresponds to generative behavior. Therefore, this stage optimizes the final reward while exploring diversity to avoid early convergence to a single mode. In the late stage, the latents core structure is largely fixed, and preference optimization mainly amplifies local details, risking overfitting. Therefore, stable convergence is preferred to avoid quality degradation. Results from 16 comparative experiments validate SGPO. Our method achieves 26.7% average gains in generative quality and 36.7% higher convergence speed.

Abstract:
Multimodal models often suffer severe performance degradation due to missing modalities during real-world inference. Although diffusion models are widely adopted for missing modality completion, they exhibit a counter-intuitive paradox: their high-fidelity reconstruction barely improves downstream accuracy, yielding a <1.3% gain over pure Gaussian noise under extreme missingness. We define this as the Terminal-State Supervision Gap: constrained by intermediate denoising optimization, existing models lack direct task-level constraints on final features, while the inherent variance of multi-step sampling triggers downstream error amplification. To bridge this gap, we propose the \underlineTrajectory-\underlineCoupled Multimodal \underlineFlow (TC-Flow) framework. Specifically, the core Prediction-Coupled Flow Dynamics (PCFD) module combines Conditional Flow Matching and x-prediction to directly predict terminal features, while employing an Ordinary Differential Equation (ODE) solver to backpropagate downstream task losses end-to-end, explicitly shaping generation dynamics with discriminative objectives. Furthermore, the Trajectory-Calibrated Fusion (TCF) module extracts the dynamic stability of the ODE trajectory as a reliability prior to adaptively calibrate fusion gating, proactively blocking error cascades from unreliable completions. Extensive experiments on CMU-MOSI, CMU-MOSEI, and Weibo-17 demonstrate that TC-Flow consistently outperforms existing methods across various missing protocols, achieving up to an 8.0% absolute performance boost on CMU-MOSEI under the fixed-missing protocol. Code and trained models will be released upon publication.

Abstract:
We present SelfieAvatar, a practical system for generating high-fidelity, animatable 3D head avatars from three casually captured, in-the-wild selfies, without requiring camera calibration, fixed facial expressions, or 3D morphable model (3DMM) fitting. While multiple selfies provide stronger geometric cues than single-image input, they also introduce significant challenges due to unknown camera poses and inconsistent facial expressions across views. To address these issues, we employ a Diffusion Transformer (DiT)-based multi-view rectifier that canonicalizes the input selfies to predefined viewpoints while normalizing them to a shared facial expression. Based on these rectified canonical views, we use a feed-forward Gaussian splatting reconstruction model that produces detailed and geometrically consistent 3D head avatars. Furthermore, we propose a 3DMM-free 3D avatar reenactment framework based on latent diffusion, in which a compact motion embedding extracted from a driving image controls facial expression via cross-attention, and multi-view consistency modules enforce coherent animation across viewpoints. This design enables consistent multi-view facial animation with fine-grained expression preservation, including challenging motions such as tongue movement. Extensive experiments demonstrate that SelfieAvatar achieves superior reconstruction quality, view consistency, and animation realism compared to existing methods, highlighting its robustness and practicality for real-world 3D avatar creation from sparse casual inputs.

Abstract:
Multi-modal object re-identification (ReID) aims to construct complementary identity features from heterogeneous modalities. Existing methods typically rely on local fine-grained perceptive modeling to fuse cross-modal information. However, such Euclidean-based methods struggle to capture intrinsic hierarchical relationships and remain susceptible to low-quality noise. To address these issues, we propose Hyper-ReID, a granularity-aware hyperbolic graph learning framework that decouples the fusion process via hierarchical modeling. Specifically, the Multi-Grained Modality Sampling (MGMS) module introduces sampling anchors to dynamically extract semantic features, constructing robust and diverse representations across different granularities. To hierarchize these features, the Granularity-Aware Hyperbolic Graph (GHG) module projects them into hyperbolic space, organizing the representations into the multi-order graph that explicitly models hierarchical correlations across part-level, intra-modal, and inter-modal semantics. To enforce this hierarchy, the Structured Hierarchical Order Loss regularizes the feature space, anchoring abstract identities near the origin and concrete semantic parts toward the boundary. Finally, the Hierarchical Multi-Hop Fusion (HMF) mechanism propagates contextual information across the constructed graph nodes, effectively suppressing modality noise to yield discriminative identity representations. Extensive experiments on four public benchmarks demonstrate that Hyper-ReID achieves state-of-the-art performance.

Abstract:
In trained object tracking methods, annotating object-tracking datasets is labor-intensive, and existing datasets are dominated by limited scenarios, both of which constrain tracking applications. Recent work has shifted the rich scenes in Large Visual Models (LVM) to the tracking field by using the training-free methods. The time instance segmentation capability of Segment Anything Model 2 (SAM2) has shown significant potential in advancing training-free methods. However, existing training-free methods have not achieved optimal results in handling distractors and rely overly on the segmentation results of the segmentation model itself, leading to certain limitations. To this end, we introduce FreeTrack, a training-free anti-distractor object tracker. Specifically, we first proposed Focused Box Prompts (FBP), which processes the Intersection over Union (IoU) of the mask obtained from the segmentation model of the object to generate reliable object and distractor positions, and then uses box prompts to focus on the object. Then Auxiliary Feature Extraction (AFE) performs semantic image feature extraction on the object area to be tracked, which facilitates auxiliary semantic knowledge and more accurate selection of object masks. Lastly, the Knowledge Gallery (KG) stores semantic knowledge and uses this knowledge base to balance template drift and template calibration for tracking objects. These three modules have been seamlessly optimized to achieve optimal results without increasing training costs. Numerous experiments have shown that compared to the SAM2 model, the Area Under Curve (AUC) value in the LaSOT dataset has increased by 5.8% and in the LaSOT\textsubscriptext dataset by 6.0%. It achieved competitive results with fully trained methods on multiple other datasets.

Abstract:
Spiking Neural Networks (SNNs) are promising for energy-efficient perception, yet their temporal advantage remains underexploited in high-precision object detection. In practice, existing directly trained SNN detectors often operate on replicated static frames or long-window integration, which suppresses temporal diversity, drives neurons toward homogeneous firing patterns, and limits the use of spike timing for fine-grained regression. This results in a fundamental mismatch: although SNNs are intrinsically temporal, their deployed dynamics for precise detection are largely absent. We present Hierarchical Temporal Residual Dynamics Network, a unified framework that reactivates temporal modeling at the neuron, feature, and decoding levels. Specifically, we introduce Temporal Residual Neurons to induce state-dependent firing dynamics under static inputs, Intra-/Inter-block Temporal Residual Connections to propagate temporal richness throughout the network, and a Dual-Path Spike Decoder to exploit fine-grained spike interval information for sub-rate bounding-box regression without extending the simulation horizon. Extensive experiments on three public benchmarks show that Hierarchical Temporal Residual Dynamics Network consistently surpasses strong baselines on both static and neuromorphic datasets with minimal overhead. Our findings suggest that hierarchical temporal residual modeling is key to unlocking the full potential of SNNs for high-precision perception.

Abstract:
Heterogeneous video transitions are essential in professional video production, yet automatically generating structurally coherent transitions between heterogeneous scenes remains an open problem. Existing approacheswhether traditional template-based effects, adapted video inpainting, or multi-shot generation methodstreat content synthesis and transition blending as isolated sub-problems, inevitably producing ghosting artifacts or boundary-unaware compositions. In this paper, we formally define the task of heterogeneous video transition and present ZeroCut, the first end-to-end framework purpose-built for this task. Our approach comprises three key components. First, a scalable data pipeline (TransPipe) automatically constructs large-scale transition pairs from ordinary motion videos, overcoming the absence of dedicated training data. Second, we extend a pre-trained video diffusion transformer with a temporal concatenation formulation that places clean scene tokens and noisy transition tokens into a single sequence, enabling the model to generate directed cross-domain transformations while being directly conditioned on both boundary scenes within each self-attention layer. A Transition Progress Encoding further guides each frame along the cross-domain appearance axis via adaptive layer normalization. Third, a Dynamic Context Training strategy decouples the transition operation from any fixed temporal position, yielding a shift-invariant operator that, combined with a Decoupled Temporal Mask Attention, naturally generalizes to zero-shot multi-scene transitions at inference. Extensive experiments demonstrate that ZeroCut significantly outperforms adapted baselines in structural coherence, transition clarity, and visual quality.

Abstract:
Online video inpainting restores masked regions in streaming videos while preserving spatial plausibility and temporal coherence over long horizons. Unlike offline restoration, online inference proceeds sequentially without access to future frames, and previously generated outputs cannot be revised. The central challenge is therefore not only current-frame completion, but how accumulated history is reused once generated content becomes a future conditioning signal. Long-range support is often necessary for high-fidelity restoration, yet indiscriminate reuse of historical generated tokens can amplify structural drift, while overly restrictive reuse may reduce restoration accuracy and efficiency. We propose Risk-Calibrated Token Routing (RCTR), a framework for long-horizon online video inpainting under bounded memory. RCTR decomposes historical tokens into short-term continuity, evidence, and hypothesis branches, performs hole-centric token routing to retrieve role-compatible supports, resolves their corrective authority through correction-authority calibration, and updates history conservatively to prevent generated tokens from becoming unrestricted future support. Extensive experiments under strict online protocols, including native online-memory baselines, matched same-backbone attribution, calibration diagnostics, propagation analysis, and robustness tests, demonstrate that RCTR improves temporal consistency and long-horizon stability while maintaining bounded online memory.

Abstract:
Accurate object-level 3D scene representation is the foundation of embodied intelligence and immersive multimedia interactions. However, existing semantic 3D Gaussian Splatting methods typically treat semantics as passive attributes of geometry. Such weakly coupled methods are highly prone to semantic aliasing at the boundaries. To address this, we propose SynGS, a novel semantic 3DGS framework based on semantic-geometric synergistic optimization. This method leverages semantic information to actively constrain the generation, evolution, and correction of Gaussian primitives. Specifically, SynGS designs a semantic-aware adaptive proliferation strategy that guides Gaussian primitives toward a reasonable spatial distribution throughout the scene through joint macro-micro scheduling. During the evolution of Gaussian primitives, semantic modulation and second-order anchor enhancement actively reshape the distribution of their geometric features. For out-of-bounds primitives, the proposed semantic correction mechanism performs pruning and reinitialization. Experiments demonstrate that SynGS achieves state-of-the-art performance on both open-vocabulary and panoptic segmentation tasks. Our method effectively eliminates boundary bleeding issues and presents sharp physical object boundaries while maintaining high-fidelity rendering.

Abstract:
Sports video analysis is crucial for athletic analytics and broadcasting enhancement. Dense sports video reasoning, however, demands a fine-grained understanding of numerous small-scale, highly interactive, and visually homogeneous entities (e.g., players sharing identical uniforms, the ball) across long temporal contexts. Current Large Multimodal Models (LMMs) inherently struggle with such dense visual complexities. Due to the lack of fine-grained visual details, these models often over-rely on textual priors to guess answers, especially when distinguishing visually similar actions and players. To address this, we propose SportsGrounder, a framework that leverages an open-vocabulary visual expert to aid interleaved grounding specifically for dense sports video reasoning. To achieve precise spatial localization, we extract domain-guided object proposals and introduce an Interleaved Grounding Fusion (IGF) mechanism. The IGF frame-by-frame integrates explicit bounding box coordinates and implicit visual semantics with global grid features. This design preserves strict temporal alignment and prevents sequence length explosion. Furthermore, we design an Action-Aware Supervision (AAS) module that directly regularizes the model's hidden states, forcing the network to learn accurate motion representations rather than relying on language bias. Optimized with Mixed Preference Optimization (MPO) to better distinguish deceptive distractors, our extensive experiments on newly curated dense sports VQA datasets (derived from SoccerNet and FineSports) demonstrate that SportsGrounder significantly improves fine-grained reasoning and achieves state-of-the-art accuracy.

Abstract:
The Diffusion Hypernetwork (DH) integrates diffusion-based optimization with hypernetwork-driven parameter generation, offering a powerful paradigm for modeling complex and structured data. However, in heterogeneous distributed settings such as federated learning, existing DH frameworks face fundamental limitations. First, they rely on a single perturbation mechanism, which fails to jointly accommodate heterogeneous modalities continuous data typically require Gaussian noise, while discrete data depend on auxiliary latent transformations hindering unified diffusion modeling. Second, we show that hypernetwork-generated weights exhibit severe geometric anisotropy in hyperbolic space under heterogeneous federation; ignoring this anisotropy leads to distorted gradients and unstable optimization, a challenge largely overlooked by prior DH methods. To address these issues, we propose Fed-HDDH, a Federated Hyperbolic Directional Diffusion Hypernetwork. Fed-HDDH treats model weights as diffusion objects in hyperbolic space and introduces a radialangular constrained forward diffusion that preserves intrinsic anisotropy while enabling cross-client conditional reverse denoising. Extensive experiments across diverse modalities demonstrate that Fed-HDDH consistently outperforms both Euclidean and hyperbolic baselines by an average of 5.7% and improves hypernetwork adaptability by 5.2%.

Abstract:
3D Gaussian Splatting (3DGS) has revolutionized the field of 3D reconstruction and novel view synthesis, emerging as a promising representation for 3D scenes. Registering multiple 3D Gaussians across different coordinate systems is a crucial step for various downstream tasks, such as large-scale scene reconstruction. However, existing methods have either underutilized the scene information in 3D Gaussians or struggled with degraded 3D Gaussian point clouds with wrong geometry. To address these issues, we introduce SemRegGS, a novel and robust registration approach that incorporates semantic information to guide the registration process so as to bring the 3DGS models into the same coordinate. Specifically, to address the error-prone geometry of the scene in 3D Gaussians, we propose an uncertainty-aware semantic unprojection module that leverages the 2D foundational vision models to assign high-level semantic information to off-the-shelf 3DGS models. Building on this, we introduce a semantic-guided point cloud registration module that is resilient to the less accurate geometry in the 3D Gaussian point clouds. Moreover, to fully exploit the scene information in the 3D Gaussians, we propose an image-guided refinement module that enhances registration precision. Extensive experiments demonstrate that our method not only outperforms existing approaches in standard scenarios but also achieves a high success rate in extreme cases where other methods fail.

Abstract:
Deep learning models in medicine increasingly depend on structured clinical knowledge for complex reasoning and diagnosis. While prior work has mainly leveraged unimodal resources such as the Unified Medical Language System (UMLS), multimodal knowledge graphs remain underexplored due to the lack of datasets linking medical images with structured clinical concepts. To fill this gap, we introduce MedMKG, a medical multimodal knowledge graph that integrates radiological chest X-ray images from MIMIC-CXR with an ontological knowledge graph from UMLS via a multi-stage construction framework. MedMKG employs a hybrid pipeline combining rule-based extraction and large language model-based disambiguation to generate reliable concepts and relations. To improve graph quality, we propose Neighbor-aware Filtering (NaF), an efficient method for reducing redundancy during multimodal knowledge graph construction. We establish MedMKG as a benchmark by evaluating it on 5 medical tasks across 2 experimental settings, using 24 baseline methods and 4 visionlanguage backbones across 6 datasets. Results demonstrate consistent performance gains across tasks, highlighting MedMKGs value for benchmarking and advancing multimodal knowledge integration in radiology.

Abstract:
Today's Internet traffic is dominated by video content, which uses adaptive bitrate (ABR) algorithms and buffers to improve streaming quality. However, many traditional ABR designs mainly consider and evaluate linear viewing, overlooking common interactions such as seeking, replaying, and short skips. In traditional ABRs, these nonlinear interactions lead to buffer waste and increased stalls. To study these effects, we built an interactive streaming platform that supports progress-bar seeking, fast-forward/replay, and keyboard shortcuts for fixed-distance skipping. We further implement a stall-minimizing seek mechanism that trades seek accuracy for reduced post-seek stalls, allowing us to quantify how stall reduction changes user behavior. Using 5448 viewing traces, we observe strongly heavy-tailed viewing dynamics: while most sessions exhibit limited interaction, a small fraction accounts for a disproportionate number of seeks. On average, users seek 5.91 times forward and 0.76 times backward per session, and the stall-minimizing mechanism experiences 10.9% more jumps. We also observe early-abandonment patterns in a subset of traces. Overall, our results show that user interaction is a first-order factor that is often missing from modern streaming pipeline design and evaluation.

Abstract:
With the rapid development of e-commerce and digital fashion, image-based virtual try-on (VTON) has attracted increasing attention. However, existing VTON models often suffer from artifacts such as garment distortion and body inconsistency, highlighting the need for reliable quality evaluation of VTON-generated images. To this end, we construct VTONQA, the first multi-dimensional quality assessment dataset specifically designed for VTON, which contains 8,132 images generated by 11 representative VTON models, along with 24,396 mean opinion scores (MOSs) across three evaluation dimensions (i.e., clothing fit, body compatibility, and overall quality). Based on VTONQA, we benchmark both VTON models and a diverse set of image quality assessment (IQA) metrics, revealing the limitations of existing methods and highlighting the value of the proposed dataset. We believe that the VTONQA dataset and corresponding benchmarks will provide a solid foundation for perceptually aligned evaluation, benefiting both the development of quality assessment methods and the advancement of VTON models. The dataset we proposed in this paper is publicly available at: https://huggingface.co/datasets/weixiny0408/vtonqa.

Abstract:
Autoregressive models have recently shown great promise in visual generation by leveraging discrete token sequences akin to language modeling. However, existing approaches typically rely on fixed raster-scan orderings and token-by-token decoding or the complexity of multi-scale representations, which impose suboptimal factorizations of the image distribution and limit both efficiency and scalability. In this work, we propose Expanding Autoregressive Representation (EAR), a novel autoregressive factorization framework that reformulates image generation as a permutation-dependent decomposition of the joint distribution. Specifically, EAR defines a structured center-outward spiral ordering that better aligns with spatial dependencies, preserving local continuity while progressively expanding the receptive field. Based on this structured factorization, we further introduce a length-adaptive decoding strategy that enables block-wise parallel generation under a conditional independence approximation, significantly reducing the number of generation steps. Extensive experiments on ImageNet demonstrate that EAR achieves state-of-the-art trade-off between fidelity and efficiency among single-scale autoregressive models, and provides a competitive alternative to recent few-step generative paradigms.

Authors: Tianyi Gao (Xi'an Jiaotong University), Han Fang (China Telecom Artificial Intelligence Technology (Beijing) Co., Ltd.), Tianyi Ding (Beijing University of Post and Telecommunications), Hao Li (Shanghai Jiaotong University), Xin Wei (China Telecom Artificial Intelligence Technology (Beijing) Co., Ltd.), Hongbo Sun (China Telecom Artificial Intelligence Technology (Beijing) Co., Ltd.), Xiaodong Dong (China Telecom Artificial Intelligence Technology (Beijing) Co., Ltd.), Ye Yuan (China Telecom Artificial Intelligence Technology (Beijing) Co., Ltd.), Jinglin Xu (University of Science and Technology Beijing), Kongming Liang (Beijing University of Posts and Telecommunications), Hao Sun (China Telecom Artificial Intelligence Technology (Beijing) Co., Ltd.), Jingmin Xin (Xi'an Jiaotong University)

Abstract:
Multimodal Large Language Models have made great progress in grounding tasks, yet existing methods still struggle to unify precise localization and complex reasoning. For one thing, text-based methods rely on coordinates or index prediction, severely limiting the perceptual capabilities of the model for dense visual objects. Meanwhile, latent token-based methods employ special tokens without inherent spatial references and use a decoding mechanism that lacks thinking steps, weakening high-level reasoning capabilities. Consequently, developing a unified framework that excels in both perception and reasoning remains challenging. To address this, we propose Mixture-of-Thought-Tokens (Motto), a new free-form multimodal grounding method that bridges the perception-reasoning gap, enabling MLLMs to empower diverse, arbitrary grounding queries. Specifically, we introduce Spatially-Grounded Thought Tokenization to explicitly align special tokens with spatial locations for clear spatial correspondence and visual interpretability. We further design a Context-Adaptive Chain-of-Tokens that dynamically switch grounding modes within an interleaved reasoning chain, achieving robust grounding across tasks of varying complexity. In addition, we construct PR-Bench, a new referring expression comprehension benchmark to evaluate the perception-reasoning gap. Extensive experiments demonstrate that Motto achieves state-of-the-art performance across diverse free-form grounding tasks.

Abstract:
With the rapid growth of e-commerce, exploring general representations rather than task-specific ones has attracted increasing attention. Although recent multimodal large language models (MLLMs) have driven significant progress in product understanding, they are typically employed as feature extractors that implicitly encode product information into global embeddings, thereby limiting their ability to capture fine-grained attributes. Therefore, we argue that leveraging the reasoning capabilities of MLLMs to explicitly model fine-grained product attributes holds significant potential. Nevertheless, achieving this goal remains non-trivial due to several key challenges: (i) long-context reasoning tends to dilute the model's attention to salient information in the raw input; (ii) supervised fine-tuning (SFT) primarily encourages rigid imitation, limiting the exploration of effective reasoning strategies; and (iii) fine-grained details are progressively attenuated during forward propagation. To address these issues, we propose MOON3.0, the first reasoning-aware MLLM-based model for product representation learning. Our method (1) employs a multi-head modality fusion module to adaptively integrate raw signals; (2) incorporates a joint contrastive and reinforcement learning framework to autonomously explore more effective reasoning strategies; and (3) introduces a fine-grained residual enhancement module to progressively preserve local details throughout the network. Additionally, we release a large-scale multimodal e-commerce benchmark MBE3.0. Experimentally, our model demonstrates state-of-the-art zero-shot performance across various downstream tasks on both our benchmark and public datasets.

Abstract:
Multi-view clustering aims to exploit complementary information across multiple views to improve clustering performance. However, recent tensor-based methods often suffer from high computational costs due to tensor singular value decomposition (t-SVD). The outliers in the original data may lead to the degradation of clustering performance. In this paper, we propose a tensor-based multi-view clustering method named Self-representation Tensor Tri-Factorization with View Consistency (VC-STTF), which enhances both computational efficiency and robustness. Specifically, the self-representation tensor is decomposed into the product of three smaller tensors with orthogonal constraints. This decomposition shifts the time-consuming nuclear norm minimization from the large original tensor to a much smaller core tensor, significantly reducing the computational burden while preserving the ability to capture high-order correlations among views. Moreover, outliers are eliminated by using self-representation learning, which improves the robustness of the model. Extensive experiments on several benchmark datasets demonstrate that the proposed method generally outperforms most state-of-the-art methods.

Abstract:
Cross-view geo-localization (CVGL) aims to estimate geographic locations by matching images captured from different viewpoints, such as drone and satellite views. Existing methods mainly rely on visual representations, but often fail to jointly model fine-grained structural correspondences and semantic priors, making them prone to confusion between visually similar but semantically different regions, and thus limiting robustness under large viewpoint variations. To address these challenges, we propose S^3Geo, a structure-semantic synergistic learning framework for cross-view matching. Specifically, we first introduce a Decoupled Query Pooling (DQP) module to extract a compact set of region-aware features from dense tokens, enabling explicit modeling of local structural patterns. We then design a query-level contrastive learning scheme with an optimal transport (OT)-based formulation to establish soft correspondences under cross-view spatial misalignment. Furthermore, we incorporate a Semantic Knowledge Distillation (SKD) strategy from a frozen CLIP teacher to transfer semantic priors and relational structures, thereby improving discrimination on hard negatives. By operating synergistically, the semantic priors provide robust contextual filtering, which guides the structural module to establish precise spatial alignments. Experiments on the University-1652 and SUES-200 datasets demonstrate that S^3Geo consistently outperforms state-of-the-art approaches without increasing inference complexity, validating the effectiveness of jointly modeling structural and semantic information for CVGL.

Abstract:
Existing watermark decoders struggle under composite attacks, where multiple degradations jointly obscure the embedded signal. We formulate robust watermark decoding under composite attacks as a compositional recovery problem rather than a monolithic one-decoder task. To this end, we propose MoE-Shield, a robust watermark decoder that couples an attack-aware feature extractor with a shared-specialized Mixture-of-Experts (MoE) architecture. The extractor processes the distorted watermarked image through image, DCT, FFT, edge, and CNN branches to expose complementary distortion cues. A distortion-supervised yet inference-blind router then composes always-active shared experts, which capture distortion-invariant recovery priors, with selectively activated specialized experts, which model attack-specific residual corrections. Extensive experiments across 20+ distortion types demonstrate compositional generalization to unseen higher-order composite attacks: MoE-Shield achieves 98.43% average bit accuracy under 13 composite-attack combinations, surpassing the strongest baseline by +2.21 percentage points, while maintaining competitive parameter efficiency and high visual fidelity (～39\,dB PSNR).

Abstract:
Despite the rapid progress of Multimodal Large Language Models (MLLMs), they still suffer from untruthfulness issues, such as visual hallucinations, content fabrication, and unfaithful reasoning, which substantially undermine their faithfulness and practical utility. Alignment methods based on human preference, such as Direct Preference Optimization (DPO), have been widely adopted to address these issues. However, multimodal reasoning errors often propagate across stages, and final-answer errors can often be traced to mistakes in early grounding stages, yet standard DPO typically applies preference optimization at the final-answer level. This credit-assignment challenge means that supervision for early grounding stages is indirect rather than stage-specific, making it difficult to suppress error propagation arising from grounding drift and context inconsistency. To address this, we propose Grounded Context Preference Optimization (Groc-PO), a grounded preference optimization framework for MLLMs. We further construct the Grounded Context Preference Dataset (GCPD), organizing multi-stage preference samples around three stages of Object Grounding, Contextual Grounding, and Grounded Reasoning, to capture the formation, integration, and utilization of grounded context. By introducing more explicit preference supervision over multiple grounded stages, Groc-PO strengthens context-dependent reasoning and mitigates cross-stage error propagation. Extensive experiments show that, compared with standard DPO and other strong baselines, Groc-PO achieves improved performance in hallucination mitigation, faithful reasoning, and overall reliability, supporting the value of more explicit grounded supervision for trustworthy multimodal reasoning.

Abstract:
Continual visionlanguage learning is essential for long-lived multimodal systems, yet existing methods still struggle to retain query-relevant grounding under non-stationary task streams. We study the core question of what is forgotten first in continual multimodal reasoning and aim to close the gap between semantic retention and evidence preservation. We propose SSBS, a saliency-ordered memory framework that separates evidence identification from semantic prediction, stores only a small set of query-critical evidence units, and preserves their relative priority during replay. The key insight is that continual forgetting often begins from drift in where to lookand especially in what order to lookbefore it becomes visible in the final answer, so the replay state should explicitly encode evidence priority rather than only targets, prompts, or global hidden states. On the VQACL benchmark, SSBS achieves the best results on all eight evaluation metrics across VQAv2 and NExT-QA, improving the strongest baseline by up to +0.74 AP while reducing forgetting by up to 0.29, and it also generalizes well to alternative continual VQA protocols and a LLaVA-based continual VQA setting. These results suggest that evidence-centric memory is a practical and effective direction for continual visionlanguage learning.

Abstract:
Vision-and-Language Navigation (VLN) requires an agent to ground different parts of an instruction as navigation progresses. Existing VLN agents typically revisit the full instruction at each step, leaving the locally executable span to be inferred implicitly through cross-modal attention or history representations. We propose SHIFT, a lightweight instruction-side module that explicitly tracks a shared instruction focus state over tokens. This state consists of a focus center, a focus spread, and a focus summary, and is reused across two stages of step-wise grounding. Focus-Guided Context Adaptation biases language-side reading toward the currently relevant region while preserving access to the remaining instruction context, and Focus-Consistent Gating gates language hidden features according to their consistency with that same focus. By explicitly estimating and reusing one shared focus state, SHIFT coordinates instruction-side context exposure and focus-consistent language feature gating under a unified step-wise control signal. The module can be integrated into transformer-based VLN backbones without changing the task formulation. On R2R, REVERIE, and RxR-English, SHIFT yields consistent gains over strong baselines across standard navigation, goal-oriented object grounding, and long-horizon instruction following.

Abstract:
Efficient token pruning is essential for deploying Large Vision-Language Models (LVLMs) in virtual reality (VR), where excessive visual tokens create a critical bottleneck. Existing pruning methods either discard user-relevant content or rely on ambiguous textual queries, limiting their effectiveness. Eye-tracking signals, naturally available in VR, capture users visual focus and can guide token pruning. However, these signals are often noisy, occasionally reflect attention to irrelevant regions, and exhibit temporal redundancy, where prolonged fixations lead to repeated computation. To address these challenges, we propose GazePruning, a gaze-guided video token pruning framework with both intra-frame and inter-frame mechanisms. We introduce multi-granularity intra-frame pruning that preserves gaze-relevant tokens based on semantic similarity while retaining global and representative tokens for holistic understanding. Furthermore, we propose gaze-guided inter-frame merging, which merges consecutive frames when the gaze-targeted objects remain visually consistent, reducing redundant computation while preserving key visual information. Experimental results on real-world egocentric datasets demonstrate that our method substantially improves efficiency while retaining user-relevant details, highlighting its high practical value for real-time and user-aligned LVLM applications in immersive VR systems.

Abstract:
Multiple-choice questions(MCQs) serve as a fundamental benchmark of multimodal machine comprehension, particularly when augmented with diagrams that serve to explicate textual information and visualize abstract concepts. A primary challenge in this domain is the manual effort and scarcity associated with creating such multimodal data. In this paper, we address this issue by formulating the problem as the joint generation of MCQs and accompanying diagrams. We propose a hierarchical restructuring strategy to segment and reorganize multimodal contexts into distinct multimodal pairs, thereby enhancing the consistency and diversity of the generated MCQs. Subsequently, building upon the generated MCQs, we design a diagram generation module that synthesizes accompanying diagrams with rich textual content by utilizing a text-image alignment mechanism to integrate semantic elements into designated positions. Experimental results on the TextbookQA and ScienceQA benchmarks demonstrate the superiority of our method in generating both MCQs and accompanying diagrams that are essential for reasoning.

Abstract:
Despite the remarkable capabilities of Large Vision-Language Models (LVLMs), they remain vulnerable to jailbreak attacks. The integration of visual modalities further expands the attack surface due to its high-dimensional nature. To ensure the secure development of LVLMs, it is essential to develop more potent jailbreak attacks to probe their safety boundaries. This work focuses on universal attacks as they better reflect real-world scenarios. However, achieving robust universality remains challenging due to the tendency to overfit specific semantics within the training corpus. To bridge this gap and enhance generalization across diverse instructions, we propose Learning Universal Semantic Simplex (LUSS), a novel jailbreak method that shifts the optimization paradigm from point-wise perturbations to continuous adversarial simplices. By incorporating a sharpness-aware penalty and a simplex-based update strategy, LUSS captures a broader range of adversarial semantics within a convex hull to mitigate overfitting. To facilitate more rigorous evaluation, we introduce HarmInstBench, a manually curated benchmark with hierarchical annotations for diverse malicious scenarios. Extensive experiments on multiple LVLMs, including MiniGPT-4, LLaVA-v1.5, and Qwen2.5-VL, demonstrate that LUSS significantly outperforms state-of-the-art baselines, achieving up to a 60.8% relative improvement in Attack Success Rate (ASR), 31.8% in toxicity score, and a 44.7% gain in Malicious Intent Fulfillment Rate (MIFR). Warning: this paper contains harmful text!

Abstract:
\beginabstract In recent years, multimodal large models have made significant progress in video understanding and reasoning. However, most of them rely on an offline setting, which assumes that both videos and questions can be fully accessed in advance. In contrast, streaming video question answering scenarios require capturing dependencies across sequential questions and adaptively controlling reasoning depth under real-time constraints. To address these limitations, we propose Streaming-o1, an adaptive reasoning framework for streaming video understanding. By explicitly modeling dependencies across questions and adaptively selecting reasoning depth, it achieves efficient and consistent real-time interaction. In the supervised fine-tuning stage, we construct a streaming video question answering dataset that incorporates multi-turn reasoning and difficulty-aware sampling, featuring chain-of-thought with multidimensional signals such as confidence. In the reinforcement learning stage, we introduce a reinforcement learning method based on Group Relative Policy Optimization (GRPO) combined with multi-dimensional reward signals such as confidence. This approach guides the model to generate multi-turn responses that are consistent, reliable, and adaptive in reasoning depth. Experimental results show that Streaming-o1 improves both general and streaming video question answering, achieving a strong balance between multi-turn consistency, complex reasoning, and efficiency. \endabstract

Abstract:
RGB-infrared object detection exploits the complementarity of visible appearance and thermal contrast, yet the two modalities are seldom equally reliable in practice. Existing methods predominantly treat them as peer semantic sources, leaving contribution control to implicit fusion weights that risk entangling stable evidence with degraded auxiliary cues under severe modality imbalance. We depart from this symmetric paradigm and reformulate multimodal detection as main-conditioned residual learning: one modality anchors a stable base representation while the other is confined to supplying bounded conditional correction. We instantiate this formulation in AIR-Det, an asymmetric two-stream detector with three components: Asymmetric Conditional Innovation Extraction (ACIE) strips auxiliary content already predictable from the main branch, retaining only genuine cross-modal innovation; Main-Conditioned Selective Residual Fusion (MCSF) injects this innovation via reliability-aware residual routing governed by low-frequency structural agreement; and Adaptive Score Calibration (ASC) contrasts fused predictions against a main-only reference during training to sharpen confidence on hard positives at zero inference overhead. Extensive experiments on four RGB-IR benchmarks (M3FD, FLIR, DroneVehicle, LLVIP) demonstrate that AIR-Det achieves state-of-the-art or highly competitive accuracy with favorable efficiency, and its successful transfer to RGB-SAR detection on M4-SAR confirms that the asymmetric residual principle generalizes beyond thermal imagery.

Abstract:
Direct Preference Optimization (DPO) has emerged as an effective approach for aligning large language models (LLMs) with human preferences. However, its adaptation to multimodal settings remains unexplored. Through representational analysis, we identify a key limitation in multimodal preference optimization, which we term visual insensitivity: models often fail to distinguish between images and those with critical visual context removed. Our theoretical analysis further uncovers two manifestations of this problem, namely Across-Image Insensitivity and Within-Image Insensitivity. To address these challenges, we propose Perception-Enhanced Alignment DPO (PEA-DPO), a framework for multimodal LLMs alignment, which explicitly leverages visual preference signals to overcome visual insensitivity. We further provide a theoretical analysis demonstrating that PEA-DPO provably mitigates both failure modes. Empirical results demonstrate that PEA-DPO enhances sensitivity to visual context while preserving the language modeling capacity of the base model. Evaluations across three hallucination benchmarks using MLLMs of varying scales show that PEA-DPO effectively mitigates visual insensitivity, achieves stronger multimodal alignment, and substantially reduces hallucinations.

Abstract:
2D3D registration remains brittle in challenging scenarios such as low overlap, occlusion, repetitive structures, and severe cross-modal ambiguity. A key reason is that existing methods improve representation learning, correspondence estimation, or pose computation in isolation, while the dominant failure mode is inherently cross-level, where errors propagate between features, correspondences, and pose. To address this limitation, we propose DMM-Align: Diffusion-based Matching Matrix Alignment, a closed-loop framework that couples correspondence refinement, pose estimation, and representation learning through a shared differentiable geometric state. Our method leverages diffusion in two coordinated roles: a geometry-aware diffusion process refines the soft matching matrix for robust correspondence estimation, while a geometry-conditioned diffusion teacher injects pose-induced supervision back into feature learning. These processes are connected via a differentiable geometric hinge that converts correspondences into a global pose and exposes geometric inconsistency to upstream modules. Extensive experiments on 7-Scenes and RGB-D Scenes V2 demonstrate that DMM-Align consistently outperforms strong baselines, especially under low-overlap and heavy-occlusion conditions, highlighting the effectiveness of closed-loop geometric feedback for robust 2D3D registration.

Authors: Kai Ye (Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China), Xiaotong You (Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China), Jianghang Lin (Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China), Jiayi Ji (Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China), Pingyang Dai (Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China), Liujuan Cao (Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China)

Abstract:
Reasoning Segmentation requires models to interpret complex, context-dependent linguistic queries to achieve pixel-level localization. Current dominant approaches rely heavily on Supervised Fine-Tuning (SFT) or Reinforcement Learning (RL). However, SFT suffers from catastrophic forgetting and domain dependency, while RL is often hindered by training instability and rigid reliance on predefined reward functions. To break the bottleneck of static inference paradigms, we propose EVOL-SAM3, a zero-shot framework that reformulates reasoning segmentation as a dynamic inference-time optimization problem. A fundamental challenge in this formulation is the non-differentiable modal gap between discrete semantic reasoning and continuous pixel-level execution. EVOL-SAM3 iteratively refines prompt populations via a Generate-Evaluate-Evolve loop, employing a Visual Arena for tournament-based fitness assessment and Semantic Mutation for error correction. A Heterogeneous Arena integrates geometric and semantic priors to ensure robust selection. Extensive experiments demonstrate that EVOL-SAM3 not only outperforms baselines in terms of primary metrics and token consumption, but also surpasses fully supervised state-of-the-art methods on the ReasonSeg benchmark in a zero-shot setting.

Abstract:
Modality missingness in deployed multimodal systems is structured, not random: sensor failures correlate with environmental conditions and privacy constraints selectively suppress semantically linked inputs. This structure induces Geometric Covariate Shift (GCS), a class-heterogeneous representational drift that deforms inter-class boundaries in the latent space. Under a local mean-matching analysis, marginal adversarial alignment can be insufficient under GCS: class-asymmetric missingness pulls incomplete embeddings toward a class-biased domain mean rather than their class-specific complete-domain neighbourhoods, with the resulting offset governed by class-prior imbalance and inter-class geometry. We propose A-STAR (\underlineAdversarial-\underlineSemantic \underlineTransfer and \underlineAdaptive \underlineRectification), a two-stage framework that couples macro-level geometric rectification with micro-level feature synergy. Stage~I corrects GCS by anchoring adversarial alignment to semantic prototypes estimated exclusively from complete samples. Stage~II restores instance-level structure via dual-stream fusion whose blending weight is derived analytically without added parameters from the geometric confidence of Stage~I, creating a principled bridge between alignment quality and restoration reliability. Experiments on CMU-MOSI and CMU-MOSEI across IID, class-asymmetric, and noise-induced missingness show consistent improvements over ten state-of-the-art baselines, with gains of up to 5.43% in binary accuracy under extreme Non-IID conditions.

Abstract:
Speech enhancement is an always-on component in mobile multimedia applications, where it must operate under stringent latency budgets alongside other computation-intensive vision tasks, such as real-time HD video calling and concurrent computational photography. Although existing lightweight speech enhancement models deliver strong acoustic performance with minimal parameter counts, their hard real-time streaming deployment on on-device NPU incurs substantial inference latency, which reduces the scheduling budget for concurrent tasks and degrades the overall real-time responsiveness of multimedia systems. To address this issue, we present Align-TCN, a hardware-aligned fully convolutional streaming architecture for on-device deployment. The proposed architecture is co-designed from both representation and computation perspectives, by constructing an aligned spectral representation that better matches NPU execution characteristics and by introducing time-frequency 2D causal convolutions for efficient parallel modeling of streaming speech features. In this way, Align-TCN achieves a better balance between deployment efficiency and enhancement performance under stringent real-time constraints, with a per-frame inference latency of only 1 ms on edge NPU. Extensive experiments on both standalone deployment and end-to-end concurrent multimedia pipelines validate the effectiveness of the proposed design. Compared to existing lightweight baselines, Align-TCN maintains a similar parameter budget while substantially reducing per-frame inference latency on edge NPU. It further improves system-level real-time performance in concurrent multimedia pipelines with 60 FPS guaranteed for concurrent tasks, while preserving competitive acoustic quality.

Abstract:
Deploying Diffusion Models (DMs) in federated settings has recently attracted growing attention for their ability to harness the strong generative capacity of diffusion models. However, their practical deployment remains hindered by privacy risks, as sharing local models can potentially leak sensitive training data. To mitigate this, we propose PrivFDM, a differentially private federated diffusion model for image synthesis that generates high-fidelity synthetic images while ensuring rigorous privacy guarantees. PrivFDM presents a local pre-training method with autoencoder?based data perturbation that injects DP noise directly into the low?dimensional representations of the training data. This design reduces data sensitivity and significantly lowers privacy budget consumption. To support effective and private learning under heterogeneous client data, we further develop a local fine-tuning strategy with model perturbation and common knowledge sharing. By integrating DP?SGD and conditioning diffusion models through label embeddings, our approach mitigates the disruptive effect of client divergence and enhances the utility of the synthesized images. We theoretically prove that PrivFDM ensures DP guarantees across the entire federated training pipeline. Extensive experiments on multiple datasets demonstrate the strong performance and practical advantages of the proposed PrivFDM in federated learning.

Abstract:
Vision-Language Models (VLMs) such as CLIP are now foundational to multimodal systems, yet their robustness to spurious correlations remains poorly understood at scale. We present the first large-scale empirical study of 194 publicly available VLMs, including 16 model families, covering a wide range of model sizes, 24 training datasets, and three evaluation benchmarks, namely ImageNet (overall performance), CelebA (typical single-attribute bias), and UrbanCars (complex multi-attribute biases). Across these settings, the Spearman correlation between model scale and performance weakens as evaluation shifts from ImageNet (\rho=0.68) to single-attribute (\rho=0.48) and further to multi-attribute (\rho=0.05) bias benchmarks. In contrast, properties of the training data (size and quality) show more consistent relationships with worst-group accuracy across both bias benchmarks. Notably, curated datasets yield improvements of up to 25% over uncurated alternatives at a comparable scale. Finally, the effect of architectural choices (e.g., patch size, image resolution) is highly context-dependent, varying with the nature of the benchmark, including the type of bias and its spatial distribution within images.

Abstract:
Despite the success of diffusion models in Video Frame Interpolation (VFI), existing methods still suffer from two critical limitations. First, latent diffusion inevitably loses fine-grained details when reconstructing images from latent representations back to the pixel space. Second, multi-step sampling incurs prohibitive memory consumption and inference latency. To address these issues, we propose \modelname, a one-step pixel diffusion framework for high-quality VFI. Specifically, \modelname~employs a progressive multi-stage architecture with dynamic patch scaling to effectively learn multi-scale motion, structural, and appearance representations. Furthermore, we propose a novel Noise-Update-Only Attention mechanism to prevent semantic degradation of the clean condition frames while reducing the computational overhead by nearly 50%. Besides, we introduce a Drift-aware Timestep Sampling strategy coupled with a tailored training objective to directly predict images in the pixel space, enabling one-step inference without compromising the quality of the generated frames. Extensive experiments show that \modelname\ achieves state-of-the-art performance. On SNU-FILM, SPEED reduces LPIPS by 8.8% while delivering 63.3% faster inference and 10.6% lower memory usage. On challenging 4K benchmarks, it further surpasses prior methods by up to 51.5% in LPIPS.

Abstract:
Query-based object detectors typically employ one-to-many matching mechanisms to enhance training supervision. However, in long-tailed scenarios, such methods are hindered by unreliable query representations and rigid positive sample assignment strategies, leading to imbalanced supervision intensity across different classes. In this paper, we present an object-level analysis of the interplay between positive sample quantity and matching quality, revealing significant disparities in supervision intensity among classes of varying frequencies. Motivated by these insights, we propose SIGMA (Supervision-Intensity Guided one-to-many MAtching), a novel framework for long-tailed object detection. Specifically, we design an adaptive matching strategy that constructs positive sample sets based on the supervision intensity of each object, effectively enriching the training supervision, particularly for objects of rare classes. Furthermore, we introduce Semantic-Aware Query Initialization and Region-Aligned Query Refinement to enhance query quality for tail classes and ensure reliable matching. Extensive experiments on LVIS v1.0 and COCO-LT demonstrate that SIGMA achieves state-of-the-art performance, with significant improvements in tail class detection.

Abstract:
Contrastive audio-language pre-training typically applies uniform optimisation pressure to all matched audio--caption pairs, despite substantial differences in caption specificity and semantic ambiguity. We propose Dynamic Margin CLAP, a training-time intervention that assigns each positive pair a margin derived from the normalised inverse document frequency of its caption tokens. The margin raises the alignment requirement for semantically specific pairs and relaxes it for generic or ambiguous descriptions, requires no external ontology, and is removed entirely at inference. We evaluate the proposed margin across multiple audio-caption corpora, Euclidean and Lorentz-hyperbolic embedding spaces, two hyperbolic parameterizations, and a multi-positive retrieval protocol. The margin consistently improves audio--text retrieval regardless of the underlying geometry, whereas the benefit of curvature is narrower, corpus-dependent, and accompanied by additional retrieval cost. Comparisons with long-tail re-weighting show that changing the positive logit inside the softmax is not equivalent to rescaling whole examples. Margin-source ablations further indicate that most of the improvement comes from introducing a positive-pair margin itself, while the IDF-based specificity signal provides a smaller additional benefit. These results position adaptive positive-pair margins as the primary contribution and hyperbolic geometry as a complementary option.

Abstract:
Open-domain open-vocabulary detection (ODOVD) requires detectors to generalize to both novel categories and unseen domains, making it more challenging than open-vocabulary detection. Existing methods typically train open-vocabulary detectors together with domain generalization modules from scratch, leading to high training cost. we propose ExDet, a lightweight category-domain collaborative generalization framework for ODOVD that enhances the cross-category and cross-domain generalization of existing detectors. ExDet consists of Text-Guided Extrapolation (TGE), a lightweight Detector-Compatible Rectification (DCR) module, and ExRPN. Specifically, TGE exploits the DeltaSpace property of vision-language models (VLMs) to infer category- and domain-aware proxy visual prototypes from text. DCR is learned from the TGE-generated prototypes in a detector training-free and real-data-free manner, and is inserted after the classification head at inference to rectify representations toward a detector-compatible source-domain visual distribution, thereby enhancing classification for targets from novel categories and unseen domains. ExRPN recalibrates proposal scores by combining semantic similarity with RPN confidence, improving recall for novel and domain-shifted objects while providing better support for subsequent classification and DCR. ExDet achieves SOTA performance on OD-LVIS, OV-LVIS, Objects365, and MSOSB.

Abstract:
With the rapid development of image editing tools and generative artificial intelligence techniques, digital image forgery has become an increasingly serious threat. Although existing image tampering localization methods can identify manipulated regions at the pixel level, their predictions may become plausible but unreliable when encountering input degradations or distribution shifts in real-world scenarios. In such cases, existing localization models lack the ability to determine whether their predicted masks remain trustworthy, raising a fundamental question: can the predicted tampering masks be trusted? To address this issue, we propose TRUST-Mask, a reliability assessment framework for image tampering localization. TRUST-Mask estimates the reliability of predicted tampering masks from two aspects: the reliability of the predicted mask itself and whether this reliability estimate remains trustworthy for the current sample. Accordingly, the framework includes a Mask Reliability Estimation (MRE) module to estimate predicted-mask reliability, a Distribution Applicability Assessment (DAA) module to evaluate whether the current sample still provides sufficient evidential support for trusting the MRE estimate, and an Adaptive Trust Calibration (ATC) module to adaptively calibrate the MRE estimate and produce a unified reliability score. Extensive experiments show that TRUST-Mask provides useful reliability cues for deployment-oriented decision-making under both distortions and distribution shifts, enabling practical strategies such as selective rejection, confidence-aware filtering, and reliability-guided model selection.

Abstract:
Vision-Language Models (VLMs) have achieved significant success by employing homogeneous Transformer architectures to process multimedia information, specifically visual and textual data. Recent studies on Large Multimodal Models (LMMs) indicate that heterogeneous structures interleaving efficient mechanisms, such as linear attention, have demonstrated improvements in both performance and inference latency compared to homogeneous designs. However, these efforts rely on handcrafted designs with static mixing patterns, which are inherently sub-optimal and difficult to adapt to specific hardware deployment targets. To bridge this gap, we propose Multi-Objective Search for Adaptive Inter-layer Composition (MOSAIC), a hardware-aware search method that automatically transforms homogeneous models into optimized heterogeneous architectures. MOSAIC integrates diverse efficiency mechanisms, including linear, sparse, and low-rank operators, into a unified search space. By formulating the selection process as a multi-objective Mixed Integer Programming (MIP) problem, our method identifies optimal configurations that maximize downstream performance under strict hardware latency constraints. To mitigate the performance degradation arising from structural transitions, we introduce a two-stage parameter recovery process. We first perform global off-policy distillation to stabilize the model's internal representations, followed by a dual-teacher on-policy distillation strategy that leverages a 235B oracle teacher for knowledge expansion while utilizing the original 4B teacher to maintain distributional stability. We validate the effectiveness of MOSAIC through MOSAIC-4B, a heterogeneous model derived from Qwen3-VL-4B-Instruct. Experimental results demonstrate that MOSAIC-4B matches the performance of the Qwen3-VL-4B-Instruct baseline across multiple benchmarks while requiring less than 2% of the training cost of the original model. Furthermore, MOSAIC-4B substantially improves inference efficiency, achieving a 1.76× prefilling speedup and 2.54× decoding acceleration. The model is publicly available at \urlhttps://huggingface.co/LiAuto-DSR/MOSAIC-4B.

Abstract:
Multi-modality image fusion (MMIF) seeks to generate images that are suitable for human viewing while also supporting downstream visual understanding. Existing approaches usually learn fusion and segmentation together, either through cascaded task supervision or by exchanging intermediate features. In most cases, however, the two tasks are coupled directly in the spatial domain, even though they depend on different forms of representation. Fusion focuses on preserving local structure and integrating complementary cross-modal details, whereas segmentation depends more on regional coherence and object-level semantics. Since these signals are largely associated with different frequency bands, direct interaction in the spatial domain often cannot effectively translate high-level semantic cues into guidance for low-level fusion. This mismatch can lead to blurred boundaries, local artifacts, and limited overall performance. To address this issue, we propose SDFMod, a semantics-guided dynamic frequency modulation framework for multi-modality image fusion. Rather than enforcing task interaction in the spatial domain, SDFMod models the collaboration between fusion and segmentation in the frequency domain, where semantic structures and perceptual details can be more explicitly coordinated. The framework consists of three progressive components. The Task-aware Frequency Separator (TFS) decouples features into task-relevant frequency components using dynamic kernel decomposition. The Frequency Filter (FreFil) then uses modality-specific frequency priors to extract complementary information. Finally, the Dynamic Frequency Prompt (DFP) adaptively balances semantic representation and perceptual detail based on task goals. Extensive experiments show that frequency-domain reformulation is a promising direction for task-oriented fusion.

Abstract:
Deploying a single traffic forecasting model across shifting conditions requires adaptation to unseen scenarios at inference time, yet existing methods demand either scenario labels, parameter updates, or prohibitive computational overhead. We formulate this challenge as zero-shot scenario adaptation. To this end, we propose PROBE (\underlinePROtotype-guided deferred task \underlineBinding for z\underlineEro-shot scenario adaptation), a unified framework that threads a single soft prototype assignment through three stages. Specifically, PROBE first discovers composable traffic prototypes via deviation-aware contrastive learning, separating anomalous patterns from routine ones without scenario annotations. Second, a posterior alignment stage grounds each prototype in a concrete energy configuration through soft prototype-to-scenario mapping. Third, at inference time, the prototype assignment composes sample-specific energy weights to refine predictions through output-space gradient steps without modifying any model parameter, enabling efficient and scenario-adaptive inference. Extensive experiments on six benchmarks show that PROBE achieves competitive accuracy with existing methods while reducing physical constraint violations by up to 12.6% through energy-guided refinement, with sub-millisecond scenario-transition latency.

Abstract:
Multimodal-attributed graphs (MMAGs) provide a unified framework for modeling complex relational data by integrating heterogeneous modalities with graph structures. While centralized learning has shown promising performance, MMAGs in real-world applications are often distributed across isolated platforms and cannot be shared due to privacy concerns or commercial constraints. Federated graph learning (FGL) offers a natural solution for collaborative training under such settings; however, existing studies largely focus on single-modality graphs and do not adequately address the challenges unique to multimodal federated graph learning (MMFGL). To bridge this gap, we present MM-OpenFGL, the first comprehensive benchmark that systematically formalizes the MMFGL paradigm and enables rigorous evaluation. MM-OpenFGL comprises 19 multimodal datasets spanning 7 application domains, 8 simulation strategies capturing modality and topology variations, 9 downstream tasks, and 57 state-of-the-art methods implemented through a modular API. Extensive experiments investigate MMFGL from the perspectives of necessity, effectiveness, robustness, and efficiency, offering valuable insights for future research on MMFGL.

Abstract:
When a single visual tokenizer serves both understanding and generation, optimizing for one objective degrades the other, a phenomenon reported across Janus, TokenFlow, and Janus-Pro but lacking formal explanation. We show this tradeoff is an information-theoretic necessity: under rate-distortion theory, understanding and generation impose fundamentally different minimum token rates on any multimedia tokenizer. Formalizing cross-modal tokenization as lossy source coding, we derive single-modality and joint rate-distortion bounds, prove that shared codebooks save rate proportional to cross-modal mutual information, and quantify the \emphprice of universality, i.e., the excess rate a single tokenizer must pay to serve both tasks. Experiments on five visual tokenizer families and SoundStream across six datasets validate these predictions: understanding accuracy saturates well before generation quality; estimated bounds rank tokenizers consistently with their operational rates; and cross-modal savings scale linearly with mutual information. As a practical payoff, theory-guided codebook sizing yields 21.3% lower reconstruction error at 2.9--5.9× less compute than grid search.

Abstract:
Multimodal large language models (MLLMs) can generate descriptions that contain objects absent from the input image. Existing training-free attention interventions often apply the same adjustment throughout decoding or treat all previously generated tokens alike, although hallucination risk and the role of response-history tokens vary across decoding steps. We first analyze these variations along two axes. Temporally, hallucinated object mentions occur more frequently in later portions of generated captions. Semantically, hallucinated generation is associated with increased attention to content-bearing response-history tokens. Based on these observations, we propose Temporal-Semantic Attention Intervention (TSAI), a single-forward-pass method that reallocates post-softmax attention weights. TSAI recovers attention from shallow-layer system tokens and suppresses response-history tokens using a temporal-semantic modulation: the penalty increases as decoding proceeds, while a static function-word list serves as a lightweight lexical proxy for distinguishing function tokens from content-bearing tokens. The recovered attention is reassigned to image and instruction tokens, with optional prefill-guided selection of image tokens. Across CHAIR, AMBER, and POPE, TSAI consistently improves hallucination-related metrics over greedy decoding and provides competitive results against recent training-free baselines without requiring an additional model forward pass.

Abstract:
Autoregressive image generation has evolved from raster-scan token prediction to structured paradigms such as multi-scale and frequency-based generation. These approaches rely on predefined spatial scales or frequency partitions to organize the generation process. Such hand-crafted orderings do not explicitly separate different types of information, causing global structures and fine details to be entangled across stages and leading to inefficient or unstable modeling. To address this problem, we propose Next-Rank Prediction, a new paradigm that organizes visual information based on matrix rank. By leveraging the singular value decomposition, an image is decomposed into rank components, where the rank-1 terms are ordered by singular value energy. This provides a data-dependent, energy-ordered organization of latent content, reducing reliance on fixed spatial scales or frequency partitions. To realize this paradigm, we introduce a Rank-Residual VAE (RR-VAE) for rank-aware tokenization and a Rank Autoregressive Transformer (RankAR) that generates tokens sequentially from high-energy components to low-energy residuals. Under the same training conditions, our proposed paradigm surpasses the strong baseline in both generation quality (gFID 4.12 vs.\ 5.04) and reconstruction fidelity (rFID 1.17 vs.\ 1.57). Moreover, we evaluate rank priors for image restoration, yielding improvements in comparisons with multi-scale baseline across different tasks.

Abstract:
Vision foundation models have achieved remarkable success in monocular depth estimation, yet they tightly couple semantic and structural cues within a single RGB observation. This coupled representation makes them fragile to real-world imaging degradations such as motion blur and exposure failure, and inherently incompatible with complementary sensing modalities that physically decouple these cues. Complementary Vision Sensors (CVS) provide a dual-pathway representation, consisting of a slow semantic pathway and a high-speed structural pathway that remain robust under adverse imaging conditions. However, directly applying RGB-foundation model to this decoupled representation leads to a fundamental modality gap, while limited CVS data prevents effective training from scratch. Here, we present a complementary-pathway distillation framework that transfers RGB foundation-model priors into the dual-pathway representation. Our approach leverages a modality-driven learning principle, where the distinct physical properties of the two pathways naturally induce semantic-structural specialization. We further introduce an asymmetric cross-modal alignment mechanism to inject stable semantic context into high-speed structural features, together with a progressive distillation strategy that stabilizes training under degraded inputs. In addition, we identify spatial-difference sensing as an effective yet underexplored modality for geometric perception. Crucially, our distillation paradigm bridges emerging brain-inspired sensing with modern computer vision. Experiments on synthetic and real-world datasets demonstrate robust, generalizable depth estimation, achieving a 757-FPS equivalent temporal resolution and maintaining a 0.26 Absolute Relative error (Abs Rel) even under severe imaging degradations.

Abstract:
Novel view synthesis from unposed multi-view images remains challenging, as the model must jointly learn scene representations and camera parameters without pose supervision. Existing approaches largely fall into two extremes: implicit latent-space rendering is flexible and easy to optimize, but often yields weakly grounded camera estimation; explicit 3D representations provide stronger geometric grounding, but introduce heavier parameterization and more fragile optimization. In this paper, we present IRIS, a fully self-supervised framework that provides a practical middle ground between these two paradigms. Instead of decoding free latent tokens or reconstructing fully explicit 3D primitives, IRIS represents the scene as a latent neural field and renders novel views by querying this field under self-predicted cameras. Specifically, projected features from reference views are aggregated at sampled 3D points to form point-wise latent features, which are then composed along target rays for rendering. This design preserves the flexibility and optimization stability of implicit modeling, while introducing stronger geometric structure than unconstrained latent rendering. Extensive experiments show that IRIS achieves strong novel view synthesis quality with competitive pose accuracy under fully self-supervised learning.

Abstract:
While Multimodal Retrieval-Augmented Generation (MM-RAG) has shown promising results, it still struggles with complex multi-hop reasoning tasks. Existing methods primarily focus on independent instance-level matching, which often fails to capture explicit relationships across modalities and documents. Although Graph-enhanced methods introduce structural modeling, they face a fundamental challenge in multimodal scenarios: incorporating fine-grained visual features leads to rapid graph expansion and retrieval noise, whereas coarse-grained representations cause the discarding of critical local evidence. To address this dilemma, we propose DualG-MRAG, a Dual-tier framework that introduces a decoupled architecture comprising Macro-reasoning and Micro-matching Graphs for Multimodal RAG. Specifically, to suppress retrieval noise by isolating global structural reasoning from fine-grained evidence matching, we construct a Macro Graph for global topological routing and a Micro Graph for precise local verification. Subsequently, to enable dynamic relevance propagation across heterogeneous evidence sources, we formulate retrieval as a query-driven message passing process via a GNN Retriever. Furthermore, to provide the generative model with coherent structural guidance, we introduce a dynamic programming decoding mechanism that extracts explicit reasoning paths directly from the GNN's forward pass, replacing the standard input of isolated document chunks. Extensive experiments demonstrate that DualG-MRAG outperforms baselines in both evidence recall and complex QA accuracy.

Abstract:
Zero-shot image captioning aims to generate image descriptions without annotated image-text pairs. Recent approaches exploit text-to-image models to synthesize training data from text-only corpora, but most focus on improving overall data quality. In contrast, we observe that synthetic image-text misalignment is often structured and fine-grained: pairs may remain globally plausible while containing missing entities or misgrounded attributes, thereby degrading supervision fidelity. As a result, methods based on global similarity for image rematching or regeneration may improve apparent plausibility, but cannot systematically repair entity-level misalignment. To address this issue, we propose ReCap, a plug-and-play framework that shifts synthetic data refinement from implicit global matching to explicit fine-grained realignment. Specifically, ReCap enforces entity-level correspondence by using detected image-supported entities to guide caption rewriting, yielding more faithful synthetic supervision. In addition, we introduce an adaptive dynamic weighted learning strategy to downweight unreliable synthetic pairs during training. As a general framework, ReCap can be integrated into existing synthetic-data pipelines. Extensive experiments show that ReCap consistently improves image-text consistency and achieves state-of-the-art performance on both in-domain and cross-domain zero-shot image captioning benchmarks.

Abstract:
Source-free domain adaptation (SFDA) aims to adapt to the target distribution by relying only on a pre-trained source model. Despite its growing importance in the medical field, the transfer process is still hindered by feature-space shifts due to distribution discrepancies across different domains. By analyzing the influence of domain-related superficial cues, we observe that current models tend to establish spurious correlations between input images and predictions. As a response to these challenges, we propose Causality-inspired Pseudo-Labeling (CausPL), a novel SFDA method that drives model predictions from unreliable guesses misled by confounders to cross-domain inference rooted in stable discriminative semantics. Specifically, to enhance the ability of pseudo-labels to transfer domain-invariant knowledge, we propose a causality-guided pseudo-label generation scheme. By constructing a structural causal model to characterize the pseudo-label generation process, we incorporate a front-door adjustment to identify the actual causal effect between features and pseudo-labels. To stabilize the learning of target domain-specific knowledge, we design a counterfactual thinking-driven balancing strategy. It mitigates the interference of domain-related non-causal confounders by comparing the counterfactual and factual branches, thereby preventing model degradation. Extensive experiments on three widely used datasets show that CausPL achieves state-of-the-art performance.

Abstract:
Visual tracking faces significant challenges due to the non-continuous variations of target states, such as occlusions, deformations, illumination changes, and background distractions. Traditional state-space models, which rely on continuous state evolution, often struggle to adapt to such abrupt changes, leading to tracking drift or target loss. To address this, we propose Spike in Mamba (SiM), a temporal-centric framework built upon an Event-driven State Space Model (ESSM) to capture non-continuous dynamics. Specifically, we introduce an Event-driven State Modulation (ESM) mechanism that selectively enhances or suppresses hidden states in response to abrupt events, enabling rapid adaptation while preserving long-term temporal consistency. In addition, an Adaptive Multi-Granularity Template Fusion (AMTF) module is incorporated to provide complementary spatial cues for more reliable temporal modeling. The proposed SiM is modality-agnostic and generalizes to both single-modality (real-time and non-real-time RGB) and multi-modality (RGB-T/D/E/Lang) tracking scenarios. Extensive experiments on eleven tracking benchmarks demonstrate that SiM achieves state-of-the-art performance, consistently improving both target perception and temporal state modeling across diverse tracking scenarios and backbone architectures.

Abstract:
Deep Imbalanced Regression (DIR) is pervasive in continuous prediction tasks across diverse modalities, such as age estimation, depth prediction, and protein mutation activity prediction, where label-scarce tail samples often carry higher practical value. However, most existing methods still learn deterministic point mappings under mean squared error or its simple variants, implicitly assuming a uniform uncertainty level across all samples and thereby overlooking the instance-wise heteroscedasticity that is widespread in long-tailed data. We further point out that even heteroscedastic negative log-likelihood suffers from a gradient coupling issue, which, under DIR scenarios, weakens the learning signal of hard tail samples and leads to optimization inertia as well as tail underfitting. To address this, we propose DUO, an uncertainty-aware long-tailed regression framework. Specifically, the proposed method models the regression target as a conditional Gaussian distribution to explicitly characterize instance-level predictive uncertainty, and transforms uncertainty into a dynamic enhancement signal for tail samples through decoupled mean-variance optimization. Furthermore, we design a distribution-guided contrastive learning mechanism that adaptively constructs positive and negative pairs based on the overlap between sample distributions, thereby alleviating feature looseness and cross-label semantic entanglement. Across visual and biological DIR benchmarks, DUO achieves the best few-shot bMAE and GM on IMDB-WIKI-DIR, AgeDB-DIR, and AAV2-DIR while remaining competitive on few-shot MAE.

Abstract:
Prized for its high-quality real-time rendering, 3D Gaussian Splatting is a key novel view synthesis technology for fields such as virtual reality and autonomous driving, where it is widely used for outdoor scene reconstruction. However, extensive unbounded and texture-scarce regions in outdoor scenes, such as the sky, pose challenges for 3DGS modeling due to the lack of initial point clouds, severely limiting rendering quality. To address this issue, this paper proposes a novel optimization approach for 3DGS in unbounded scenes. We first reveal the strong coupling effects between the sky and foreground during optimization and subsequently design a prior-based decoupled optimization strategy. This method is primarily implemented through two key components: prior-based skybox initialization (PBSI) and sphere-aligned skybox densification (SASD). PBSI employs semantic cues to effectively eliminate the coupling effects. SASD introduces a spherical alignment constraint to enable high-fidelity optimization of the sky regions. Experimental results demonstrate that the proposed method can effectively eliminate the aforementioned coupling effects.

Abstract:
Computer vision models for generated facial content, such as face editing and privacy protection, increasingly affect people, requiring similarity metrics that serve as faithful proxies for human perception. While perceptual evaluation has progressed from signal-based heuristics to representation-based metrics, current approaches are limited to behavioral modeling without cognitive alignment. They rely on implicit and spurious relations while assuming a universal observer, failing to account for inherent variations across diverse human populations. This leads to inaccurate evaluative models of stakeholders and misleading guidance for generative model debugging. Rather than treating perception as a black box, we leverage scientific findings from cognitive psychology of human face similarity perception: dependence on facial featural and configural attributes, nonlinear psychophysical response scaling, and own-group biases. We introduce the FACETS dataset and propose AlignFace, an interpretable, human-aligned, face similarity metric that encodes these cognitive principles through ante-hoc modeling. It employs visual-language modeling (VLM) to encode paired face images and text-based attributes, gated cross-attention (CA) to extract attribute-specific facial difference representations, concept bottleneck modeling (CBM) to constrain reasoning via interpretable face attributes, and neural generalized additive model (GAM) to model their nonlinear influence. Experiments found AlignFace significantly improves alignment with human subpopulation perceptions compared to baseline metrics, including recent domain-free learned perceptual metrics. By bridging learned representations and human cognitive processes, this work enables more transparent and aligned perceptual evaluation metrics for face images.

Abstract:
Current text-based person retrieval (TBPR) systems are typically built on English-pretrained vision-language backbones and evaluated on English-only, single-domain benchmarks. This setting limits their applicability to multilingual queries and encourages reliance on dataset-specific shortcuts, resulting in weak cross-domain generalization. To address these issues, we propose GRAM (Generative Reranking and Alignment for Multilingual text-based person retrieval), a compact MLLM-based TBPR framework built around a generative paradigm. By jointly optimizing contrastive retrieval and caption generation, the model is encouraged to encode fine-grained visual attributes rather than holistic image-text alignment, yielding more transferable representations across languages. We introduce cross-lingual knowledge distillation to transfer retrieval capability from English to eight target languages. This distillation stage improves English retrieval, suggesting a multilingual regularization effect. During inference, the decoder is reused to rerank top-K candidates with conditional generation probability, providing fine-grained verification with low additional latency. We also construct M-PEDES, a pioneering multilingual TBPR benchmark with eight target languages and stratified human auditing of translation quality. Extensive experiments demonstrate zero-shot multilingual transfer, further gains from distillation, and improved cross-dataset generalization over strong English and multilingual baselines.

Abstract:
Agents are increasingly susceptible to hijack attacks, where malicious instructions embedded in cross-modal inputs (e.g., adversarial images or retrieved web content) are used to deceive the agent's execution. Existing defenses typically rely on strict filtering or refusal mechanisms, which suffer from a critical limitation: over-refusal, prematurely terminating valid agentic workflows. We propose INSURE, a probing-to-mitigation framework that neutralizes cross-modal attacks while preserving task continuity. Our key insight is that multimodal IPI attacks leave distinct over-focusing signatures in the vision-language latent space. We introduce a Latent Space Trace Prober to detect inter-modal attacks based on high-intensity activation scores. Subsequently, a Mitigating Rectifier performs surgical attention steering that selectively manipulates adversarial visual-textual dependencies while amplifying task-relevant elements to restore the large model's functional trajectory. Extensive evaluations on multiple multimodal backbones show that our method achieves a competitive 0.3% ASR (average), matching commercial-grade detectors, while yielding over 50% task utility gain. Furthermore, INSURE demonstrates robust Out-of-Distribution (OOD) generalization and seamlessly secures diverse multimodal interactions, establishing a superior balance between security and efficiency.

Abstract:
Cross-view image geo-localization (CVGL) aims to determine the geographic location of a street-view query image by matching it with aerial reference images. Existing methods typically employ a single aggregator for global feature aggregation, which struggles to adapt to image semantic structures in diverse scenarios. The Mixture-of-Experts (MoE) mechanism has shown great generalization ability in diverse scenes. However, its application to CVGL is hindered by issues of expert collapse and routing inconsistency. To address these challenges, we propose a stable MoE feature aggregation method for cross-view image geo-localization (SmartGeo). SmartGeo integrates MoE into the global feature aggregation, replacing a single aggregator with a heterogeneous multi-expert aggregator. We further design a Multi-Scale Attention Router (MSAR) that adaptively assigns expert weights based on the input image semantics, thereby enhancing both scene adaptability and feature representation capability. Moreover, we design two constraints related to the CVGL task: a load-balancing loss that promotes balanced expert utilization at the batch level to mitigate expert collapse, and a routing consistency loss that enforces consistent routing distributions between the pairwise ground and satellite images effectively suppressing routing shifts and stabilizing training. Extensive experiments demonstrate that SmartGeo outperforms state-of-the-art (SOTA) methods on three benchmarks. On the challenging VIGOR dataset, SmartGeo improves R@1 by 1.21% and 19.22% in the "Same-Area" and "Cross-Area" settings, respectively, demonstrating its strong generalization ability.

Abstract:
With the rapid advancement of fine-grained speech synthesis, partial deepfake speech has emerged as an increasingly concerning threat. Recent end-to-end speech editing systems further exacerbate this challenge by enabling acoustically and perceptually seamless partial forgeries, leading to the systematic failure of existing partial deepfake detectors. Our investigation of existing partial deepfake detection models reveals that these models exhibit a critical learning bias toward model-specific shallow decoding artifacts, which fundamentally undermines their generalization. To address this issue, we propose PDetective, a framework that mitigates artifact-dependent decision-making through targeted training constraints, encouraging the model to learn more intrinsic and stable representations. We further introduce a large-scale partial editing spoofing dataset comprising more than 250k utterances to underpin and validate our framework. Experiments demonstrate that PDetective significantly improves cross-domain generalization, reducing EER by over 11% on unseen editing-based forgeries, and achieving an average EER reduction of over 30% relative to the best-performing baselines in cross-dataset evaluations.

Abstract:
The ubiquitous deployment of Large Language Models (LLMs) necessitates robust provenance verification to mitigate potential misuse. While multi-bit watermarking enables embedding detailed metadata, existing schemes face a fundamental trade-off: they either compromise text quality to ensure robustness or suffer from low extraction accuracy when maintaining statistical unbiasedness. To address this, we propose CheckMark, a novel unbiased watermarking framework. By leveraging Exponential Minimum Sampling (EMS), CheckMark ensures the watermarked distribution remains statistically identical to the original model output, thereby preserving generation quality. A key innovation is the integration of a Cyclic Redundancy Check (CRC) mechanism with a hierarchical Candidate Pool decoding strategy. This synergy significantly enhances robustness against text edits while minimizing redundancy. Extensive experiments demonstrate that CheckMark achieves superior extraction accuracy and robustness compared to state-of-the-art baselines, while maintaining zero statistical distortion.

Abstract:
The increasing multimodal disinformation, where deceptive claims are reinforced through coordinated text and visual content, poses significant challenges to automated fact-checking. Recent efforts leverage Large Language Models (LLMs) for this task, capitalizing on their strong reasoning and multimodal understanding capabilities. Emerging retrieval-augmented frameworks further equip LLMs with access to open-domain external information, enabling evidence-based verification beyond their internal knowledge. Despite their promising gains, our empirical study reveals notable shortcomings in the external search coverage and evidence quality evaluation. To mitigate those limitations, we propose \Name, an end-to-end framework for automated multimodal fact-checking. It introduces a novel evidence retrieval strategy that improves evidence coverage and filters useless information from open-domain sources, enabling the extraction of high-quality evidence for verification. Extensive experiments demonstrate that \Name achieves an accuracy of 88.3% on two public multimodal disinformation datasets and 90.2% on newly emerging claims. Compared with existing evidence retrieval strategies, our approach improves verification accuracy by up to 30.8%, highlighting the critical role of evidence quality in LLM-based disinformation verification.

Abstract:
End-to-end planning has shown strong promise for autonomous driving, but most existing methods are designed for structured urban roads and generalize poorly to unstructured mining environments. In such settings, weak road structure, terrain-induced occlusions, degraded visibility, and large unobserved regions make safe planning particularly challenging. To address these challenges, we propose UnsDrive, an end-to-end planner designed for unstructured mining scenes. UnsDrive builds an unknown-aware occupancy representation that explicitly models occupied, free, and unknown space using multi-frame visibility cues, and conditions a flow-matching planner on this representation to generate multimodal future trajectories. To improve safety under partial observability, we further introduce an occupancy trajectory consistency loss and an uncertainty-aware trajectory scorer that penalize trajectories entering non-traversable or unobserved regions. We also present MineLoop, a mining-oriented closed-loop simulator for evaluating autonomous driving under irregular road geometry, degraded visibility, heavy-vehicle interactions, and mining-specific operational constraints. Experiments in both open-loop and closed-loop settings show that UnsDrive consistently outperforms strong baselines in trajectory accuracy, collision avoidance, and long-horizon driving robustness. These results demonstrate the value of explicit unknown-space reasoning for autonomous driving in unstructured mining environments.

Abstract:
Due to its exceptional efficiency in processing large-scale data, anchor-based multi-view graph clustering has emerged as an important paradigm in multimedia analysis. However, the performance of existing methods often degrades when multi-view data are corrupted by severe noise or exhibit strong cross-view heterogeneity. Specifically, enforcing a strictly shared anchor graph may weaken the expression of view heterogeneity under rigid constraints, whereas independently learning view-specific anchor graphs is highly susceptible to noise when capturing cross-view consistency. Consequently, constructing reliable anchor graphs while robustly integrating cross-view information remains a challenging problem. To tackle this challenge, we propose a novel multi-view clustering method termed Noise-Aware Tensorized Principal Anchor Graph Learning (NTPAG). Specifically, we formulate a unified framework that integrates intra-view principal structure decomposition and cross-view tensor consistency modeling. Within this framework, the anchor representation of each view is decomposed into a principal anchor graph and a deviation term, explicitly disentangling the major clustering structure from noise and disturbances. To further characterize cross-view relationships, the principal anchor graphs from different views are stacked into a tensor. We then design a high-order tensor rank surrogate based on Gaussian Error Tensor Rank (GETR) to capture latent cross-view consistency, which helps preserve principal spectral structures and suppress residual noisy components. Extensive experiments on benchmark datasets demonstrate the effectiveness and robustness of the proposed method.

Abstract:
Multi-step scientific reasoning in K-12 STEM domains demands that Multimodal Large Language Models (MLLMs) maintain rigorous logical coherence and robust self-correction. However, conventional models often falter due to the scarcity of high-quality process-level supervision and the inherent fragility of token-level autoregressive generation, which lacks global planning. To address these challenges, we propose EduFlow, an end-to-end framework that unifies task-aware data selection, search-based trajectory synthesis, and process-guided test-time scaling. At its core, EduFlow establishes a co-evolutionary feedback loop between a process-oriented reward model, EduPRM, and a domain-adapted search framework, EduMCTS. Unlike traditional binary rewards, EduPRM provides fine-grained, step-level critiques through structured [Action, Content, Label, Reason] quadruples. This evaluator guides EduMCTS, which utilizes a pedagogical action space including a dedicated Self-Reflection node to facilitate mid-trajectory error recovery. This integration fosters a closed-loop refinement process: EduMCTS explores diverse solution spaces to synthesize complex reasoning traces that sharpen EduPRM's diagnostic precision, while EduPRM provides calibrated guidance to steer subsequent search rollouts. Through this pipeline, we curate the EduPRM-420K and EduMCTS-160K datasets. Extensive experiments demonstrate that EduFlow significantly enhances reasoning consistency and sample efficiency across various scales from 2B to 72B, enabling 7B-scale models to rival state-of-the-art proprietary systems on challenging, reasoning-intensive benchmarks. Our code and datasets will be publicly released.

Abstract:
Recent video representation methods that rely on fixed-grid, patch-wise tokenization often exhibit limited versatility. Spatially, uniformly allocating a fixed number of tokens often leads to over-encoding in low-information regions. Temporally, reducing redundancy remains challenging without explicitly distinguishing between static and dynamic content. In this work, we introduce the Gaussian Video Transformer (GVT), a versatile video representation framework built on a feed-forward 2D Gaussian Splatting (2DGS) tokenization scheme. We first extract latent rigid features from a video clip and represent them with a set of 2D Gaussians generated by our proposed Spatio-Temporal Gaussian Embedding (STGE) mechanism in a feed-forward manner. Such 2D Gaussians not only enhance spatial adaptability by assigning higher (resp., lower) rendering weights to regions with higher (resp., lower) information content during rasterization, but also improve generalization by avoiding per-video optimization. To enhance the temporal versatility, we introduce a Gaussian Set Partitioning (GSP) strategy that separates the 2D Gaussians into static and dynamic sets, which explicitly model static content shared across different time-steps and dynamic content specific to each time-step, enabling a compact representation. We evaluate GVT across four tasks: video reconstruction, video action recognition, video compression, and video generation, on the UCF101, Kinetics, and DAVIS datasets. The results demonstrate state-of-the-art reconstruction and compression performance, improved action recognition, and video generation performance comparable to the baseline MAGVIT-v2.

Abstract:
Tensor-based multi-view clustering with anchor graphs effectively exploits complementary information across views while reducing computational complexity, leading to impressive results. Existing methods typically impose low-rank constraints to capture global consistency, but often overlook the local smoothness prior inherent in real-world data, where neighboring samples on a manifold should share similar attributes. To address this limitation, we propose Multi-View Clustering via Low-Rank and Smooth Tensorized Label Learning (MVC-LSTLL). Our framework unifies low-rank and smoothness objectives through a novel tensor total variation Schatten p-norm with a non-convex surrogate, which simultaneously captures both priors and offers a tighter rank approximation than traditional tensor nuclear norm (TNN). We further propose an adaptive anchor graph learning strategy that dynamically aligns data-anchor relationships with the low-rank smooth representation to enhance clustering accuracy. We develop an optimization approach based on the Augmented Lagrangian Multiplier (ALM) algorithm for the proposed model. Extensive experiments on multiple benchmarks confirm the effectiveness and superiority of our approach.

Abstract:
Multiple Kernel Clustering (MKC) has emerged as a mainstream unsupervised technique for multi-modal fusion. It constructs a consensus clustering structure by linearly combining base kernels, which relies on the idealized assumption that input views possess both consistency and complementarity. However, in real-world applications, this assumption is frequently violated by severe kernel-level noise and redundancy. We find that even purely redundant kernels can degrade clustering performance rather than enhancing consensus, while kernel-level noise causes even severe impacts. Nevertheless, existing MKC algorithms generally lack a robust architecture and suffer from severe weight diversion, failing to mitigate such kernel-level corruption and ultimately leading to severe clustering instability. To address this limitation, we propose the Kernel-Level Noise and Redundancy Robust Multiple Kernel K-Means (KNR-MKKM) method, which explicitly decouples kernel purification from weight learning. First, we quantify base kernels using spectral features and global similarity, applying Otsu's method to discard severely corrupted kernel-level noise. Next, we employ hierarchical clustering to merge highly overlapping kernels, effectively mitigating kernel-level redundancy. Finally, utilizing the purified kernel set, we introduce a Consensus Reinforcement Regularizer within a Min-Max optimization framework to explicitly reward high-consensus valid views. Extensive experiments on 8 benchmark datasets demonstrate that KNR-MKKM significantly outperforms state-of-the-art MKC baselines in both clustering accuracy and robust stability.

Abstract:
%% Abstract \beginabstract Vision-language-action (VLA) driving models can generate fluent natural-language rationales that sound plausible even when their predicted trajectories are driven by different evidence. We argue that this failure is fundamentally architectural: the problem is not insufficient language supervision, but insufficiently shared decision evidence between the language and action branches. To reduce this decoder-side mismatch, explanation and trajectory must be forced to depend on the same compact decision substrate. We instantiate this principle in \ours with a lightweight structured bottleneck. First, we encode the scene with ego, agent, and map tokens connected by simple relational edges, and extract a Decision-Critical Subgraph \Gcrit that serves as the only decoder-side input to both heads. Second, we introduce Counterfactual Consistency Learning (CCL), which perturbs \Gcrit during training and penalizes mismatched responses across the two outputs. On nuScenes, \ours reduces contradiction rate to 6.2% while maintaining competitive planning accuracy (ADE 0.356\,m). Under supervision-matched comparisons, the gain remains substantial, showing that it is primarily architectural rather than supervisory. Controlled interventions further show that editing the shared bottleneck induces synchronized changes in both language and action. Together, these results support shared intervenable bottlenecks as a practical route to faithful multimodal action generation. \endabstract

Abstract:
On-device LLM agents are increasingly attractive yet remain bottlenecked by the prefill stage, where attention compute and memory access dominate latency under long agent contexts. Existing sparse attention designs are poorly suited: hard block routing may drop critical constraints, while token pruning yields fragmented accesses that conflict with block-based KV paging. We propose LinkCLAW, a spatiotemporal collaborative sparse attention framework for long-context, tool-augmented on-device agents. Spatially, LinkCLAW introduces Link Sparse Attention (LSA): learnable link tokens appended to each block compress block-level semantics and mediate cross-block interaction, enabling heterogeneous connectivity that combines high-fidelity link-to-token access for hotspot blocks with low-rank link-to-link interactions for background blocks. Temporally, Link-guided Asynchronous Prefill (LAP) removes the ``wait-for-all'' barrier by splitting each job into a block prefill (ordinary tokens, run upon arrival) and a chained link prefill (in logical order), overlapping compute with I/O under out-of-order arrivals without requiring special position encoding. We also introduce an asynchronous agent benchmark with heterogeneous job latency and out-of-order returns. Experiments on seven benchmarks show that LinkCLAW matches or exceeds dense-attention accuracy while achieving 2.64×, 3.11×, and 1.47× speedups in single-layer attention latency, tail prefill, and end-to-end TTFT, respectively, for on-device agent workflows on a single consumer-grade GPU.

Abstract:
Answering complex questions about long videos requires dynamically acquiring cross-modal multimedia evidence across visual scenes, speech, and environmental audioyet existing multimodal methods perform a single round of retrieval followed by answer generation, without planning what evidence to seek or verifying whether retrieved evidence truly supports the conclusion. We propose Plan-Retrieve-Verify (PRV), an LLM-driven framework that reformulates long-video audio-visual reasoning as an iterative evidence acquisition process. The LLM controller first generates an evidence plan specifying which modalities and temporal regions to query, then orchestrates six multimodal toolsvisual retriever, audio retriever, ASR parser, OCR extractor, temporal checker, and contradiction verifierto gather and cross-validate evidence, and finally synthesizes a calibrated answer with principled abstention when evidence remains insufficient or contradictory. The framework operates through a closed-loop PlanRetrieveVerifyCounter-Retrieve cycle, where each iteration refines the hypothesis based on newly gathered supporting and contradicting evidence stored in a typed evidence workspace with explicit support and contradiction edges. We further introduce counterfactual tool-use trace supervision that directly trains the LLM's planning, tool selection, and verification decisions using five targeted perturbation types (audio-swap, visual-swap, temporal-shuffle, same-entity distractor, and pseudo-support), rather than only supervising the final answer. Experiments on three complementary benchmarksLongVALE (omni-modal events), CinePile (long-form movie QA), and AVUT (audio-centric comprehension)demonstrate that PRV outperforms eleven strong baselines spanning long-context models, retrieval-augmented methods, and agentic approaches in answer accuracy by 3.85.1 points while achieving 11.612.9 points higher evidence recall and up to 63% lower calibration error, with competitive calibration on all benchmarks.

Abstract:
GUI grounding requires precise identification of interface elements from high?resolution screenshots, but processing dense visual inputs with large visionlanguage models incurs substantial computational overhead. Existing visual token compression methods typically allocate a fixed budget, ignoring the large variability of visual redundancy across different GUI layouts. We show that a fixed visual budget is suboptimal, as the necessity of high?resolution visual evidence varies substantially across interface contexts. We propose FoveaUI, a dynamic high?resolution budget allocation framework for GUI grounding. The agent first reasons over a low?resolution global view and selectively requests high?resolution crops only when fine?grained evidence is required. FoveaUI introduces three coordinated components that progressively address when to seek high?resolution detail, where to acquire it, and how to leverage it effectively: a balanced training strategy for on?demand resolution allocation, a mix?tool strategy paired with a smooth process?level reward for region and granularity selection, and locality-prioritized attention masking for efficient utilization of acquired high?resolution information. Experiments on GUI grounding benchmarks show that our method maintains strong grounding performance while substantially reducing high?resolution visual token costs.

Abstract:
We present PanelLayout, a novel framework for story-aware hierarchical panel layout generation based on synchronized diffusion. Existing methods represent layouts as independent bounding boxes, a coordinate-based set prediction paradigm that struggles to model structured panel partitioning in story and comic pages, often leading to panel overlaps, poor boundary alignment and disorganized layouts. Moreover, they lack the ability to parse panel-wise spatial proportions from long-form text and maintain consistency between page-level structure and panel-level layout. To address these limitations, PanelLayout adopts a signed distance field (SDF) at the page level to represent panel structure, directly modeling spatial partitioning as a continuous geometric field. Unlike discrete bounding box prediction, SDF enforces seamless connectivity, prevents overlap, and encourages cleaner and more consistent panel boundaries. For intra-panel elements (e.g., character boxes and speech bubbles), we still use independent bounding boxes to allow reasonable overlaps that reflect real-world comic styles. We further introduce a Semantic-to-Structure Planner to extract high-level structural semantics from long story text and incorporate them as spatial priors. Building on this design, we couple page-level and panel-level generation through shared-t synchronized diffusion, enabling hierarchical joint generation that preserves consistency across levels. Experimental results show that PanelLayout outperforms existing methods on story-page layout generation and exhibits strong plug-and-play capability, serving as a page-level structural foundation for various downstream panel-level tasks.

Abstract:
Automatic depression detection (ADD) from multimodal behavioral signals requires modeling both long-range temporal dynamics and heterogeneous cross-channel patterns. We address these challenges from a dual-asymmetry perspective. First, bidirectional temporal modeling captures complementary past-to-present and future-conditioned context, but forward and backward representations can become directionally misaligned, making direct fusion suboptimal. Second, fused representations exhibit sample-dependent channel-activation variability that fixed-selectivity recalibration may not fully accommodate; multimodal and bidirectional fusion each tend to amplify this variability. To address both issues, we propose BSSSE-ATECA. The Bidirectional Selective State Space Encoder (BSSSE) introduces a Temporal Alignment module that refines the backward stream with a temporally local 1D convolution before fusion. The Adaptive Temperature-Enhanced Channel Attention (ATECA) computes a coherence score to modulate attention temperature, enabling sample-adaptive channel selectivity with batch-size-independent inference via a frozen EMA prototype. Experiments on AVEC2013, AVEC2014, DAIC-WOZ, and E-DAIC show competitive performance in unimodal and multimodal settings. Ablations show that Temporal Alignment and ATECA are complementary, with the largest MAE gains obtained when both are used together.

Abstract:
Key frame selection is essentially a set-level optimization problem: the quality of the selected subset depends on the interactions among frames, rather than the score of any single frame. Existing methods generally exhibit three major limitations. Point-wise methods score each frame independently and ignore inter-frame dependencies. Although the training-free set-level methods explicitly model the inter-frame relationships, their selection criteria are fixed and cannot be adapted through downstream task feedback. Learnable methods can leverage data-driven training; however, they lack an explicit, differentiable set-quality objective and rely on offline-generated supervision signals. To address these limitations, we propose an end-to-end trainable and task-adaptive framework for frame selection. A Chain-of-Thought prompt conditions a Small Language Model (SLM) to extract task-specific latent query vectors, which are combined with multimodal features to enable dynamic, query-aware frame scoring. We further formulate a continuous set-level objective function that jointly accounts for relevance, coverage, and redundancy, enabling differentiable set-level optimization via Gumbel-TopK for selecting optimal frame combinations. Finally, we employ a student-teacher mutual learning strategy, in which the student selector (SLM) and teacher reasoner (MLLM) are trained to align their frame-importance distributions via KL divergence. Combined with cross-entropy loss, this design enables fully end-to-end optimization, eliminating reliance on static pseudo-labels. Experiments across multiple benchmarks, including Video-MME, LongVideoBench, MLVU, and NExT-QA, demonstrate that our method significantly outperforms existing frame-selection approaches.

Abstract:
Scanpath prediction for omnidirectional images (ODIs) aims to efficiently sample complex environments by sequentially fixating on task-relevant regions, thereby emulating human-like viewing trajectories. However, conventional methods typically process entire scenes in a passive, global manner, leading to computational costs that scale with spatial-temporal resolution, while lacking the flexibility to accommodate trajectories of varying lengths. Inspired by the active vision mechanism in humans, we propose AdaptiveGaze, a novel model that shifts gaze modeling from passive full-scene processing toward active and adaptive scanpath prediction. Specifically, AdaptiveGaze formulates scanpath prediction as a coarse-to-fine sequential decision-making process, which progressively identifies and attends to salient regions by incrementally integrating information across fixations. To simulate the human glance-and-focus behavior, AdaptiveGaze integrates two synergistic components: a global overview module for rapid contextual scanning, and a local focus module for precise fixation on high-information regions. The former extracts dynamic top-down attention via lightweight global context modeling, while the latter performs fine-grained perception over a series of localized salient patches. To further enhance these modules, we introduce a hierarchical feature sampler that adaptively modulates feature extraction based on the evolving gaze state. Comprehensive experiments on multiple eye-tracking benchmarks demonstrate that AdaptiveGaze not only achieves state-of-the-art performance but also significantly reduces computational overhead compared to existing methods.

Abstract:
Human reaction generation can be understood as a dynamic process where one person acts and the other reacts, inherently requiring the generation of the current reaction based on information from past time steps. However, existing approaches tend to overlook the exploration of reaction history as well as spatial dependence of body parts, during motion generation. In this paper, we propose InterGPT, a novel two-stage framework for generating human reactions with high realism and faithful adherence to real-world dynamics. In the first stage, we employ a Vector Quantized Variational Autoencoder (VQ-VAE) to discretize and reconstruct motions, thereby obtaining spatio-temporal latent representations of motions. In VQ-VAE, a graph-based encoder is designed to capture the skeletal dependence, which adopts hierarchical skeleton pooling, adaptive graph convolution and spatio-temporal modeling together to learn compact latent vectors. These latent vectors are temporally compressed and quantized to facilitate efficient autoregressive modeling. In the second stage, we adopt a spatio-temporal autoregressive architecture to generate reactions, conditioned on both action sequence up to the current time step and previously generated reaction sequence. A spatio-temporal block is further designed to model body-part dependence within each time step while preserving temporal causality, enabling structured online generation. Extensive experimental results on multiple datasets demonstrate that InterGPT achieves substantial improvements over previous methods.

Abstract:
Multimodal Large Language Models (MLLMs) have advanced video temporal grounding (VTG) by formulating timestamp prediction as autoregressive token generation. However, this discrete formulation is susceptible to temporal hallucination. When visual evidence is insufficient, models default to linguistic priors, yielding plausible but ungrounded timestamps. While Contrastive Decoding (CD) offers a promising training-free strategy for mitigating such hallucination, it operates uniformly across the entire vocabulary. Since accurate temporal grounding hinges on probability distinctions among a small set of numerical tokens, this uniform operation dilutes fine-grained numerical distinctions, often causing format collapse or numerically invalid outputs. In this paper, we diagnose this limitation through the grouping property of Shannon entropy, revealing that the entropy contribution from non-numerical tokens dominates the total predictive entropy. This entropy imbalance pushes the task-relevant numerical uncertainty to the margin, limiting the effectiveness of full-vocabulary CD. Building on this analysis, we propose Subspace Entropic Contrastive Decoding (SECD), which dynamically isolates the numerical subspace and confines contrastive operations to tokens that determine temporal boundaries. Since the strength of linguistic priors varies across the layer hierarchy, a single-layer reference alone is insufficient. SECD further computes renormalized entropy of the numerical subspace at multiple intermediate layers. These entropy values serve as dynamic coefficients to aggregate layer-wise predictions into a unified contrastive reference. Extensive experiments across five widely used VTG benchmarks demonstrate that SECD, as a training-free and plug-and-play module, effectively mitigates temporal hallucination and yields precise and visually grounded timestamp predictions.

Abstract:
Real-time robot control demands fast action generation. Diffusion and flow matching policies for robot control require multi-step sampling, limiting their deployment in real-time scenarios. Natively reducing the sampling steps to one sacrifices representation quality and task performance, creating a trilemma among speed, fidelity, and performance. We present One-Step Generative Policy Optimization (OGPO), a systematic framework to resolve this trilemma. OGPO first pairs a lightweight architecture with the interval velocity principle for distillation-free one-step inference, while representation spreading prevents representation quality degradation. It then performs on-policy reinforcement learning (RL) fine-tuning on this fast, stable policy to break the imitation learning ceiling. Experiments on RoboMimic and OpenAI Gym benchmarks show that OGPO matches or exceeds multi-step baselines while achieving 5--20x inference speedup and over 120Hz control frequency. Physical deployment on a Franka-Emika-Panda robot validates real-world applicability.

Abstract:
Modeling multimodal time series has emerged as an important research direction, aiming to incorporate auxiliary modalities such as textual descriptions to enhance temporal understanding. However, existing methods often treat external modalities as supplementary features, limiting their applicability to specific tasks and resulting in high costs in terms of design, training, and maintenance. In this paper, we propose GenTS, a unified generative framework for multimodal time series analysis that integrates multiple downstream tasks within a single model. GenTS is jointly trained for text-conditioned time series generation and multimodal forecasting, enabling effective modeling of the joint distribution between time series and external modalities. This unified formulation supports diverse applications, including conditional generation, forecasting, and time series editing. Moreover, by incorporating efficient time series captioning, GenTS alleviates the challenge of multimodal data scarcity. Extensive experiments on diverse real-world datasets demonstrate the effectiveness and generality of our approach across multiple tasks.

Abstract:
Open-vocabulary object detection (OVOD) relies on vision-language foundation models to connect visual regions with open-ended textual concepts, yet their globally aligned representations are better suited to holistic image--text alignment than fine-grained local reasoning. Recent distillation-based methods partially alleviate this mismatch by transferring cropped region features from CLIP, but they largely treat regions independently and therefore underuse the semantic-spatial context that is often crucial for subtle category discrimination. We address this issue with Object-Focused Neighborhood Distillation (OFND), which constructs object-focused neighborhood structure from patch-, instance-, and part-level regions and transfers it from a frozen CLIP teacher to a student through node-level, edge-level, and fused-representation consistency. A cross-granularity fusion module further aggregates complementary cues across heterogeneous region partitions. Experiments on OV-COCO, OV-LVIS, and Objects365 show that OFND consistently improves over direct region-feature distillation, reaching 47.3% AP_50^\textnovel on OV-COCO and 36.2% mAP_r on OV-LVIS. The results indicate that modeling object-focused neighborhood structure is an effective way to adapt globally-aligned CLIP to fine-grained alignment in OVOD.

Abstract:
Text-to-image cross-view geo-localization aims to retrieve target locations from aerial or ground image databases using descriptive textual queries. A fundamental challenge in this task is perceptual aliasing, where visually homogeneous but geographically distinct locationsoften conceptualized as "urban clones"cause semantic ambiguity. Existing methods predominantly rely on global feature aggregation, which struggles to establish the fine-grained cross-modal correspondences necessary to distinguish these distractors. In this paper, we propose the Dual-Perspective Alignment Network (DPAN) to tackle this bottleneck. Specifically, we introduce an Optimal Transport Fusion (OTF) module to dynamically route macroscopic contextual structures and microscopic semantic details across modalities. By enforcing global marginal constraints, OTF provides precise multi-granularity semantic routing while mitigating the noise inherent in conventional cross-attention mechanisms. Furthermore, to separate indistinguishable spatial clones, we design an Adaptive Intra-modal Dissimilarity (AID) mechanism. AID leverages geometric knee-point detection on distance distributions to adaptively penalize the embedding proximity of hard negatives, expanding the discriminative margin. Experiments on CVG-Text and GeoText-1652 demonstrate DPAN's effectiveness. On CVG-Text, it achieves 61.67% and 58.67% R@1 on New York and Tokyo satellite retrieval, outperforming the state-of-the-art by +15.42% and +21.84%, respectively. Consistent improvements on GeoText-1652 confirm cross-scale generalization. Code and pre-trained models will be released upon acceptance.

Abstract:
Diffusion models have general generative abilities but struggle to align with specific objectives. Fine-tuning can improve alignment, yet its training cost is often prohibitive. This led to training-free methods that apply objective-guided terms in sampling to bias the generation distribution toward designated regions, e.g., high-reward areas. However, these methods face two issues: (1) the strong directional bias narrows the pretrained distribution and generation diversity, and (2) indiscriminate constant guidance fails to prune redundant signals, hurting both quality and efficiency. To address the above challenges, we propose SwiftExplorer, a plugin that mitigates distribution collapse caused by excessive diversity loss and reduces compute costs. First, we adopt an Inheritance-Restart exploration mechanism to avoid early convergence, while exploration also increases the likelihood of high-reward trajectories. Additionally, it balances diversity and fidelity, adding diversity without causing a distribution over-shift. Second, our QualityEfficiency arbitration mechanism improves guidance by removing incorrect signals, and it reduces computation by dynamically stopping generation when completeness and marginal reward gain are optimal. In an extensive number of experiments and different types of evaluation metrics, the proposed SwiftExplorer achieves excellent performance on all metrics, including preference, fidelity, diversity, and richness.

Abstract:
Text-driven controllable dance generation remains under-explored, primarily due to the severe scarcity of high-quality datasets and the inherent difficulty of articulating complex choreographies. Characterizing dance is particularly challenging owing to its intricate spatial dynamics, strong directionality, and the highly decoupled movements of distinct body parts. To overcome these bottlenecks, we bridge principles from dance studies, human anatomy, and biomechanics to propose Choreographic Syntax, a novel theoretical framework with a tailored annotation system. Grounded in this syntax, we combine professional dance archives with high-fidelity motion capture data to construct DanceFlow, the most fine-grained dance dataset to date. It encompasses 41 hours of high-quality motions paired with 6.34 million words of detailed descriptions. At the model level, we introduce DanceCrafter, a tailored motion transformer built upon the Momentum Human Rig. To circumvent optimization instabilities, we construct a continuous manifold motion representation paired with a hybrid normalization strategy. Furthermore, we design an anatomy-aware loss to explicitly regulate the decoupled nature of body parts. Together, these adaptations empower DanceCrafter to achieve the high-fidelity and stable generation of complex dance sequences. Extensive evaluations and user studies demonstrate our state-of-the-art performance in motion quality, fine-grained controllability, and generation naturalness.

Abstract:
Language-conditioned humanoid motion generation should be both semantically faithful and spatially controllable. Existing text-to-motion models can generate plausible whole-body behaviors but provide limited control over root trajectories, body heading, and end-effector placement, while direct geometric conditioning may disrupt the learned motion prior. We present PACE, a robot-native goal-conditioned flow-matching framework for spatially controllable humanoid motion generation and autoregressive execution. PACE provides a unified seven-slot interface covering the pelvis, two feet, two wrists, root trajectory, and heading. It combines goal-consistent offline teacher construction, global routing for root-level constraints, and zero-gated Decoupled Spatial-Semantic Attention for local end-effector goals. Geometry-aware objectives and lightweight runtime refinement improve path and target accuracy, while prefix conditioning and learned termination support continuous multi-segment generation. Extensive experiments demonstrate improved root and end-effector targetability, full-path tracking, and simultaneous multi-goal control while preserving text-motion quality. Simulation and physical Unitree G1 experiments further demonstrate stable execution across diverse whole-body behaviors.

Abstract:
Zero-shot anomaly detection (ZSAD) aims to identify anomalous images and segment anomalous regions in unseen target domains without access to target-domain training data. Although CLIP provides strong zero-shot transferability, its pretraining primarily focuses on aligning images with object-level textual semantics, whereas anomaly detection depends more on normal/abnormal states and subtle local visual patterns. Object-agnostic prompting alleviates this mismatch by replacing category-specific names with generic object placeholders, thereby encouraging the prompts to represent transferable normal and abnormal semantics rather than object identity. However, a fixed set of static prompts cannot adapt anomaly semantics to substantial instance-level variations. Visual-conditioned prompt learning provides a natural way to achieve such adaptation, yet global image representations may dilute the cues of small and localized anomalies because they are often dominated by normal content. To address this issue, we propose CoPA, an object-agnostic contextual prompt learning framework that keeps the normal prompt branch static and dynamically adapts only the abnormal branch. CoPA explicitly models interactions between local visual patch features and abnormal context tokens, producing anomaly semantics conditioned on image-specific local evidence while preserving category independence. Experiments on diverse industrial and medical benchmarks demonstrate competitive cross-domain performance in both image-level anomaly detection and pixel-level anomaly segmentation, with limited additional computational overhead.

Abstract:
Developing expressive and responsive conversational digital humans is a cornerstone of next-generation human-computer interaction. While large language models (LLMs) have significantly enhanced dialogue capabilities, most current systems still rely on cascaded architectures that connect independent modules. These pipelines are often plagued by accumulated errors, high latency, and poor real-time performance. Since these modular pipelines often decouple semantic reasoning from motion synthesis, they face an inherent risk of information loss, which may lead to sub-optimal emotional depth compared to unified architectures. To address these challenges, we propose A\textsuperscript2-LLM, an end-to-end conversational audio avatar large language model that jointly reasons about language, audio prosody, and 3D facial motion within a unified framework. To facilitate training, we introduce FLAME-QA, a high-quality multimodal dataset designed to align semantic intent with expressive facial dynamics within a question-answering (QA) format. By leveraging deep semantic understanding, A\textsuperscript2-LLM generates emotionally rich facial movements beyond simple lip-synchronization. Experimental results demonstrate that our system achieves superior emotional expressiveness while maintaining real-time efficiency (latency within 600 ms and a 0.7 real-time factor (RTF)).

Abstract:
In long-horizon open-world multi-agent systems, existing methods often treat local anomalies as automatic triggers for communication. This default design introduces coordination noise, interrupts local execution, and overuses public interaction in cases that could be resolved locally. To address this issue, we propose a partitioned information architecture for MLLM agents that explicitly separates private execution states from public coordination states. Building on this design, we introduce two key mechanisms. First, we develop an event-triggered working memory based on system-verified outcomes to maintain compact and low-noise local state representations. Second, we propose a cost-sensitive gated escalation mechanism that determines whether cross-region communication should be initiated by jointly considering node criticality, local recovery cost, and downstream task impact. In this way, communication is transformed from a default reaction into a selective decision. Experiments conducted on long-term construction tasks in open environments demonstrate that, compared to baseline models based on strong communication and planned structures, the introduction of gated communication and a partitioned information architecture results in superior performance in terms of blueprint completion quality and execution chain length. It also improves local self-recovery, reduces ineffective escalations, and increases the utility of public communication.

Authors: Jiannan Ye (United Imaging Intelligence), Wenqing Liu (Shanghai Jiao Tong University), Xiaoyang Li (Ruijin Hospital, Shanghai Jiao Tong University School of Medicine), Zixuan Guo (Shanghai Jiao Tong University), Ruqi Huang (United Imaging Intelligence), Aixi Guo (United Imaging Intelligence), Jiaqi Wang (United Imaging Intelligence), Haozheng Tang (United Imaging Intelligence), Shineng Tang (United Imaging Intelligence), Zhihong Yan (United Imaging Intelligence), Jingyi Wu (United Imaging Intelligence), Yinru Chen (Ruijin Hospital, Shanghai Jiao Tong University School of Medicine), Ying Liu (Ruijin Hospital, Shanghai Jiao Tong University School of Medicine), Zhian Bai (Ruijin Hospital, Shanghai Jiao Tong University School of Medicine), Xubo Yang (Shanghai Jiao Tong University)

Abstract:
The Objective Structured Clinical Examination (OSCE) is a global benchmark for assessing clinical competence. However, it remains costly and difficult to scale due to its heavy reliance on expert raters. This limitation is particularly evident in procedural and skill stations, which demand consistent and comprehensive evaluation of technical maneuvers, workflow integrity, and doctor-patient communication. To address these challenges, we propose AIMRater, an AI-powered mixed-reality (MR) rater system designed to automatically assess clinical skills within existing OSCE setups. By integrating the sensing channels of off-the-shelf MR HMDs with customized medical instruments, we construct multimodal components that evaluate both the integrity of the clinical workflow and the execution quality of each procedural step. Furthermore, we introduce an interactive Virtual Standard Patient (VSP) to provide a communication context and bring humanistic-care-related considerations into procedural assessment beyond conventional manikin-based setups. Results demonstrate a strong positive correlation between AIMRater and expert human examiners. This confirms that the system reliably differentiates between high and low clinical proficiency. Finally, participant feedback also indicates that the system delivers an immersive and satisfactory examination experience. These findings suggest that AIMRater effectively standardizes clinical skill evaluations and provides important insights into how multimodal artificial intelligence can scale objective assessments for medical education.

Abstract:
Text-to-CAD generation, where a large language model produces executable parametric programs from natural language shape descriptions, is a cross-modal task spanning language, code, and 3D geometry. We analyse the token-level entropy landscape of CadQuery programs and uncover a pronounced structural dichotomy. A small fraction of tokens encode geometric design choices, such as API selection and Boolean operations. These tokens exhibit high prediction entropy. In contrast, the vast majority of syntactic tokens are predicted with near certainty. This structure has critical implications for training. Standard supervised fine-tuning compresses the entropy landscape uniformly, disproportionately flattening uncertainty at these geometry-critical positionsa failure mode we term entropy collapse. The resulting policy generates nearly identical rollouts with negligible reward contrast, causing downstream reinforcement learning to stagnate. We propose EntropyCAD, a framework that reshapes training around this entropy structure. Entropy-Stratified Distillation (ESD) assigns per-token distillation temperatures by entropy rank, selectively preserving geometric-decision uncertainty during supervised fine-tuning. Geometry-Gated Policy Optimization (GGPO) then allocates reinforcement learning gradient via a per-token weight that fuses the entropy prior with a geometric-consistency signal derived from rendered rollout shapes. On the Text2CAD benchmark, EntropyCAD establishes state-of-the-art performance in Chamfer Distance, F1, and IoU, demonstrating that entropy-aware training is a principled and effective paradigm for bridging language and geometry in CAD code generation.

Abstract:
Multimodal image alignment is an essential step for vision tasks that rely on multiple modalities. Most existing methods use deep networks to warp one image to match another. Although effective, such warping is not 3D-aware and often alters the image structure, causing distortion. To address this problem, we propose ThermalGaussian-X, a 3D-aware method that aligns multimodal images via 3D Gaussian Splatting (3DGS). Our model constructs a shared 3DGS model from unaligned multiview RGB and thermal images, automatically learning the relative camera poses between modalities. Specifically, we treat the relative poses as learnable parameters, which are optimized during 3DGS training. We design a dynamic coordinate re-anchoring to localize the coordinate systems, resulting in a stable gradient. We also design an iterative pose regularization to ensure the global consistency of relative camera poses across all viewpoints. Finally, we introduce a multi-task regularization term to improve training stability. Experiments demonstrate that our method achieves state-of-the-art results in both unaligned multimodal reconstruction and multimodal alignment. The code will be released.

Abstract:
Image Difference Captioning (IDC) focuses on generating natural language descriptions of the differences between similar image pairs. While methods based on Multimodal Large Language Models (MLLMs) have significantly advanced the development of IDC, their generalizability to unseen environments and novel change types remains limited. Motivated by the generalizable knowledge acquisition of rule-based Reinforcement Learning (RL), we explore this paradigm to learn robust difference perception policies. However, applying rule-based RL to IDC presents two primary challenges: (1) Ambiguous rewards for open-ended outputs. The diversity of difference captions makes it difficult to provide verifiable rewards for policy optimization. (2) Inefficient difference perception policy learning. The limitations of MLLMs in difference-aware cross-modal alignment result in a suboptimal initial policy space for subsequent RL training. To overcome the aforementioned challenges, we propose IDC-RFT, a Reinforcement Fine-Tuning framework that utilizes rule-based Reinforcement Learning (RL) for IDC. First, to provide verifiable feedback for open-ended captions, we introduce a dual-dimensional caption reward mechanism, comprising the Multi-Change Structural Reward (MCSR) and Semantic Reward (SR). Second, aiming to improve policy learning efficiency, we develop a dual-modality contrastive optimization strategy by incorporating Text- and Vision-oriented Contrastive Difference Identification (T-CDI and V-CDI) as auxiliary tasks. Experiments across seven benchmarks demonstrate that IDC-RFT significantly outperforms state-of-the-art methods with identical backbones.

Abstract:
Real-time interactive video streaming applications (e.g., cloud gaming, AR/VR) require high visual quality and low end-to-end latency with smooth playback. Yet, despite substantial optimization of packet delivery in modern RTC systems (e.g., FEC, congestion control), a key bottleneck remains overlooked: the frame receive-to-render gap. We find that although 94% of frames arrive before their deadline, only 76% render on time, as unresolved dependencies block decoding. This limitation fundamentally stems from the serial dependency structure of codecs, in which decoding a frame requires its reference frames to be decoded first. To address this issue, we control frame references to approximate the effect of independent decoding by steering frames to reference already received frames. Specifically, we introduce feedback-driven reference control, in which ACK-guided referencing uses receiver-side decode feedback to steer frames toward decode-confirmed references, bounding stall propagation while preserving inter-frame compression. However, realizing this conceptually simple idea into a practical RTC system is non-trivial. We present FireBreak, which combines frame-level decode acknowledgment, ACK-aware slot state control, reference-aware rate control, and dynamic dependency control. Implemented with libaom and WebRTC in 3,200 C++ lines, FireBreak reduces deadline miss rate during loss-recovery phase by 77% compared to P-only coding with less than 2\,dB average PSNR loss on real traces.

Abstract:
Vision-Language Models (VLMs) have demonstrated remarkable zero-shot reasoning capabilities, offering a scalable alternative to manual annotation. However, due to the systematic semantic bias inherent in VLMs, existing methods still rely heavily on manual verification or partial annotation to ensure quality, incurring high labor costs. To address this bottleneck, we propose RL-Rectify, a reinforcement learning framework designed to automate label curation with minimal human oversight. In this framework, the VLM generates candidate labels, while an autonomous RL agent acts as an auditor to selectively Accept, Rectify, or Reject these predictions. To alleviate VLM bias effectively, the agent utilizes a multi-view state space that integrates VLM semantic priors, human-annotated task priors, and intrinsic geometric manifold distances. Furthermore, it stabilizes training via a dual-granularity reward, balancing dense manifold consistency with sparse anchor validation. Extensive experiments show that RL-Rectify outperforms existing paradigms under extreme low-shot regimes: using only 1-shot supervision, it attains 81.91% and 80.10% accuracy on CIFAR-100 and Tiny-ImageNet respectively.

Abstract:
Image inpainting detection remains a major challenge in visual forensics, especially as modern tampering techniques produce increasingly realistic results, making the already subtle traces even harder to capture. Existing methods exploit diverse forensic cues but often rely on local feature aggregation or implicit binary feature separation, without explicitly constructing image-conditioned references for both pristine and tampered regions. Consequently, they may struggle to capture manipulation-specific patterns and suffer from severe feature confusion along ambiguous tampering boundaries. To address these limitations, we propose BiPro-Net, a Boundary-Aware Dual-Prototype Network for robust image inpainting detection. Its core component, a Dual-Prototype Memory (DPM) module, leverages cross-attention to dynamically construct image-conditioned memory tokens for both authentic background and tampered foreground, enabling explicit dual-reference discrimination. To further resolve feature confusion in transitional regions, we introduce a Prototype-Anchored Contrastive Loss, which constructs foreground and background class anchors during training to selectively optimize hard pixels along tampering boundaries and sharpen the decision boundary. Additionally, we design an Uncertainty-Aware Dual-Role mechanism guided by predicted uncertainty maps. It reduces the contribution of unreliable evidence when refining class-conditioned contexts while applying a learned spatial gate to emphasize ambiguous regions. Extensive experiments across multiple benchmarks demonstrate that BiPro-Net achieves leading overall performance in both detection accuracy and localization precision.

Abstract:
Multimodal continual instruction tuning can be viewed as a constrained optimization problem: updates should improve current tasks while preserving prior knowledge. Existing data-free methods address this via proxy-based regularization or architectural isolation. However, proxy methods rely on surrogate signals that do not explicitly model forgetting, often leading to suboptimal guidance, while architectural methods introduce task-specific components and added complexity. In this work, we propose a shift from proxy-based regularization to learned loss-aware optimization, where update-induced forgetting is explicitly modeled and directly optimized. Concretely, we introduce GUIDE, a predictor-guided continual tuning framework that learns a lightweight model to estimate the old-task loss increase caused by parameter updates and incorporates this prediction into the training objective to penalize harmful update directions. The predictor operates on structured representations of parameter changes and is trained to approximate long-horizon update effects in a recursive manner, enabling it to capture cumulative forgetting across tasks while preserving a unified, deployment-friendly architecture without replay or model expansion. Extensive experiments demonstrate that GUIDE consistently outperforms existing methods, achieving superior knowledge retention while maintaining strong adaptation to new tasks.

Abstract:
Volumetric streaming remains difficult to scale because receivers with overlapping fields of view are often served independently, causing repeated transmission of shared content. We present MD2G-Cast, a relay-coordinated multicast framework over Media over QUIC with an application-aware control layer for scalable multi-user volumetric delivery. MD2G-Cast jointly uses viewing overlap, receiver capability, and bandwidth conditions to form reusable multicast groups, share common Base content, and selectively admit Enhanced delivery. We formulate grouping and Enhanced admission as a sequential control problem, realize it with Proximal Policy Optimization (PPO), and train a compact relay model with teacher guidance for Enhanced admission. We implement MD2G-Cast with real MoQ processes and evaluate it with measured access traces and head-motion traces derived from a public 6DoF dataset for up to 100 users. At 20 and 100 users, MD2G-Cast keeps the receiver-side P_99 delivery interval below 40\,ms across all seven access profiles, while Rolling reaches the 500\,ms reporting cap in most cases. Across the evaluated user scales, MD2G-Cast achieves the highest or tied-highest mean system utility under homogeneous access and the highest mean utility under heterogeneous access, while reducing aggregate link load by about 27% relative to Clustering at 100 users. A matched relay-control ablation separates the control structure from its optimizer, showing that random feasible actions reduce utility while deterministic control remains competitive with PPO. Together, the results support relay coordination and selective Enhanced admission, rather than a particular policy optimizer, as the central design contribution.

Abstract:
Graphic designs nowadays predominantly contain human images, where people are posed and framed in a delicate way to guide viewers' visual exploration and evoke desired feelings. Despite recent progress in automatic graphic design generation, there has been no work on investigating the role of the human representations in design creation. In this paper, we seek to emphasize the importance of visual human representations in graphic design generation, by studying a novel task, dubbed as design-conditioned 3D human adding. To solve this task, we propose a model that adds 3D human representations into an existing partial design, by predicting their 3D poses, their 2D composition in the image containing them, as well as the spatial arrangement of the image in the input design, to generate a complete design.Our experiments show that our model outperforms baselines by predicting the 3D human poses, 2D framing and human image layout that better harmonize with the other existing design elements, and generating higher-quality holistic designs. Our work is the first to explore connections between 3D human generation and 2D graphic design generation, which would inspire future research in leveraging 3D information to build more powerful generative models for graphic designs in the 2D domain.

Abstract:
Compositional visual question answering requires Vision-Language Models (VLMs) to execute multiple reasoning operations like object selection, spatial relation resolution, and attribute verification. Despite strong aggregate performance, the mechanistic basis of VLM failures on this task remains underexplored. To address this gap, we analyze visionoperation misalignment in VLMs by examining how failures relate to specific reasoning operations and the internal computational pathways through which they arise and propagate. We introduce an Operation-centric mechanistic framework that decomposes VLM failures by both the reasoning operation where they originate and the internal computational pathway through which they propagate. Our analysis reveals four dominant failure modes: grounding failure, reasoning failure, attribute extraction failure, and language-prior dominance, each characterized by a distinct relationship between visual grounding strength and answer correctness. Through three complementary causal interventions applied across all transformer layers, we find that object-selection failures are associated primarily with feedforward computation, multi-step relational failures with late-layer direct attention, and attribute-extraction failures with answer-position feedforward computation. Validation on VSR further shows that single-step spatial failures are concentrated at object-position encoding, distinguishing them from multi-step relational composition. These findings reveal distinct computational bottlenecks across operation types and provide a principled basis for targeted diagnosis of VLM failures in multimedia reasoning.

Abstract:
Precise 3D characterization of nanomaterials is essential for unlocking structure-property relationships. However, standard electron tomography is fundamentally limited by the missing wedge problem. Consequently, conventional algorithms suffer from severe geometric distortions, a challenge further complicated by pervasive noise interference. Current learning-based methods either rely on physics-blind post-processing or employ end-to-end architectures constrained by local receptive fields, failing to capture complex 3D topologies. We propose NanoMorph-3D, a unified end-to-end framework grounded in a comprehensive Nanomorphological Taxonomy. Powered by a large-scale synthetic dataset explicitly modeling non-linear electron attenuation, we design a Physics-Driven Unrolled Network mapping proximal gradient descent into a learnable architecture. To capture complex internal topologies, we formulate a hierarchical attention mechanism with Physics-Normalization for long-range 3D dependencies and scale invariance. Crucially, our Dual-Domain strategy leverages Sinusoidal Attention to explicitly model physical projection trajectories, enforcing strict sinogram consistency to mitigate missing wedge artifacts. Finally, an unsupervised dual-stream mechanism bridges the simulation-to-reality gap. Experiments demonstrate NanoMorph-3D reconstructs diverse topologies with superior fidelity and speed.

Abstract:
Recent works in AI-generated image detection have shown that careful training data alignment can improve generalization by removing spurious correlations. However, linear probes on frozen DINOv3 representations achieve remarkably strong performance even when trained on misaligned datasets. Motivated by this result, we analyze the underlying rationale and the limits of this generalization. We find that frozen DINOv3 performs well because its decisions rely on features that faithfully represent the space of authentic images. At the same time, its final layer is less effective at capturing the subtle pixel-artifact cues that can be emphasized by aligned training data. We further observe that naively mixing aligned and misaligned data during adaptation improves sensitivity to such cues but at the cost of distorting the pre-trained representation, limiting generalization. To address this issue, we propose Anchor-Regularized Adaptation (ARA). We apply Low-Rank Adaptation to capture pixel-level artifacts while leveraging a frozen anchor classifier to avoid deviations from the original representation structure. This allows the model to exploit pixel-artifact cues without sacrificing generalization. Our method achieves state-of-the-art performance on nine diverse and challenging benchmarks, indicating that ARA enables complementary supervision from misaligned and aligned data for more effective detection.

Abstract:
Existing 3D anomaly detection methods are built on a rigid prior: normal geometry is pose-invariant and can be canonicalized through registration or alignment. This prior does not hold for articulated objects with hinge or sliding joints, where valid pose changes induce structured geometric variations that cannot be collapsed to a single canonical template, causing pose-induced deformations to be misidentified as anomalies while true structural defects are obscured. No existing benchmark addresses this challenge. We introduce ArtiAD, the first large-scale benchmark for articulated 3D anomaly detection, comprising 15,229 point clouds across 39 object categories with dense joint-angle variations and six structural anomaly types. Each sample is annotated with its joint configuration and part-level motion labels, enabling explicit disentanglement of pose-induced geometry from structural defects. ArtiAD also provides a seen/unseen articulation split to evaluate both interpolation and extrapolation to novel joint configurations. We propose Shape--Pose-Aware Signed Distance Field (SPA-SDF), a baseline that replaces the rigid prior with a continuous pose-conditioned implicit field, factorized into an articulation-independent structural prior and a Fourier-encoded joint embedding. At inference, the articulation state is recovered by minimizing reconstruction energy, and anomalies are identified as point-wise deviations from the learned manifold. SPA-SDF achieves 0.884 object-level AUROC on seen configurations and 0.874 on unseen configurations, substantially outperforming all rigid-based baselines. Our code and benchmark will be publicly released to facilitate future research.

Abstract:
Multimodal recommendation systems (MRS) jointly model user-item interaction graphs and rich item content, but this tight coupling makes user data difficult to remove once learned. Approximate machine unlearning offers an efficient alternative to full retraining, yet current MRS unlearning applies reverse updates largely uniformly across model components. We show that this uniform treatment is misaligned with modern MRS: deleted-data influence is distributed unevenly across ranking behavior, modality branches, and model modules. This non-uniformity gives rise to three bottlenecks in MRS unlearning: target-item persistence in the collaborative graph, modality imbalance across feature branches, and concentrated module-level sensitivity in the parameter space. To address this mismatch, we propose targeted reverse update (TRU), a plug-and-play unlearning framework for MRS. Instead of applying a uniform global reversal, TRU performs three coordinated interventions across the model hierarchy: a ranking fusion gate to suppress residual target-item influence in ranking, branch-wise modality scaling to preserve retained multimodal representations, and capacity-aware parameter-group selection to localize reverse updates to deletion-sensitive modules. Across two backbones, three datasets, and three unlearning regimes, TRU achieves a stronger retain--forget trade-off than MMRecUn in most settings. In two challenging user-level cases, TRU also attains favorable operating points among all evaluated baselines. Security audits report the lowest MIA BalAcc and a tie for the lowest ASR among approximate methods in both audited settings, while wall-clock trajectories show earlier convergence to favorable retain--forget regions.

Abstract:
Fusing a low resolution (LR) mosaiced hyperspectral image (HSI) with a high resolution (HR) panchromatic (PAN) image offers a promising avenue for video-rate HR-HSI imaging via single-shot acquisition, yet its severely ill-posed nature remains a significant challenge. In this work, we propose a novel semi-supervised flow matching framework for mosaiced and PAN image fusion. Unlike previous diffusion-based approaches constrained by specific protocols or handcrafted assumptions, our method seamlessly integrates an unsupervised scheme with flow matching, resulting in a generalizable and efficient generative framework. Specifically, our method follows a two-stage training pipeline. First, we pretrain an unsupervised prior network to produce an initial pseudo HR-HSI. Building on this, we then train a conditional flow matching model to generate the target HR-HSI, introducing a random voting mechanism that iteratively refines the initial HR-HSI estimate, enabling robust and effective fusion. During inference, we employ a conflict-free gradient guidance strategy that ensures spectrally and spatially consistent HR-HSI reconstruction. Experiments on multiple benchmark datasets demonstrate that our method achieves superior quantitative and qualitative performance over representative baselines by a significant margin. Beyond mosaiced and PAN fusion, our approach provides a flexible generative framework that can readily generalize to other real-world image restoration tasks.

Abstract:
Most Gaussian head avatar methods rely on accurately calibrated multi-view cameras. Under uncalibrated settings, where camera poses are obtained from image-based prediction, inaccurate pose initialization poses a major challenge across two coupled stages. During FLAME fitting, pose errors can undermine cross-view head observability and destabilize the geometric prior, while residual errors during Gaussian Splatting (GS) optimization can introduce projection inconsistency, leading to training-view overfitting and degraded novel-view quality. To this end, we propose AnyAvatar, a unified framework for high-fidelity and controllable Gaussian head avatars under uncalibrated settings. First, we introduce Feasibility-Guided FLAME Initialization, which localizes a feasible FLAME position from coarse predicted poses through Constrained Ray Proximity Localization, restoring cross-view head observability and providing a reliable geometric prior. We then introduce Pose-Driven Gaussian Optimization, which jointly refines camera poses, Gaussian attributes, and FLAME parameters during GS training, while using a structured triplane-based appearance module to suppress pose-error compensation and improve cross-view consistency. Experiments on multiple uncalibrated datasets and self-captured data demonstrate state-of-the-art performance and robustness across diverse camera pose predictors.

Abstract:
All-in-one image restoration aims to recover clean images from diverse unknown degradations using a unified model. Existing methods predominantly treat degradations as RGB scalar superpositions in the spatial domain, ignoring their true origin, which is optical vector interactions between light waves and intervening media. Compressing these intricate vector variations into scalar intensities inevitably entangles features. To break this bottleneck, we shift our perspective to the frequency domain, where spatially entangled degradations explicitly manifest as structured frequency perturbations. To eliminate the perturbations, we introduce polarization, which explicitly records the vector wave cues of light, as a physical key to unlock frequency decoupling. Specifically, we reveal an intrinsic mapping that the Angle of Polarization naturally encodes the directional prior of frequency biases, while the Degree of Polarization reflects the severity of amplitude attenuation. Guided by the physical priors, we propose PolaRestore, a novel unified framework integrating polarization-guided frequency priors with diffusion-based control. Specifically, it leverages polarimetric cues to drive a frequency-aware attention mechanism to precisely locate and eliminate frequency perturbations, and construct a dynamic frequency dictionary with polar-aware weights for diverse degradations. Extensive experiments on multiple degraded datasets demonstrate that our frequency-driven approach achieves state-of-the-art performance across complex coupled scenarios.

Abstract:
Ultra-high-definition (UHD) video denoising requires simultaneously suppressing complex spatio-temporal degradations, preserving fine textures and chromatic stability, and maintaining efficient full-resolution 4K deployment. In this paper, we propose UHD-GPGNet, a Gaussian-process-guided local spatio-temporal denoising framework that addresses these requirements jointly. Rather than relying on implicit feature learning alone, the method estimates sparse GP posterior statistics over compact spatio-temporal descriptors to explicitly characterize local degradation response and uncertainty, which then guide adaptive temporal-detail fusion. A structure-color collaborative reconstruction head decouples luminance, chroma, and high-frequency correction, while a heteroscedastic objective and overlap-tiled inference further stabilize optimization and enable memory-bounded 4K deployment. Experiments on UVG and RealisVideo-4K show that UHD-GPGNet achieves competitive restoration fidelity with substantially fewer parameters than existing methods, enables real-time full-resolution 4K inference with significant speedup over the closest quality competitor, and maintains robust performance across a multi-level mixed-degradation schedule. A real-world study on phone-captured 4K video further confirms that the model, trained entirely on synthetic degradation, generalizes to unseen real sensor noise and improves downstream object detection under challenging conditions.

Abstract:
Multimodal Depression Recognition (MDR) aims to infer individuals depressive states from signals such as visual and linguistic. However, existing methods rely on end-to-end learning from raw features to coarse-grained depression labels, forcing heterogeneous symptom patterns to share a unified supervision signal, which leads to conflicting optimization signals and unstable representation learning, especially under limited data conditions. To address this issue, we propose an attribute-aware MDR framework via Large-Model-Guided Structured Representation Learning (LMG-SRL), which decomposes coarse-grained supervision into multiple low conflict attribute-level supervision signals and learns structured attribute representations to improve recognition. Our framework consists of three stages: i) LM-Guided Attribute Construction, which defines clinically motivated psychological attributes and leverages a large language model as a structured teacher to construct attribute-level semantic descriptions and intensity supervision for each sample; ii) Attribute Learning from Raw Data, which trains a lightweight student model to learn attribute representations from raw multimodal signals under structured attribute supervision; iii) Attribute-Aware Decision \& Analysis, which performs decision making based on the learned representations and enables attribute-level interpretation. Experimental results demonstrate that the LMG-SRL achieves superior performance and provides meaningful analysis of individual symptom differences.

Abstract:
Remote sensing image classification is a fundamental task for Earth surface observation. Joint classification using multiple sensor data has become the mainstream approach, as the discriminative information from different modalities provides the basis of fusion gains. However, joint optimization does not guarantee that each modality's discriminative information is fully exploited. When one modality suffices for a correct fused prediction, the shared error signal shrinks, compressing the optimization pressure on all encoders. This prevents the remaining confused modalities from receiving supervision commensurate with their discriminability deficiency. More critically, the weaker modality is not globally fixed but varies across samples, since land covers differ in separability across distinct physical observation dimensions. Such suppression is therefore difficult to alleviate through globally uniform modulation strategies. To this end, we propose Intra-modal Discriminability Modulation (IDM), which allocates optimization signals on a per-sample basis according to each modalitys discriminability. Specifically, IDM estimates modality-wise ambiguity from unimodal predictions and uses a Gaussian modulation function to jointly realize intra-modal discriminative enhancement and cross-modal guidance. A self-calibrated scaling strategy further adapts the modulation to different modality combinations. Experiments on multiple benchmark datasets and network architectures show that IDM consistently improves multimodal classification performance, with gains concentrated on confused samples. Additional noise perturbation and cross-task experiments further suggest that IDM is robust to modality corruption and has promising generalization ability.

Abstract:
Existing training-based conditional generation methods typically rely on auxiliary networks for intermediate state modification, such as ControlNet, or LoRA to modulate backbone, as seen in EasyControl. These approaches share two limitations: 1) the same control parameters are applied uniformly across all inference steps, and 2) separate weights must be trained and loaded for each condition type. We hypothesize that using static weights to process noise across varying intensity levels is suboptimal. Instead, employing timestep-specific weights should yield superior generation quality and faster convergence. Furthermore, training independent weights for each condition is inefficient; a unified network capable of handling multiple conditions would significantly reduce training and storage overhead. To address these challenges, we propose HyperUC: a Flow Matching Hypernetwork for Unified Conditional Generation. Specifically, we design a hypernetwork as a weight generator to produce timestep-dependent LoRA weights for the flow matching model. By conditioning on the control type, condition image, noisy latents, and timestep, HyperUC generates tailored weights adapted to diverse scenarios. This approach substantially reduces training costs while enhancing both structural fidelity and inference speed. Experiments demonstrate that HyperUC successfully generates optimized weights for eight distinct conditions, providing precise real-time guidance within only 8 sampling steps. Compared to existing training-based methods, HyperUC achieves superior efficiency and generation performance.

Abstract:
The shortage of legally compliant data for face recognition training has sparked growing interest in using synthetic data as an alternative. While recent diffusion-based methods enable the generation of photorealistic face images with strong identity adherence and data diversity, their downstream recognition performance still exhibits a significant synthetic-real gap. This paper identifies visual tendency as a previously underexplored limitation, whereby synthetic data exhibit an unrealistic prevalence of visual attributes and thus deviate from the real-data distribution. Visual tendency can be attributed to the generator's conditioning on identity embeddings, through which co-occurring residual visual cues are unintentionally absorbed into learned identity semantics. To discourage the generator from exploiting such visual cues, this paper proposes SteerFace, a simple and efficient training framework that perturbs identity embeddings by steering them toward random orthogonal directions on the embedding hypersphere. The perturbation serves as an identity-preserving regularizer that penalizes the generator's reliance on non-identity components, as supported by theoretical analysis. This paper further introduces an adaptive strategy that learns perturbation strengths with both sample-wise preference and favorable overall statistics. Experiments show that SteerFace mitigates visual tendency, improves downstream recognition performance, and generalizes across settings.

Abstract:
Multimodal domain generalization aims to learn robust models capable of mitigating performance degradation in unseen target domains. Recent works have attempted to introduce a modality-balanced optimization paradigm into this field to enhance generalization. However, we challenge this paradigm: since different modalities exhibit varying degrees of sensitivity to distribution shifts, we argue that models should structurally prioritize modalities that remain stable across domains rather than pursuing rigid modality balance. We demonstrate through domain generalization theory that enforcing strict balance can potentially yield a looser generalization bound by compelling the network to rely on modalities susceptible to distribution shifts. Motivated by these theoretical insights, we propose a novel framework called Strategic Asymmetry to structurally shift the predictive focus towards more stable modalities. Specifically, we introduce a prototype-based Dynamic Stability Estimation module to quantify modality stability, which drives a Stability-Guided Asymmetric Modulation mechanism. This mechanism actively prioritizes stable features while retaining unstable modalities as bounded complementary signals, thus reducing over-reliance on domain-susceptible information. Extensive empirical evaluations across 28 transfer tasks demonstrate that our method achieves strong and competitive performance, with robust gains across diverse distribution shifts.

Abstract:
Delivering haptic media for immersive telepresence over adverse networks requires both closed-loop operational stability and faithful texture reproduction. Existing approaches struggle to reconcile these demands, as end-to-end models amplify prediction errors into unsafe jerk spikes in 1-kHz loops, while purely passive controllers preserve safety by heavily attenuating texture cues. To mitigate this trade-off, we propose the dynamic feasibility and texture algorithm (DynaTex), which formulates degraded transmission as a layered haptic delivery representation. DynaTex decouples signals into a transport-safe kinematic reference and a learned perceptual residual, allowing a residual-only online passivity projector (OPP) to enforce conservative passivity constraints. Moving beyond pure pointwise tracking, we introduce the Multiverse Benchmark, which emphasizes operational stability and subjective quality of experience (QoE). Under volatile networks, DynaTex reduces dangerous jerk spikes by over 90% relative to predictive baselines and improves texture envelope correlation by up to 168% over passive controllers under systemic degradation. Perceptual studies further yield an overall two-alternative forced-choice (2AFC) preference rate of 83.3% for DynaTex and a strong correlation between preserved envelopes and perceived realism.

Abstract:
As the capabilities of large language models (LLMs) continue to advance, they are increasingly deployed for complex decision-making tasks. However, users' initial queries frequently exhibit intent ambiguity, necessitating multi-turn interactions to ultimately deliver satisfactory solutions. In this context, the text-only modality reveals dual limitations: for information presentation, plain text cannot efficiently deliver key information due to lengthy and unstructured responses; for information collection, users face high expressive barriers and cognitive load when attempting to precisely articulate multidimensional preferences. To address these challenges, we introduce the generative user interface (GenUI) paradigm into multi-turn decision-making scenarios. Specifically, we propose a dual-agent architecture that introduces an intermediary UI agent (IUA) between the user and the decision making agent (DMA). The IUA acts as a dynamic bridge, translating the DMA's logical deductions into interactive GenUI components for seamless user communication. To ensure accurate synchronization between these agents, we introduce a novel GenUI decision exchange (GUIDE) protocol, which enables the DMA to precisely transmit complex decision-making states and rationale to the IUA. Furthermore, to effectively power our proposed architecture, we develop a tailored three-stage data synthesis pipeline, which generates high-quality, domain-specific data to simultaneously train and optimize both the DMA and IUA. Comprehensive automated evaluations and extensive ablation studies validate the robustness of our architecture design and training strategy. A user study demonstrates the superiority of the GenUI modality in reducing users' cognitive load and improving their overall experience in decision-making tasks.

Abstract:
Multimodal data collected by heterogeneous devices are used for collaborative training, where federated learning (FL) serves as a key paradigm for effective distributed modeling with data privacy preservation. However, local training suffers from the forgetting of previously learned global knowledge under cross-client data heterogeneity, which leads to significant declines in both performance and convergence speed. Most previous studies rely on global alignment strategies to retain global knowledge, which hinder local optimization and lead to inadequate supervision of missing classes. Some studies introduce proxy datasets to supplement supervision for missing classes. However, it remains a challenge to balance class-wise global consistency and local optimization objectives without proxy datasets. In this work, we propose FedADB, a Class AnchorDriven Dual-Branch FL framework. Specifically, the server generates class anchors optimized in a differentiable input space, which are shared across clients. These class anchors serve as global references that provide supervision for missing classes during local training. A dual-branch collaborative training mechanism is designed for clients. In this mechanism, the anchor-based global branch focuses on learning with global consistency, achieving global knowledge alignment by class-anchor balanced sampling. The local calibration branch focuses on learning discriminative local features, mitigating the degradation of local representations caused by excessive global alignment. Extensive experiments across multiple medical and natural datasets demonstrate that FedADB achieves significant improvements in both accuracy and convergence speed.

Abstract:
Understanding sentiment in social media requires integrating heterogeneous textual and visual cues. However, affective meaning is rarely expressed explicitly and is often progressively constructed from perceptual impressions, semantic content, and higher-level emotional interpretation. Most existing methods still formulate multimodal sentiment analysis as a one-step fusion or alignment problem, making it difficult to capture this progressive nature of affect understanding. To address this issue, Multi-stage Semantic-guided Prompt Reasoning (MSPR) is proposed to model multimodal affect understanding as a three-stage process over Tone, Content, and Emotion. Specifically, offline Chain-of-Thought (CoT)-guided hierarchical captions provide progressive stage-wise supervision, grounding perceptual, semantic, and emotional cues in sequence. Building on this structured supervision, MSPR introduces prototype-guided cross-modal prompt interaction to transfer affective cues across textual, visual, and prototype representations. It further employs a prototype-guided dual-path contrastive alignment strategy to regularize the learned affective space and improve representation robustness. Extensive experiments on four public benchmarks demonstrate that MSPR consistently outperforms strong baselines and learns more structured and interpretable affect representations.

Abstract:
Conventional detection methods for AI-generated music typically follow a discriminative paradigm that fits decision boundaries to known fake distributions. As generative algorithms evolve, these methods remain structurally vulnerable to non-stationary distribution shifts and may overfit generator-specific artifacts. To address this limitation, we propose LAMP, a manifold-prior-guided framework that models the stable distribution of authentic music rather than continually tracking evolving forgeries. A probing study across multiple music-understanding tasks shows that different layers of a self-supervised music encoder capture complementary attributes ranging from fine-grained spectral details to high-level semantic cues, motivating a hierarchical feature extraction strategy. Layer-wise learnable memory banks are trained exclusively on authentic samples to encode compressed manifold priors at multiple representation levels. Original features and their memory-reconstructed counterparts are fused per layer and aggregated through a state-space sequence model that captures cross-layer dependencies for final decision. Experiments on FakeMusicCaps demonstrate state-of-the-art detection performance and strong robustness under diverse acoustic degradations, while zero-shot evaluation on SONICS provides initial evidence of cross-dataset transfer to unseen music generators.

Abstract:
Recent advances in diffusion models have improved Real-World Image Super-Resolution (Real-ISR), but lack human feedback integration, risking misalignment with human preference and potentially leading to artifacts, hallucinations, and harmful content generation. To this end, we are the first to introduce human preference alignment into Real-ISR, a technique that has been successfully applied in Large Language Models and Text-to-Image tasks to effectively enhance the alignment of generated outputs with human preferences. Specifically, we introduce Direct Preference Optimization (DPO) into Real-ISR to achieve alignment, where DPO serves as a general alignment technique that directly optimizes from human preferences. Nevertheless, the pixel-level reconstruction objectives of Real-ISR are difficult to reconcile with the image-level preferences of DPO, which can lead to DPO being overly sensitive to local anomalies, leading to reduced generation quality. To resolve this challenge, we propose Direct Semantic Preference Optimization (DSPO) to align instance-level human preferences by incorporating semantic guidance, which consists of two strategies: (a) semantic instance alignment strategy, implementing instance-level alignment to ensure fine-grained perceptual consistency, and (b) user description feedback strategy, mitigating hallucinations by injecting user semantic textual feedback on instance images as prompt guidance. Our method surpasses both Real-ISR and preference alignment baselines, demonstrating superior performance. As a plug-and-play solution, DSPO performs consistently across one-step and multi-step SR frameworks, highlighting strong generalizability.

Abstract:
Reconstructing high-resolution (HR) dynamic scenes from low-resolution (LR) inputs remains challenging due to insufficient details and missing temporal cues. Decoupling super-resolution from 3D reconstruction causes view-inconsistent artifacts, breaking multi-view consistency. We propose a unified event-guided framework that directly models 3D Gaussians in the HR space and supervises them through LR observations via an analysis-by-synthesis strategy. To convert the high temporal resolution of event cameras into super-resolution capability, we introduce multi-scale event temporal slicing that provides complementary supervision across temporal frequencies. We propose HR-aware regularization, combining 2D total variation and 3D scale ratio constraints to prevent Gaussian degeneration at HR, along with a local rigidity constraint to enforce physically plausible motion in dynamic regions. Experiments on synthetic and real-world datasets show our method outperforms two-stage baselines, achieving superior detail recovery and temporal consistency in dynamic scene reconstruction.

Abstract:
While Chain-of-Thought (CoT) prompting has significantly advanced the reasoning capabilities of Multimodal Large Language Models (MLLMs), managing state updates during complex, multi-step problem-solving remains a critical bottleneck. Existing approaches typically treat reasoning history as an immutable, one-dimensional text stream. Consequently, local error corrections require either verbose downstream patches or costly full-context regeneration. This linear structure forces models to implicitly track state updates, which increases cognitive load and token consumption. This limitation is particularly detrimental in spatial domains such as geometry and SVG design, where text lacks explicit visual grounding. To address these issues, we introduce Canvas-of-Thought (Canvas-CoT), a framework that shifts multimodal reasoning from linear sequence generation to mutable, structured state manipulation. Canvas-CoT uses an HTML Document Object Model (DOM) as an explicitly updatable external reasoning substrate. This architecture allows the MLLM to perform atomic CRUD operations, enabling localized, in-place state revisions without disrupting the surrounding context. We also integrate a rendering-based critique loop that validates hard constraints by translating the DOM state into explicit visual feedback, helping to expose and resolve complex spatial or logical hallucinations. Extensive evaluations across rigorous benchmarks, including VCode, RBench-V, and MathVista, demonstrate that Canvas-CoT significantly outperforms existing baselines. By decoupling reasoning logic from state persistence and facilitating in-place corrections, Canvas-CoT establishes a highly context-efficient and visually grounded approach for advanced multimodal reasoning.

Abstract:
Class Incremental Learning (CIL) aims to enable machines to continuously acquire new knowledge without forgetting previously learned concepts. However, existing methods primarily rely on discriminative parameter expansion, making them highly prone to catastrophic forgetting due to parameter overwriting. Furthermore, their opaque black-box nature limits their reliability in high-stakes real-world applications. In contrast, humans learn by matching salient visual details with semantic structures, naturally forming traceable reasoning chains. Inspired by this cognitive process, we propose a novel Explainable Class Incremental Learning(ECIL) paradigm which shifts from opaque parametric memory to explicit, human-like reasoning. We simulate human visual attention through soft-label-guided augmentation to extract key discriminative image regions. Instead of expanding the classification head, we utilize structured semantic attributes as explicit knowledge anchors, fundamentally preventing the forgetting of prior knowledge. To bridge the visual and semantic modalities, we design an entropy-regularized optimal distribution matching (DM) strategy. This mechanism not only inherently suppresses feature interference across tasks but also explicitly quantifies the contribution of specific image regions to semantic attributes, thereby generating transparent reasoning chains for each decision. Extensive experiments demonstrate that our method achieves an excellent balance between high accuracy, low forgetting rates, and intuitive interpretability, paving the way for safe and reliable continuous learning.

Abstract:
Depression recognition from human gait provides a non-intrusive and scalable approach for mental health screening. However, existing methods largely rely on generic spatiotemporal models, overlooking clinically relevant motion patterns, a limitation further exacerbated by the scarcity of high-quality annotated datasets. To address these challenges, we propose a clinically-inspired gait representation learning framework that incorporates symptom-related priors into the learning process, enabling the model to identify where to focus and how to model depression-related gait patterns. Concretely, we incorporate joint-level significance and balance-aware topological structures within a graph-based framework to capture depression-specific gait characteristics. To alleviate data scarcity of high-quality labeled clinical data, we further leverage motion priors from large-scale action datasets via transfer learning. Extensive experiments across multiple datasets show that our method achieves superior performance and strong generalization across diverse scenarios, highlighting its potential for practical gait-based mental health assessment.

Abstract:
Scribble annotations offer an efficient alternative to costly pixel-wise labeling for medical image segmentation, yet in real clinical scenarios, scribble-annotated samples are often still limited, imposing the dual challenges of sparse supervision and annotated sample scarcity. These compounded constraints severely deprive models of the structural evidence needed for complete region recovery and precise boundary delineation. To break this bottleneck, we propose a bi-level collaborative learning framework for few-shot scribble-supervised medical image segmentation. Specifically, an upper-level learnable superpixel model is introduced to provide region-structural priors for lower-level segmentation, while superpixel-based region-wise pseudo-label propagation and a spatial-prior-guided filtering strategy are performed to generate reliable dense pseudo-labels for segmentation learning. Meanwhile, the anatomical semantics learned by the lower-level segmentation model under the guidance of the current superpixels are fed back to the upper level, further driving it to learn region-structural representations better aligned with the segmentation task. Through bidirectional interaction and collaborative learning between the upper and lower levels, the proposed framework significantly outperforms existing state-of-the-art scribble-supervised methods on the ACDC and Prostate datasets under the few-shot scribble-supervised setting.

Abstract:
Source detection in hypergraphs plays an important role in mitigating the spread of harmful rumors and ensuring social stability. Current methods frequently overlook the causal entanglement problem in observations, thereby limiting the performance of source inference. In this work, we propose a source detection approach in Hypergraphs via Causal Disentanglement (HyperCD). To decouple causal propagation semantics from confounding noise, we propose the causal disentanglement module. A dual-branch variational encoder is utilized to disentangle snapshot embeddings into causal and confounding latent spaces, ensuring that the inference module relies solely on causal propagation signals. To further enforce this separation, an adversarial inference loss term is introduced to guarantee statistical independence between the decoupled branches. Moreover, to capture diffusion dynamics in hypergraphs, an attention-based hypergraph aggregation module is designed, which comprises the node-to-hyperedge propagation and hyperedge-to-node propagation processes. Extensive experiments on six datasets demonstrate the superiority of our HyperCD in effectively mitigating the causal entanglement problem.

Abstract:
X-ray absorption spectroscopy (XAS) is a key technique for probing local atomic environments, yet learning-based modeling must bridge two heterogeneous modalities: 1D continuous spectra and 3D atomic structures. Existing approaches typically decouple forward spectrum prediction and inverse structure inference into separate regression tasks, hindering shared representation learning. Moreover, severe permutation ambiguity among identical atoms often limits inverse modeling to coarse structure descriptors rather than explicit 3D structure generation. In this work, we present Uni-XAS, a unified benchmark and learning framework that reframes bidirectional XAS modeling as a cross-modal alignment and conditional generation problem. We first propose XASLip, an alignment recipe coupling a physics-aware spectral encoder with an absorber-aware manifold optimization strategy to resolve fine-grained intra-element coordination variations. Building upon this shared latent space, we formulate forward prediction as anchored absolute-spectrum generation via retrieval-augmented decoding, effectively preventing physical scale collapse and energy drift. For the inherently ill-posed inverse problem, we introduce Permutation-Rectified Flow Matching, which integrates type-wise optimal transport into a continuous generative flow to provide a principled solution to ligand permutation ambiguity without relying on heavy high-order equivariant architectures. Evaluated on a large-scale standardized benchmark of 328,839 structure--spectrum pairs, Uni-XAS demonstrates strong performance in cross-modal retrieval, accurate absolute-spectrum prediction, and composition-conditional 3D structure generation, establishing a scalable, reproducible, and protocol-consistent foundation for multimodal learning and standardized evaluation in scientific spectroscopy.

Abstract:
Reinforcement learning (RL) for document parsing often relies on reference-based rewards rooted in edit distance (e.g., tree edit distance), yet it remains hard to optimize in the high-accuracy regime because such rewards become weakly discriminative: near-correct outputs receive very similar scores, providing limited learning signal for hard cases. We propose Step-Aware Annealing (SAA), a plug-and-play reward sharpening mechanism that progressively increases reward curvature during training, amplifying subtle quality differences among high-scoring samples while preserving stability in early learning. Built on SAA, we introduce DocPO, a document policy optimization framework with element-specific, reference-based rewards anchored by edit-distance signals: normalized string edit distance (NED) for text, tree edit distance similarity (TEDS) for tables, and a hybrid Rubric+edit reward for formulas. Experiments on OmniDocBench and DocElemHard show that SAA consistently improves GRPO-style RL across document elements over non-annealed rewards, without requiring additional human supervision for reward construction.

Abstract:
Sand painting is a process-driven art where visual appearance emerges from granular accumulation. Given a single image, reconstructing a plausible sand painting process requires modeling coherent stroke structures and material-dependent effects. Existing methods, including stroke-based optimization and diffusion-based video synthesis, often lack structural coherence and material consistency, leading to unrealistic drawing sequences. We present SandSim, a framework that reconstructs a sand painting process from a single image. We introduce a curve-guided Gaussian representation that models strokes as sequences of anisotropic primitives along continuous trajectories, whose smooth kernels capture the soft boundaries of sand strokes and enable coherent stroke formation. We further adopt a subtractive compositing scheme to model light attenuation during sand accumulation. We incorporate a semantic-guided planning module for scene decomposition and drawing order inference. Our framework jointly optimizes stroke geometry and appearance and can be integrated with a physics-based simulator for interactive sand dynamics and editing. Experiments show that our method produces temporally coherent and visually realistic results, achieving improved reconstruction quality and perceptual fidelity compared to existing approaches.

Abstract:
While Facial Recognition (FR) systems have significantly improved the efficiency of identity authentication, they concurrently pose severe privacy risks, including identity spoofing and illicit trajectory tracking. Although proactive defense mechanisms, such as adversarial perturbations and feature interventions, have been developed, existing approaches often struggle to reconcile visual fidelity with transferability, thereby limiting their real-world applicability. To address these challenges, we propose DiffDAP, a diffusion-based dual-path adversarial privacy framework that achieves semantic-level identity anonymization while preserving high-fidelity visual quality. DiffDAP synergizes a generative adversarial path with a loss-optimization adversarial path. Specifically, the generative path leverages identity prompts from target images to provide asymmetric adversarial guidance during DDIM inversion and denoising, facilitating robust identity decoupling and reconstruction. Concurrently, the optimization path incorporates visual perceptual and ensemble attack losses to fortify adversarial modifications and enhance robustness. Furthermore, DiffDAP supports localized privacy protection by confining perturbations to key facial regions, thus preserving non-sensitive content. Extensive experiments across multiple benchmarks demonstrate that our method significantly outperforms state-of-the-art alternatives in both visual quality and cross-model attack transferability.

Abstract:
Generative modeling for Computer-Aided Design (CAD) remains fundamentally challenged by the need to capture the intricate interplay between discrete topological connectivity and precise spatial geometry inherent in Boundary Representations (B-reps). Existing generative approaches largely formulate CAD construction as a sequence modeling problem, but consequently lack an explicit topological inductive bias. In this work, we present GraphCAD, a geometry-aware generative framework that centers modeling on the Edge-Adjacency Graph (EAG), a novel graph-based CAD representation tailored for generative tasks. GraphCAD adopts a hybrid graph diffusion formulation to jointly model the discrete and continuous components of EAG and introduces a geometry-aware denoising mechanism that ``renders'' node attributes into spatial point sets. These point sets are encoded by an SE(3)-equivariant backbone and fused into the node representations, thereby grounding the generative process in physical geometry. To support this study, we construct EAGBench, a large-scale dataset of engineering CAD models. Extensive experiments show that GraphCAD can generate complex CAD models with more than 50 faces and achieves state-of-the-art performance in geometric fidelity and distributional alignment compared with sequence-based baselines.

Abstract:
4D medical image interpolation aims to recover missing volumes from sparsely observed time points and is important for dynamic anatomical analysis in applications such as cardiac MRI and thoracic CT, where motion is often repetitive or near-periodic over clinically relevant intervals. A key challenge is that this structure is not always encoded directly in deformation representations for interpolation. In addition, physiological motion is often non-uniform, so equal temporal intervals do not necessarily correspond to equal amounts of anatomical change. To address these issues, we formulate interpolation as learning a continuous deformation process with a phase-structured prior. Given two endpoint volumes, we parameterize a phase-conditioned velocity field with a finite Fourier basis, which embeds near-periodic motion patterns directly into the deformation space and supports continuous querying at arbitrary target times. We further introduce a phase-aligned temporal reparameterization that maps normalized within-interval time to a latent motion phase according to deformation variation intensity, thereby better modeling non-uniform motion progression. Intermediate volumes are then synthesized by continuously warping both endpoints, followed by bidirectional fusion and lightweight residual refinement. Experiments on ACDC and 4D-Lung show that the proposed method achieves state-of-the-art performance over existing baselines while producing anatomically plausible and coherent intermediate volumes from sparse observations.

Abstract:
Benefiting from promising zero-shot capabilities of the Segment Anything Model (SAM), its medical variants have demonstrated excellent performance in a wide spectrum of medical image segmentation tasks. However, SAM-series models seek to encode diverse modalities into a unified and shared parameter space, where heterogeneous modality-specific information may lead to modality conflicts within the model, manifested as modality imbalance during multi-modal training and catastrophic forgetting during continual adaptation. To address this problem, we propose MoME-SAM, a novel medical foundation model that leverages a Mixture-of-Modality-Experts (MoME) architecture for balanced medical image segmentation and orthogonal-aware modality adaptation. Different from the Mixture-of-Experts (MoE) architecture, we take the modality prior into consideration and design the prior-guided gating mechanism by the convex combination between the data-driven gating distribution and the modality prior in MoME. Therefore, modality-specific knowledge is represented by different subsets of experts, reducing multi-modal interference and achieving balanced knowledge collaboration between modalities. For continual adaptation, we present an Orthogonal-Aware Modality Adaptation (OAMA) to preserve previous knowledge via orthogonal parameter updates and incremental expansion of experts. Extensive experiments over multiple modalities indicate that MoME-SAM achieves state-of-the-art segmentation performance on 11 public datasets for multi-modal training and 3 public datasets for continual learning.

Abstract:
Multimodal graph neural networks (MGNNs) have become an effective paradigm for modeling complex multimedia data where nodes are described by heterogeneous modalities. However, real-world multimodal graphs are often modality-incomplete, which limits the effectiveness of existing MGNNs. Meanwhile, most graph pre-training approaches assume fully observed modalities and fail to generalize well when modalities are missing. In this work, we study the problem of pre-training for incomplete multimodal graph learning, aiming to learn robust representations for modality-missing graphs. Therefore, we propose a novel and unified framework, Multimodal Graph Semantic Hub Induced Pre-training (MG-SHIP). We introduce a learnable cross-modal semantic hub graph that serves as a semantic mediator, explicitly routing information across modality-specific graphs. Through this hub-mediated interaction mechanism, the model captures transferable intra-modal knowledge and cross-modal relational dependencies. This design improves representation robustness under modality-missing conditions. To further stabilize pre-training, we incorporate a relational consistency objective to encourage robust cross-modal relations and a distribution-aware regularization to mitigate modality discrepancies. The pre-trained semantic hub further supports lightweight adaptation to downstream graphs with incomplete modalities. Extensive experiments on five real-world benchmark datasets demonstrate that MG-SHIP consistently outperforms state-of-the-art methods under different levels of label scarcity and modality incompleteness.

Abstract:
Multimodal dialogue understanding aims to integrate different modalities to infer conversational semantics. However, modality-missing problems frequently occur in real-world scenarios, which remains challenging due to uncertainty introduced by missing modalities and the difficulty of effectively adapting large language models (LLMs) under such conditions. In this paper, we propose LEAP, an LLM-EnhAnced Probabilistic framework for incomplete multimodal dialogue understanding. LEAP models multimodal features as Gaussian distributions for simultaneous reconstruction and explicit uncertainty estimation, and performs uncertainty-guided adaptive fusion that dynamically emphasizes reliable messages. On this basis, we introduce a modality-specific and layer-selective low rank adaptation strategy \textS^2\textLoRA for LLM-based semantic reasoning to provide modality-aware adaptations at varying semantic levels. Our proposed LEAP effectively mitigates the reliability gap caused by missing data and enables robust semantic understanding under incomplete data conditions. Extensive experiments on three benchmarks verify the performance and robustness of LEAP under diverse modality-missing conditions.

Abstract:
Deep hashing has demonstrated strong computational efficiency for large-scale image retrieval. However, existing deep hashing methods primarily focus on optimizing the distribution of hash codes in the Hamming space, while paying insufficient attention to the loss of rich descriptor information caused by compression and quantization during hash mapping. To address this, we propose a Mamba dynamically prompted Vision Transformer for hashing-based image retrieval. MpVH exploits the complementary sensitivities of ViT and Mamba to visual content: ViT captures global contextual semantics, whereas Mamba is more responsive to locality-biased discriminative information through selective state-space modeling and multi-path spatial scanning. Based on this observation, we introduce a Mamba prompter to generate locally discriminative prompts and a dynamic mixture-of-experts adapter to aggregate historical prompts and bridge the semantic gap across architectures. These prompts guide the ViT backbone to learn importance-aware image descriptors, preserving more discriminative information during feature compression. Extensive experiments on three benchmark datasets demonstrate that MpVH achieves competitive retrieval accuracy while maintaining high encoding efficiency.

Abstract:
With the rapid growth of multimodal data, efficient cross-modal hashing retrieval has become increasingly important. However, existing methods face difficulties in effectively capturing critical information, and the large semantic gap between modalities results in low training efficiency for multi-bit hash codes. To address these issues, we propose an Online Prompt Diffusion Hashing (OnlinePDH) method that focuses on key feature information while leveraging feature fusion techniques to reduce the semantic gap between modalities, thereby improving both training and retrieval efficiency. The OnlinePDH framework integrates prompt learning, Laplacian diffusion-based feature fusion, and a multi-bit collaborative hashing learning strategy, enabling the generation of discriminative hash codes of varying lengths in a single training run, thus saving training resources and avoiding the generation of low-quality hash codes. Extensive experiments on multiple benchmark datasets demonstrate that OnlinePDH significantly outperforms the current state-of-the-art deep cross-modal hashing methods in both accuracy and efficiency.

Abstract:
Character mobility limitations arising from injury or illness are common scenarios in video games. Currently, character animations struggle to adapt interactively to injuries at different body parts and of varying severity levels in such scenarios. Consequently, existing applications often resort to using a single generic injury animation to represent all types of damage, severely degrading players' interactive experience and immersion. To address this problem, we propose a physics-based reinforcement learning framework that generates motions for mobility-limited characters using only healthy reference motion data. Instead of conventional trajectory-tracking motion priors, we introduce a compensatory style prior based on state transitions, allowing the framework to freely explore viable strategies while preserving the original motion style. Building on this, we propose a biomechanical surrogate model, trained on high-fidelity musculoskeletal data, that maps high-dimensional muscle impairment features to dynamic joint torque limits for the reinforcement learning policy in real time. This approach avoids direct simulation of complex muscle paths, ensuring that the generated motions maintain biomechanical authenticity while satisfying the demands of real-time interactivity. Experiments demonstrate that our method successfully synthesizes authentic compensatory behaviors across a diverse range of motor skills. By maintaining stylistic tracking and ensuring biological plausibility, our approach provides a highly robust and scalable solution for real-time physics-based compensatory character control.

Abstract:
Leveraging the complementarity of visible and infrared imagery is essential for robust all-weather object detection. However, conventional low-level enhancement methods prioritize perceptual fidelity over task-oriented semantics, inevitably introducing task-irrelevant biases and degrading detection performance under severe modal-specific degradations. To address this issue, we propose a Dynamic Degradation-Aware Learning (DDAL) framework for detection-driven infraredvisible image enhancement and fusion. The core is a lightweight, task-oriented Differentiable Signal Processing (DSP) front-end with modality-specific branches, which uses explicitly parameterized physics-grounded operators to adaptively rectify asymmetric cross-modal degradations. Different from convolutional black-box enhancers, our design is lightweight, interpretable, and eliminates the requirement for paired clean supervision. By learning detection-discriminative representations from degraded inputs in an end-to-end detection-driven closed loop without paired clean supervision, our method effectively bridges the gap between low-level signal restoration and high-level visual perception. Extensive experiments on M^3FD and MSRS benchmarks demonstrate that our approach achieves state-of-the-art detection performance under diverse adverse conditions. Furthermore, it acts as a plug-and-play component with strong cross-framework transferability, facilitating practical deployment for real-world multi-modal perception systems.

Abstract:
Multimodal object detection integrates heterogeneous space-based sensors (e.g., optical, SAR, infrared) to achieve robust perception in challenging conditions, which rely on cross-modal feature alignment and fusion consistency to construct discriminative representations. However, this dependency also introduces adversarial vulnerabilities, as perturbations can disrupt cross-modal consistency and degrade fused representations. Existing adversarial attacks are primarily designed for single-modal detectors or rely on naive multimodal extensions, which either fail to effectively degrade multimodal detectors or introduce noticeable visual artifacts. In this paper, we formulate multimodal adversarial attacks as an inverse information redistribution process and propose the Information Redistribution Adversarial Attack (IRAA). IRAA adopts a dual-level disruption strategy: the Spatial Entropy Redistribution Module (SERM) disperses discriminative features to weaken cross-modal consistency at feature-level, while the Prediction Uncertainty Disruption Module (PUDM) degrades semantic confidence and localization accuracy at prediction level. To ensure effective perturbation under perceptual constraints, a Modality-Aware Optimization Strategy (MAOS) is further introduced to balance perturbations across modalities via their inherent characteristics, preserving visual quality while maintaining attack effectiveness. Extensive experiments on Optical-SAR and Optical-Infrared datasets across multiple state-of-the-art detectors demonstrate that IRAA consistently achieves significant performance degradation in both white-box and black-box settings, revealing the vulnerability of multimodal detectors.

Abstract:
Real-world multimodal knowledge graphs (MMKGs) are dynamic, with new entities, relations, and multimodal knowledge emerging over time. Existing continual knowledge graph reasoning (CKGR) methods focus on structural triples and cannot fully exploit multimodal signals from new entities. Existing multimodal knowledge graph reasoning (MMKGR) methods, however, usually assume static graphs and suffer catastrophic forgetting as graphs evolve. To address this gap, we present a systematic study of continual multimodal knowledge graph reasoning (CMMKGR). We construct several continual multimodal knowledge graph benchmarks from existing MMKG datasets and propose MRCKG, a new CMMKGR model. Specifically, MRCKG employs a multimodal-structural collaborative curriculum to schedule progressive learning based on the structural connectivity of new triples to the historical graph and their multimodal compatibility. It also introduces a cross-modal knowledge preservation mechanism to mitigate forgetting through entity representation stability, relational semantic consistency, and modality anchoring. In addition, a multimodal contrastive replay scheme with a two-stage optimization strategy reinforces learned knowledge via multimodal importance sampling and representation alignment. Experiments on multiple datasets show that MRCKG preserves previously learned multimodal knowledge while substantially improving the learning of new knowledge.

Abstract:
End-to-end autonomous driving increasingly adopts a generate-and-select paradigm, making reliable candidate evaluation critical once multiple feasible trajectories are available. Existing evaluators mainly rely on deterministic scores or single-outcome estimates, limiting their ability to capture long-horizon uncertainty and lower-tail risk. We present RiSE-Drive, a risk-sensitive long-horizon candidate evaluator that preserves the upstream perception and proposal-generation pipeline. RiSE-Drive models candidate-dependent uncertainty through scene-conditioned branches and latent consequence propagation, estimates ordered outcome quantiles, and ranks trajectories using a criterion combining the distributional location, lower-tail Conditional Value-at-Risk (CVaR), and cross-branch instability. Experiments on NAVSIM-v1, NAVSIM-v2, and the closed-loop HUGSIM benchmark demonstrate state-of-the-art performance and consistent improvements in robust candidate selection.

Abstract:
Medical Referring Expression Segmentation (MRES) aims to precisely segment specific clinical targets from medical images based on textual descriptions. Existing methods predominantly rely on explicit textual prompts and implicit cross-modal feature interactions, struggling to handle the complex and diverse referring contexts in clinical practice. We observe that medical images exhibit highly consistent anatomical priors, and textual descriptions primarily serve as an explicit conditional trigger rather than complex semantic reasoning. Motivated by this insight, we propose MRESeg, a anatomical knowledge-guided framework for MRES. Its core lies in decoupling text modeling into two distinct stages: semantic conditional triggering and explicit knowledge injection, thereby transforming diverse textual descriptions into stable anatomical guidance signals. Specifically, MRESeg constructs an anatomical knowledge graph and designs a Knowledge-guided Spatial Relation Interaction (KSRI) module to explicitly inject text-triggered spatial priors into the feature learning process. This mechanism not only improves localization accuracy for conventional referred targets but also significantly enhances the model's structured localization capability for implicit descriptions. Furthermore, a Reference-Enhanced Multi-class Optimization (REMO) strategy selectively reinforces referred target regions under global multi-class supervision, successfully synergizing semantic and referring expression segmentation within a unified framework. Additionally, we propose a data transformation strategy to rapidly adapt common medical semantic segmentation datasets into MRES training and evaluation benchmarks. Extensive experiments across five datasets encompassing diverse organs and modalities demonstrate that MRESeg achieves state-of-the-art performance in both tasks, validating its effectiveness and generalizability in complex referring scenarios. The code is available.

Abstract:
Lossless image compression fundamentally relies on accurate entropy modeling, driving a recent shift toward large language model (LLM)-based methods. However, existing approaches mainly emphasize the modeling capacity of LLMs while directly serializing images into pixel or residual sequences, leaving heterogeneous signal patterns entangled within a single prediction stream. In this paper, we argue that effective entropy modeling depends not only on model capacity, but also on the predictability of the input representation. Motivated by this insight, we propose WaveLLM, a wavelet-domain LLM-based entropy model for lossless image compression. By decomposing an image into multi-level wavelet subbands, WaveLLM converts a heterogeneous prediction problem into a set of more structured and statistically homogeneous subproblems that are more amenable to autoregressive probability estimation. To bridge numerical coding symbols with the LLM's native text token space, we introduce a dedicated numerical interface featuring value embeddings and a coefficient prediction head. Furthermore, a task-adaptive prompting strategy is designed to enable a single shared model to adapt to diverse subbands within a unified framework. Built on Llama-3.2-3B with LoRA adaptation, WaveLLM achieves 2.51, 2.24, and 2.27 bpsp on Kodak, CLIC.pro, and DIV2K, respectively, outperforming the PixelCNN-based codec using the same wavelet transform by 0.43, 0.36, and 0.46 bpsp, and surpassing recent LLM-based lossless compressors on all three benchmarks.

Abstract:
Contactless keystroke inference via WiFi sensing highlights severe privacy threats, yet its real-world feasibility is hindered by two fundamental physical and deployment bottlenecks: the strict requirement for network privileges to acquire stable sensing streams, and the inherent "waveform fusion" ambiguity of pure WiFi signals during rapid, muscle-memory typing. To overcome these limitations, we propose AirKey, a novel cross-modal sensing framework that achieves highly stealthy, zero-training PIN eavesdropping. First, to bypass network deployment barriers, AirKey exploits fundamental IEEE 802.11 mechanisms to predictably elicit Acknowledgment (ACK) responses from unmodified target devices. By passively harvesting Channel State Information (CSI) from these ACKs using a low-cost microcontroller, AirKey secures a continuous spatial sensing stream entirely without network association. Crucially, to resolve the WiFi waveform fusion bottleneck, AirKey introduces a cross-modal complementarity mechanism. By utilizing lightweight acoustic signals as precise temporal anchors, the system robustly guides the segmentation of overlapping CSI trajectories. This joint spatiotemporal fusion strictly intersects CSI-derived spatial similarities with acoustic-guided inter-keystroke timing. Extensive real-world evaluations demonstrate that AirKey achieves over 4 higher accuracy than state-of-the-art unimodal zero-training schemes, successfully recovering device-unlock PINs within 6 attempts. Ultimately, this work exposes a critical vulnerability in contemporary smart interfaces, underscoring the severe privacy implications of ubiquitous multimodal sensing.

Abstract:
Video Moment Retrieval (VMR) aims to locate specific temporal segments in untrimmed videos that semantically align with given language queries. Existing proposal-based methods suffer from cross-modal granularity mismatch and fail to fully exploit the distinct temporal cues embedded in the 2D proposal map. To address these limitations, we propose a novel framework named Learning with Temporal Coordinate Priors (LTCP) for video moment retrieval. LTCP integrates event-level semantics for high-quality proposal map initialization and explicitly models proposal relationships by leveraging the temporal cues along the coordinate axes. Specifically, we first design an event-level semantic guidance module that dynamically aggregates frame-level video features into event-level features and fuses them with multimodal features via bidirectional Mamba, yielding robust and context-rich initialization features for the subsequent proposal map. Then, we develop a temporal-coordinate axial modeling module that leverages decoupled 1D axial operations to isolate and model relationships between proposals with shared boundaries. Extensive evaluations on three benchmarks (ActivityNet-Captions, TACoS, and Ego4D-NLQ) demonstrate that LTCP achieves highly competitive and often state-of-the-art performance, validating the effectiveness of our method.

Abstract:
Large speech language models have demonstrated strong capabilities in unified cross-modal understanding and generation, yet paralinguistic cues, especially emotion, remain difficult to preserve. Existing systems typically rely on entangled acoustic representations, which allow the underlying language model to depend excessively on recovered lexical content instead of grounding its behavior in acoustic-prosodic evidence. We address this limitation with EMODE, an emotion-aware speech language model built around Dynamic Para-Semantic Experts (DPSE). DPSE decomposes continuous speech features into semantic and paralinguistic pathways, routes them dynamically, and fuses them before integration into the language model. To turn this structural decomposition into functional specialization, EMODE is trained with a three-stage curriculum consisting of semantic warm-up, paralinguistic activation, and joint refinement, guided by Orthogonal Expert Guidance (OEG), Semantic-to-Acoustic Alignment (SAA), and Gating Diversity Regularization (GDR). Experiments on SER test, empathetic response evaluation, and the newly constructed bilingual MEPA benchmark show that EMODE improves the balance between lexical fidelity and emotional sensitivity, strengthens affect-grounded response generation, and exposes the value of explicit para-semantic factorization for robust cross-corpus emotion understanding.

Abstract:
Large Vision Models (LVMs) have emerged as a compelling backbone for RGB-based gait recognition, offering rich universal representations learned from large-scale pre-training. However, generic LVMs remain ill-suited for this fine-grained biometric task because their global self-attention readily captures spurious background correlations, and their pre-training objectives bias the representation toward static, visually salient regions at the expense of the subtle limb dynamics that are most informative for identification. We present PriorGait, a framework that bridges this gap by grounding LVMs in structured semantic spatial priors. PriorGait derives gait-discriminative weight maps from background masking and human body-part parsing, then injects them through two lightweight, plug-and-play modules at complementary stages of the encoding pipeline. The Prior-Guided Patch Embedding (PGPE) module modulates patch token representations before the first Transformer block, redirecting representational capacity toward dynamic limb regions from the outset. The Channel-Guided Spatial Adaptation (CGSA) module, positioned at the encoder output, selectively amplifies gait-responsive feature channels while suppressing appearance-correlated noise. Comprehensive evaluations on mainstream RGB gait benchmarks show that PriorGait consistently surpasses prior methods and establishes new state-of-the-art results, confirming semantic spatial grounding as the critical missing inductive bias for robust LVM-based gait recognition.

Abstract:
Out-of-distribution (OOD) detection remains challenging for image classifiers, especially when near-OOD samples lie close to in-distribution (ID) class boundaries. Recent vision-language detectors improve OOD detection through class semantics, local prompting, or LLM-generated outlier concepts, but seldom use language as explicit boundary evidence between confusing ID classes. We propose Pairwise Witness Local Rejection (PWLR), which uses an MLLM offline to describe visible local cues that favor one ID class over a specific rival class. These cue phrases are then screened with ID-only data under a frozen vision-language backbone, so that only reliable local verifiers are kept. At inference, PWLR first retains a small set of globally plausible classes, then checks whether any of them is locally supported against its most relevant rivals, and finally combines this pairwise local evidence with the global class score through calibration. Experiments on ImageNet-100 far-OOD, cleaner/challenging OOD and near-OOD benchmarks show that PWLR consistently improves strong vision-language baselines across multiple backbones. Source code will be released.

Abstract:
Recent generative models can produce images with few obvious visual artifacts, weakening detectors and explanations that rely only on surface appearance. We present LaP-Forensics, a multimodal framework that augments RGB semantics with reconstruction-based forensic evidence. A frozen Stable Diffusion DDIM inversionreconstruction model provides a fixed reconstruction reference, and its residual map measures local compatibility with that reference. Independent projectors encode the RGB image and residual map before a structured WhereWhatWhy model predicts a textual analysis and an artifact mask. Supervised fine-tuning is followed by Group Relative Policy Optimization (GRPO), whose reward combines mask overlap with output-structure and evidence-reference terms. These text-side terms encourage the model to refer to the consistency map but do not constitute a verifier of free-form textual truth. A separate image-level head fuses RGB and DDIM-residual class features. Experiments show cross-generator detection on UniversalFakeDetect and competitive artifact localization on the official SynthScars benchmark. Controlled cue-construction, inversion-horizon, component, reward-term, and counterfactual analyses support the utility of the residual stream under the evaluated settings, while free-form textual faithfulness and reliability under post-processing remain open limitations.

Abstract:
Transfer-based adversarial attacks reveal models' vulnerabilities under unseen noise, serving to deceive models and assess their robustness. Among transfer-based adversarial attacks, input transformation-based attacks are popular due to their simplicity and effectiveness. However, their mechanisms remain unclear, hindering advancements in DNNs. Thus, we explore the mechanism and argue that the so-called input transformation-based attacks are actually model transformation-based attacks. The attacks utilize model transformations to train adversarial examples for a wide range of models, thereby generalizing them to more models. Furthermore, we classify transformations into 2 categories based on whether they are used during model training: in-domain model transformations and out-of-domain model transformations, and explore their roles respectively. Three findings are obtained: 1) In-domain model transformations can construct composite functions to boost adversarial transferability with a minimal number of transformations; 2) Out-of-domain transformations can also improve transferability via a large number of operations by balancing the trade-off between domain shift and surrogate model diversity; 3) Integrating L2-normalization into the attack mitigates gradient imbalance during adversarial example generation. We propose a Model Transformation-based Attack (MoTA), which fuses in-domain and out-of-domain model transformations and incorporates L2-normalization to address gradient imbalance. Extensive experiments demonstrate that MoTA achieves a mean attack success rate of 95.4% on 12 models.

Abstract:
Personalized video generation faces a significant scalability challenge: as the number of custom subjects increases, existing methods suffer from severe identity leakage and semantic entanglement, resulting in videos with blended attributes and incoherent narratives. We identify that this stems from a lack of explicit binding between textual entity phrases and their corresponding visual patterns prior to fusion in diffusion models. To address this, we propose MultiBinder, a novel framework built on a ``bind-then-narrate'' paradigm. Our approach introduces a two-stage conditioning process before video diffusion. First, Cross-modal Entity Binding devises an entity-aware Q-Former to extract identity-centric features from each subjects reference image and tightly align them with its unique entity phrase, yielding disentangled ``identity tokens.'' Second, a Multimodal Large Language Model (MLLM) acts as a ``narrative director'' in the Narrative-Aware Subject Arrangement stage, capturing the action of each subject from the text prompt and synthesizes contextualized subject features that encapsulate ``who is doing what'' and preserve role-specific distinctions. These features are then injected into a video diffusion model via dedicated conditioning pathways, ensuring precise attribute allocation. Extensive experiments on both general and newly curated multi-subject benchmarks show that MultiBinder significantly outperforms existing methods in multi-subject identity fidelity.

Abstract:
Image deblurring aims to recover sharp images from blurred ones. Recently, diffusion models have shown strong potential by providing powerful generative priors that mitigate the ill-posedness of the task. However, as high-frequency details are generated from stochastic noise, diffusion models may introduce spurious details or misaligned structures. This inherent nature limits the fidelity of restored images. To address this issue, we propose DF-Diff, a novel framework that incorporates generative priors into the Window-based Dynamic Filtering Transformer (WDFT) through the physical model of blur degradation. Specifically, diffusion priors are transformed to window-based dynamic parameters that correspond to the blur degradation model, effectively avoiding hallucinated details contaminating the restored image. Moreover, to better adapt diffusion priors for dynamic parameter estimation, we propose a new Spatial Blur-aware Diffusion Model (SBDM) that generates sharp priors consistent with the distribution of clean images. By expanding compact vector features into cubic spatial feature representation, we explicitly model spatially aware sharp priors, allowing them to improve parameter estimation especially under spatially non-uniform blur conditions. Extensive experiments on synthetic and real-world blur datasets demonstrate that DF-Diff outperforms state-of-the-art methods across widely used benchmarks.

Abstract:
Domain adaptive point cloud completion (DA PCC) aims to bridge the geometric discrepancy between labeled source domains and unlabeled target domains for accurate 3D shape completion. Existing methods based on CNNs, Transformers, and Mamba-like sequence models still face challenges from limited receptive fields, high computational complexity, insufficient geometric bias, and serialization sensitivity. Recently, Receptance Weighted Key Value (RWKV) provides a linear-complexity alternative for long-range sequence modeling. However, adapting RWKV to DA PCC remains challenging due to the lack of explicit geometric priors, the mismatch between recurrent modeling and coarse-to-fine generation, and unstable target-domain adaptation. To address these challenges, we propose DAPointRWKV, a novel RWKV-based framework for DA PCC. DAPointRWKV consists of three components: Geometry-Aware RWKV (GA-RWKV) Backbone, RWKV-Oriented Progressive Decoder (RWKV-PD), and Dual Structural Regularization (DSR). Specifically, GA-RWKV introduces geometry-aware recurrent representation learning, RWKV-PD bridges RWKV token representations with coarse-to-fine reconstruction through local token-guided refinement, and DSR stabilizes target-domain transfer through visible-structure anchoring and complete-shape hypothesis stabilization. Extensive experiments on synthetic and real-world benchmarks demonstrate that DAPointRWKV consistently outperforms state-of-the-art methods.

Abstract:
Fine-grained motion generation aims to produce motions that accurately align with user intents via detailed text, coordinates or sketches. However, existing methods often suffer from either abrupt sequential transitions or spatial misalignment. To address these limitations, we propose StickFree, a stickman-based, plug-and-play framework for text-to-motion diffusion models. It enables fine-grained motion generation via flexible and highly compatible stickman guidance and can be integrated seamlessly into existing models without requiring retraining. Given a textual description and hand-drawn stickmen, StickFree can flexibly incorporate stickman constraints at arbitrary frames within the motion sequence, generating semantically coherent and natural motions. Specifically, we introduce cluster-aware implicit guidance to alleviate the abruptness caused by guiding only the frames where constraints are applied. Moreover, we propose structure-aware textual saliency guidance to adaptively balance global text and local stickman constraints, enabling the generation of textually unspecified limb movements. Extensive experiments demonstrate that StickFree achieves zero-shot transferability across multiple diffusion-based backbones, exhibiting strong generalization and consistently producing motion sequences that align with stickman representations across diverse scenarios. Our code, demos, and relevant data will be released to support community development.

Abstract:
A visual tokenizer supporting both understanding and reconstruction is the fundamental cornerstone for Unified Multimodal Models (UMMs). Recent works typically decouple the visual tokenizer from Vision-Language Models (VLMs) to optimize it for both tasks. However, the resulting tokenizer fails to support plug-and-play reuse of the VLM's understanding capabilities, demanding costly LLM retraining. This occurs because the decoupled training inevitably disrupts the inherent alignment between the tokenizer and the LLM. To overcome this limitation, we propose JoinTok, a novel LLM-training-free paradigm that jointly optimizes LLM's understanding and pixel-level reconstruction through a dual-projection mechanism. Instead of compromising existing capabilities, JoinTok explicitly preserves the pre-trained VLM's comprehension capacity as an optimization objective while injecting reconstruction skills, ultimately leading to enhanced understanding proficiency. Crucially, we uncover the inherent orthogonality between reconstruction and understanding, providing a theoretical basis for their joint optimization without mutual interference. Extensive experiments show that JoinTok consistently surpasses state-of-the-art methods on comprehensive visual understanding and reconstruction benchmarks.

Abstract:
Existing interactive world models have demonstrated remarkable capabilities in generating visual content via user-driven actions. However, these models are fundamentally constrained to single-view generation, severely limiting their potential for multi-user, multi-agent interactive environments. Transitioning to multi-view interactive generation poses a critical challenge: maintaining strict spatiotemporal synchronization across diverse views. To address this, we present MultiWorld, a novel multi-view world model framework. Unlike previous methods that rely on restrictive explicit geometric constraints (e.g., epipolar lines) or frame-wise synchronizations, MultiWorld constructs a unified sequence-dimension attention space, enabling unrestricted token-level interactions across all views and frames. To explicitly distinguish continuous views within this flattened sequence without introducing parametric bottlenecks, we propose a View-Aware Rotary Positional Encoding (RoPE), which injects a cross-view temporal phase shift. To achieve the stringent parallel output requirements of interactive environments, we design a Parallel View KV Cache architecture, allowing synchronous and consistent multi-view decoding. Importantly, MultiWorld is not tied to a fixed number of views: trained on only two views, it naturally generalizes zero-shot to N-view inference. Extensive experiments demonstrate that MultiWorld achieves state-of-the-art cross-view spatiotemporal consistency and multi-user interactive video generation.

Abstract:
Current multimodal industrial anomaly detection methods face three main challenges: poor early cross-modal interaction, structural information loss during deep feature alignment, and high sensitivity to sensor noise when using simple distance metrics. To solve these problems, we propose the Structure-Aware Geometry-Appearance Network (SAGA-Net), an effective multimodal anomaly detection framework. First, we introduce a unified prior and spatial-geometric filtering module. This module adaptively adds high-frequency structural cues to RGB features for safe early interaction. Second, we design a context-aware synergistic attention mechanism. It combines local spatial context with bidirectional cross-attention to build detailed semantic-geometric connections while keeping structural shapes intact. Finally, we propose a generative residual discriminator. It evaluates joint reconstruction errors to separate real defects from sensor noise, replacing rigid distance scoring. Extensive experiments show that our model achieves 99.2% I-AUROC on MVTec 3D-AD and 91.8% on Real-IAD D^3, outperforming state-of-the-art methods.

Abstract:
While recent range-image generative models have learned strong spatial priors for LiDAR synthesis, forecasting future scans from past observations under ego motion remains an open challenge for autonomous driving: unlike single-frame generation, future prediction must jointly exploit heterogeneous historical evidence and preserve semantically meaningful 3D structure under growing temporal uncertainty. We propose STRIDE, a trajectory-conditioned LiDAR world model in range-image space built on a past-isolated spatial--temporal backbone trained with conditional flow matching, where observed frames serve as deterministic context and future frames are modeled stochastically. Central to our design is Semantic-Stratified Historical Retrieval (SSHR), which leverages historical semantics as a structural prior to partition history into static- and dynamic-dominant branches for targeted fusion, rather than aggregating all cues homogeneously. We further introduce REPA3D, a LiDAR-native representation alignment objective that regularizes intermediate world-model latents against a transferable self-supervised LiDAR prior, promoting semantically coherent 3D structure beyond output-space supervision alone. Across four diverse benchmarks, nuScenes, KITTI Odometry, SemanticKITTI, and KITTI-360, STRIDE establishes new state-of-the-art future prediction quality, temporal consistency, and generation realism.

Abstract:
Partially Relevant Video Retrieval (PRVR) seeks to retrieve untrim-med videos containing a moment that matches a text query, without temporal annotations. The relevant moment may last only seconds within a video spanning several minutes, creating an extremely low signal-to-noise ratio that makes PRVR more challenging than standard full-video retrieval. This task presents two intertwined challenges: (1) \emphsignal dilution, where coarse global representations blur the brief relevant signal into the dominant irrelevant surroundings; (2) \emphcurvature rigidity, where embedding all videos in the same fixed-geometry space distorts representations for videos that range from flat atomic events to deep compositional hierarchies. Existing PRVR methods have improved moment selection and cross-modal matching, but they still typically encode all videos in a single fixed-curvature retrieval space, limiting their ability to model diverse video structures. To address both challenges, we propose Curv\-Spec, a framework that learns content-adaptive curvature for video retrieval representations rather than imposing a fixed geometric prior. Curv\-Spec processes features through parallel Euclidean and hyperbolic attention layers, with independently learned curvatures assigned to the hyperbolic layers, and a content-aware fusion mechanism routes each input to its most suitable geometric regime. To further suppress signal dilution, Curv\-Spec represents each video with semantic centroids whose number is determined by the video's content complexity, projects them onto the learned manifold, and matches each query against its nearest centroid by geodesic distance. Experiments on ActivityNet Captions, TVR, and Charades-STA demonstrate state-of-the-art retrieval performance.

Abstract:
High quality 3D asset generation remains a long-standing challenge in computer graphics, with the core difficulty lying in accurately recovering geometry and appearance that are consistent with a given input under limited single view conditions. However, existing generation paradigms struggle to impose reliable structural and appearance constraints from a single image, often resulting in inconsistencies in geometry and texture. To address these challenges, we propose HiQu3DGen, a consistency-driven framework that guides 3D asset generation through the synthesis of view-consistent multi-view representations. The proposed framework adopts a bidirectional recursive multi-view generation strategy to explicitly strengthen inter-view dependencies during generation, thereby reducing the emergence of mutually contradictory views. Leveraging these structurally consistent multi-view observations, it constructs a robust geometric intermediate and further refines geometry and texture by tightly coupling multi-view appearance cues, ultimately producing high quality textured 3D assets. Extensive quantitative and qualitative evaluations on both simple and complex 3D asset generation benchmarks demonstrate that the proposed framework consistently outperforms existing methods in terms of geometric consistency and texture fidelity.

Abstract:
Balancing scalability and structural integrity in multi-view multi-label feature selection remains a significant challenge. Current anchor-based frameworks typically rely on isolated pairwise graphs. This inherently separates structural modeling from semantic fitting, failing to integrally couple cross-view high-order correlations with sample-label dependencies. To break this bottleneck, we propose a unified scalable framework which strategically migrates high-order structural modeling from the full-sample space to a compact anchor space. Specifically, we first introduce a mask-guided joint learning paradigm with bidirectional information propagation, enabling pseudo-labels to simultaneously incorporate precise supervision and topological structures. Furthermore, a tensor-coupled consensus mechanism employs low-rank constraints to extract latent pure structures and explicitly capture global high-order correlations across views. Building upon this cross-view consensus, a novel scalable hypergraph-bipartite regularization factorizes the hypergraph Laplacian to seamlessly bridge the multi-view structures with the multi-label semantics. Consequently, it captures nonlinear manifolds with linear computational complexity with respect to the sample size, providing reliable structural guidance for accurate feature evaluation. Extensive experiments demonstrate the superiority of our method, validating its effectiveness in characterizing high-order correlations for discriminative feature selection.

Abstract:
Large-scale incomplete multiview clustering (IMVC) is a fundamental yet challenging task in multimedia analysis. Existing methods often suffer from topological distortions caused by missing instances and are typically constrained by cubic computational complexity, rendering them inapplicable to massive datasets. To address these, we propose DTMC, a novel dual-domain optimization framework that emphasizes spatial-domain manifold reshaping. Rather than relying on isolated view-specific structures that are highly vulnerable to data incompleteness, DTMC designs an efficient cross-view voting strategy to mine static, reliable structural priors. By leveraging these mined triplet priors to guide the mean embedding of multiple views, our framework generates dynamic gradient forces that pull intra-cluster samples and push inter-cluster samples. This mechanism effectively rectifies topological distortions in corrupted views, coupled with a lightweight frequency-domain tensor filter to preserve global cross-view low-rankness. Furthermore, we devise a robust optimization strategy that circumvents the construction of memory-intensive quadratic global graphs. Benefiting from vectorized gradient accumulation and dimensionality reduction techniques, DTMC maintains near-linear time and linear space complexities. Extensive experiments on several large-scale incomplete datasets demonstrate that our method significantly outperforms state-of-the-art baselines in both clustering accuracy and computational efficiency.

Abstract:
On-device multimedia retrieval is vital for smartphones, enabling applications like cross-modal semantic search and multimodal personal AI agents. However, realizing efficient retrieval on mobile SoCs remains a critical systems challenge. Unlike text-based search, multimedia applications depend on high-dimensional cross-modal embeddings (often exceeding 1024 dimensions) and continuously expanding media streams (e.g., video lifelogs and ambient audio). Mobile devices must process these intensive workloads under strict latency, energy, and memory constraints. Existing vector retrieval systems, designed primarily for static server datasets, are poorly suited for the dynamic execution models of mobile SoCs, where interactive queries must coexist with continuous background media ingestion and index maintenance. To address these challenges, we propose MUSE, a MUltimedia Search Engine comprehensively co-designed for mobile SoCs. MUSE tackles these issues through an architectureindex co-design. It introduces a three-stage asynchronous execution pipeline that overlaps DMA transfers, NPU matrix computation, and hardware-vector data adaptation to overcome the memory wall of high-dimensional embeddings. Furthermore, it redesigns the Inverted File Index (IVF) and introduces workload-aware heterogeneous scheduling across the CPU, GPU, and NPU for hybrid interactive-and-ingestion workloads. Evaluated on Snapdragon 8-series SoCs using real-world multimodal datasets, MUSE improves query throughput by up to 1.4 at matched recall, achieves up to 7 faster index construction, and delivers up to 6 higher insertion throughput under concurrent streaming. Crucially, its accelerator-native design ensures strict physical compliance, capping peak device temperatures at 38C and reducing total energy consumption by up to 4.6 compared to CPU-bound baselines.

Abstract:
Recent advances in speech language models leverage discrete speech representations from pretrained codecs to enable scalable training and generation. However, existing codecs are primarily optimized for compression without accounting for the autoregressive nature of language model training, resulting in suboptimal performance when modeling compressed speech tokens. In this work, we revisit speech discretization from a generative modeling perspective and propose a novel framework that explicitly aligns speech tokenization with autoregressive training. Our approach introduces autoregressive-compatible constraints during codec training, encouraging token sequences that exhibit temporal consistency and predictability. In addition, we propose a heterogeneous downsampling strategy for different layers of speech tokens, distinguishing semantic from acoustic layers, to improve the alignment between semantic tokens and corresponding textual content. Extensive experiments across multiple benchmarks demonstrate that our method bridges the gap between speech compression and generative modeling, enabling more effective continued pretraining of existing language models on speech data. The approach consistently improves performance across multiple codecs, validating its generality and applicability to diverse speech modeling scenarios.

Abstract:
We investigate descriptive video degradation assessment (VDA) for granular diagnosis beyond the interpretability limits of conventional video quality assessment (VQA). Unlike scalar mean opinion score (MOS) regression, descriptive VDA identifies spatio-temporal degradations and explains their collective perceptual impact, turning black-box judgments into actionable feedback for video enhancement, compression optimization, and AIGC refinement. Adapting multimodal large language models (MLLMs) to this task is hindered by scarce quality-specific video-text supervision and the tension between fine-grained spatial perception and long-range temporal reasoning. We address both challenges with score-anchored semantic synthesis, which uses MOS as semantic anchors for consistent quality-oriented descriptions, and hierarchical spatio-temporal adaptation, which aligns spatial degradation primitives on images before learning temporal quality dynamics from videos. Experiments show more accurate spatial attributions and more stable temporal judgments than prior VQA baselines, demonstrating strong diagnostic utility. A preliminary extension study further suggests that the framework can generalize beyond technical degradations to aesthetic, temporal, and AIGC-specific quality dimensions. Code and data will appear at Descriptive-VDA.

Abstract:
We introduce \method, the first POI recommendation framework that generates \emphcontext-conditioned multimodal representations at recommendation time, rather than relying on static POI embeddings pre-computed independently of context. % Existing multimodal systems encode each POI once as a static embedding, a design that precludes reasoning about why the same caf\'e affords solo work on Monday but group celebration on Friday evening. We formally prove that no pre-computed encoder can satisfy context-sensitive ranking under standard bilinear scoring, motivating inference-time item-side representation. % \method inverts this computation: given a situational context, a four-phase LLM agent generates dynamic, context-specific affordance queries (Phase~1) and executes a five-step cross-modal chain-of-thought over image, review, and metadata evidence (Phase~2). The resulting uncertainty-aware affordance representation is grounded in Gibsonian affordance theory. These cross-modal verdicts form a structured, uncertainty-adjusted affordance representation (Phase~3), which is aligned with user preferences via a semantic caching system for low-latency ranking (Phase~4). % On three POI benchmarks and three evaluation configurations (standard, cold-start, context-shift), \method achieves a 23.2% relative gain over the strongest baseline and degrades by only 7.5% under context-shift versus 16--17% for the strongest baselines.

Abstract:
Recent progress in compressing large-scale 3D Gaussian Splatting (3DGS) data has substantially reduced storage footprint, network transmission bandwidth, and memory traffic to GPU caches before rendering. Yet decoding with advanced 3DGS codecs still takes seconds, making them unsuitable for interactive applications. To systematically address this challenge, we propose SpeedyGS, a Content-Aware 3DGS Compressor that separately optimizes the structural formation and statistical coding. First, in structural formation, we jointly optimize adaptive quantization and pruning under a unified ratedistortion objective, where the rate term is replaced by a lightweight rate proxy that estimates entropy coding cost of the next stage, thereby efficiently regulating Gaussian density and precision to yield a compact scene representation. Then, in the statistical coding phase, Gaussian geometry is converted into sparse octree tokens and subsequently undergoes multi-stage coding, while Gaussian attributes are serialized into a 1D token stream for entropy coding via a complexity?controllable local autoregressive model. SpeedyGS achieves a favorable balance among optimization efficiency, compression performance, decoding latency, and rendering speed. Compared to vanilla 3DGS, SpeedyGS achieves up to 160× model size reduction with negligible quality degradation across common datasets. Compared to state-of-the-art compression methods, it also offers significantly faster decoding and accelerates optimization by 9× on consumer-grade hardware. To further reduce decoding overhead, the statistical coding stage also supports channel-wise, fixed-length coding for Gaussian as a simpler alternative, enabling SpeedyGS to better adapt to the underlying application and reduce decoding latency to nearly zero.

Abstract:
Sign languagetext alignment remains a fundamental challenge for text-driven sign language understanding. Existing methods predominantly rely on appearance-heavy RGB representations, which entangle motion semantics with visual variations and lead to ambiguous motionlanguage grounding. In this paper, we reformulate sign languagetext alignment in a structured kinematic space and propose a kinematics-centric framework that adopts 3D motion parameterized by SMPL-eXpressive (SMPL-X) as the primary representation. By explicitly modeling the kinematic dynamics of signing in a unified motion space, our approach reduces reliance on appearance signals and yields more semantically consistent representations. To capture the compositional nature of sign language, we introduce a gloss-guided local alignment mechanism that leverages gloss temporal spans as automatically estimated training-time supervision to decompose continuous motion into coherent segments and establish fine-grained motiontext correspondences, thereby reducing ambiguity in localizing word-level semantics in continuous signing. Furthermore, we develop a visual distillation strategy, where RGB signals serve as privileged supervision during training to provide complementary contextual cues, and retrieval is performed using only SMPL-X motion and text representations. Extensive experiments on standard benchmarks demonstrate that our method achieves state-of-the-art bidirectional retrieval performance on CSL-Daily and competitive results on PHOENIX-2014T. These results highlight the effectiveness of kinematic representations and explicit local grounding for sign languagetext alignment.

Abstract:
Video conferencing has become an essential medium for modern communication, serving billions of users worldwide in daily interactions. Despite this pervasiveness, the scalability of its underlying infrastructure has received limited attention from the research community, largely due to the lack of a standardized and reproducible benchmarking approach. Realistic evaluations require many participant clients, yet conventional browser-based emulation incurs heavy client-side media processing costs and often makes benchmark hosts themselves the CPU bottleneck before the conferencing systems become saturated. Our empirical analysis shows that video capture/rendering and encoding/decoding dominate endpoint CPU overhead, revealing a critical barrier to large-scale benchmarking. To address this challenge, we present RTCBench, a lightweight and extensible framework for stress-testing large-scale video conferencing systems. RTCBench employs lightweight sender and receiver designs that eliminate unnecessary runtime media processing while preserving the packet-level transport behavior, delivery statistics, and feedback loops needed for realistic conferencing dynamics. Its configuration-driven orchestration and plugin abstraction further support systematic evaluation across heterogeneous conferencing platforms. Our experiments demonstrate that RTCBench substantially outperforms Chromium-based endpoints, sustaining up to 86 more emulated participants on the same hardware and enabling scalable and reproducible benchmarking on modest resources.

Abstract:
In high cognitive-demand scenarios such as learning or social interactions, an individual's emotional responses are not only triggered by external stimuli but also modulated by their cognitive processing. This inherent subjectivity leads to label shift, where identical stimuli may produce significantly different emotion labels across individuals due to variations in their cognitive states. Standard cross-subject domain adaptation methods often overlook this issue; merely aligning physiological signals fails to eliminate such subjective appraisal biases, making generalization to naturalistic interaction scenarios difficult. To address this issue, we propose the Cognitive Probe Network (CP?Net), a cognitive-guided approach that models the subject's cognitive appraisal process by constructing a Gaze?Guided Gate Residual Fusion (GRF) module. This module simulates the subject's selective attention and serves as a dynamic "cognitive probe" to regulate the model's attentional selection: filtering multimodal redundancy while preserving emotional cues, thereby improving both recognition performance and interpretability. To further disentangle subject-related variability, we design an Adversarial Gaze?Emotion Disentanglement (A?GED) module, which orthogonally separates subject?invariant emotional representations from individual cognitive patterns. Additionally, we introduce a bias-adaptive calibration strategy to handle subject-specific label biases. Extensive experimental results under the rigorous LOSO setting demonstrate that CP-Net effectively handles complex appraisal dynamics across the VLMED and MAHNOB-HCI datasets, achieving competitive performance with an accuracy of 45.50% on VLMED and 56.94% on MAHNOB-HCI.

Authors: Tam Nguyen (University of Science, Vietnam National University Ho Chi Minh City), Hoang-An Vo (University of Information Technology, Vietnam National University Ho Chi Minh City), Truong Phu Le (University of Science, Vietnam National University Ho Chi Minh City), Khoa Anh Ta (University of Information Technology, Vietnam National University Ho Chi Minh City), Khanh Tran (University of Information Technology, Vietnam National University Ho Chi Minh City), Duyen Thi My Ngo (University of Science, Vietnam National University Ho Chi Minh City), Bao-Thi Trong Dang (University of Science, Vietnam National University Ho Chi Minh City), Le Hoang Uyen Thu (University of Science, Vietnam National University Ho Chi Minh City), Lizi Liao (Singapore Management University), Binh Nguyen (University of Science, Vietnam National University Ho Chi Minh City)

Abstract:
Automated oversight of medical advertising requires authorization-based compliance checking rather than conventional truthfulness verification. This is especially challenging under cross-modal regulatory evasion, where benign text conceals images of unauthorized procedures. Existing vision-language models are brittle here, often over-relying on the safe modality and exploiting spurious correlations. We present MM-RegCheck, a training-free multimodal framework that extracts and grounds clinical procedures from image-text pairs using retrieved statutory evidence. To enforce compliance, we introduce TF-BEAR, a Bayesian evidence aggregation method. Using evidence-conditioned log-odds attribution, TF-BEAR isolates the regulatory evidence's contribution, enforcing strict liability when hidden visual evidence contradicts the text. We also introduce MedReg-MM, an expert-validated benchmark featuring adversarial evasion subsets. Using a 9B backbone, MM-RegCheck achieves 0.8985 Macro-F1 (Beauty) and 0.6556 (Healthcare), yielding state-of-the-art explanation quality (3.7 FIDES), and demonstrates robust cross-modality within-family generalization. These results highlight inference-time evidence adjudication as a practical foundation for high-stakes compliance monitoring.

Abstract:
Federated adaptation of pre-trained vision-language models offers a promising paradigm for learning from privacy-sensitive distributed data. However, under client heterogeneity, local updates are often biased toward frequent local categories and domain-specific visual patterns, which gradually distort the shared cross-modal semantic space between visual features and text semantics. As a result, the adapted model not only struggles to preserve underrepresented semantic concepts but also exhibits degraded open-vocabulary generalization to unseen classes. To address this issue, we propose FedSA-OV, a semantic-preserving framework designed for heterogeneous clients and open-vocabulary generalization. FedSA-OV employs globally shared text semantic anchors to regularize local representation learning, introduces a frequency-aware mechanism to strengthen semantic preservation for underrepresented classes, and performs semantics-aware aggregation on the server. Furthermore, for native image-text data, FedSA-OV exploits paired text as an auxiliary sample-level semantic signal during local training. Extensive experiments on both image benchmarks instantiated with textual class descriptions and a native image-text benchmark demonstrate that FedSA-OV consistently improves novel-class generalization, low-frequency category recall, and image-text alignment, while maintaining competitive base-class recognition. These results, especially on the native image-text benchmark N24News, highlight that preserving a shared cross-modal semantic structure is essential for robust open-vocabulary adaptation under heterogeneous federated settings.

Abstract:
3D single object tracking is a key technology in robotic vision and intelligent perception systems. In complex scenarios, common challenges such as target absence, occlusion, and deformation often lead to tracking drift or failure. For LiDAR point clouds, their inherent sparsity and texture-less nature often result in insufficient target feature representation, further exacerbating tracking instability in the aforementioned complex scenarios. Although existing Siamese tracking paradigms introduce memory mechanisms to propagate geometric and mask features, they are still confined to pure geometric information and lack the utilization of intrinsic material properties, leading to insufficient discriminability when facing geometrically similar distractors and sparse appearances. To address this, we propose the RP2Track framework, which innovatively integrates LiDAR reflectance as a physical prior into the 3D single object tracking pipeline. It comprises three modules: a Selective Reflectance Fusion Module (SRFM) for adaptive physical-geometric enhancement; a Decoupled Adaptive Propagation Module (DAPM) separating dynamic appearance from static intrinsic maintenance to suppress drift; and a Reflectance-Guided Attention Fusion Module (RAFM) utilizing material semantics to resolve geometric ambiguity. Experimental results on KITTI, NuScenes, and Waymo Open datasets demonstrate that RP2Track significantly enhances tracking robustness under sparse point clouds and achieves competitive performance.

Abstract:
Reinforcement learning (RL) has shown strong potential for improving reasoning in vision-language models, yet its application to medical visual question answering (VQA) remains limited to close-ended tasks with binary verification. Open-ended medical reasoning---where models must produce free-form diagnostic interpretations---poses a reward-design challenge, as valid answers vary in terminology and granularity. We propose Med-OR, a curriculum-driven RL framework that unifies close-ended and open-ended medical VQA training. Med-OR applies a consistency refinement protocol to three public benchmarks, producing 27k training pairs (14k close-ended, 13k open-ended) with tight question--answer alignment. We identify a gradient conflict in joint training: low-variance discrete rewards from close-ended tasks suppress the smoother open-ended signal. To address this, we introduce a two-stage curriculum---Stage~1 builds diagnostic accuracy; Stage~2 refines generative reasoning using the Stage-1 checkpoint as a KL anchor. A lightweight hybrid reward drives training, while independent LLM-as-Judge and blinded expert review serve as evaluation. Experiments on eight benchmarks and four architectures show consistent improvements, with 12.7% in-domain and 6.4% out-of-distribution (OOD) gains over the Qwen2.5-VL-7B baseline. Using only 27k training pairs (3% of leading alternatives), Med-OR achieves competitive results confirmed by blinded expert evaluation (78.6/100), demonstrating that principled curriculum design can compensate for data scale in medical multimodal reasoning.

Abstract:
While Multimodal Large Language Models (MLLMs) exhibit strong generalization, visual instruction tuning for downstream tasks inevitably causes catastrophic forgetting, impairing overall generalization. While existing methods regulate weight updates to reduce forgetting, they overlook the fundamental cross-modal alignment in MLLMs. Based on prior work and our observations, we argue that cross-modal alignment is implicitly captured in the information-compression trajectory. To preserve the alignment flow embedded in the trajectory, we propose LLaVAFlow, an information-theoretic distillation framework. First, we compress the mutual information between the extracted relations and MLLM embeddings, encouraging a learnable module to produce a refined alignment flow that benefits downstream tasks. Second, we maximize the mutual information between the extracted alignment flows of the pretrained and fine-tuned MLLMs, enabling the transfer of compact alignment information. Extensive experiments show that LLaVAFlow is an effective plug-and-play framework that preserves alignment flow and enhances both downstream performance and generalization.

Abstract:
Multimodal fake news detection has attracted increasing attention due to the widespread dissemination and strong misleading potential of online news content. Although existing methods have achieved promising performance, they mainly focus on identifying predictive cues for classification, while paying limited attention to the intrinsic regularities underlying how fake news is fabricated from real news. To address this issue, we propose a novel framework called Misleading Pattern-guided Deviation Modeling (MPDM), which revisits multimodal fake news detection from the perspective of deviation modeling. Specifically, we introduce a deviation modeling branch that is activated only for real samples during training, aiming to capture the intrinsic deviation regularities associated with fake news fabrication through deviation trend distillation and directional adaptation. First, we propose Deviation Variant Generation (DVG), which leverages a large language model to generate deviation variants from original news text under predefined misleading patterns. Since the representation deviations induced by different variants vary substantially in both direction and strength, Deviation Trend Distiller (DTD) is introduced to adaptively aggregate these variants in the embedding space and distill a stable dominant deviation trend. Finally, we design a Low-rank Direction Adapter (LDA) to inject the distilled dominant deviation trend into the fused embedding via directional adaptation, thereby avoiding semantic distortion caused by directly injecting the deviation trend and guiding the classifier to learn a more precise decision boundary. Extensive experiments on four benchmark datasets demonstrate that MPDM consistently outperforms state-of-the-art baselines.

Abstract:
In remote sensing, the continuous emergence of new land-cover categories and evolving application demands necessitate models that can incrementally learn novel concepts from only a few annotations, making few-shot class-incremental learning (FSCIL) essential. However, unlike natural images, remote sensing imagery is inherently multi-label, characterized by severe label sparsity, high inter-class visual similarity, and complex category co-occurrence, which jointly compress the probability space and lead to ambiguous decision boundaries. Despite its importance, FSCIL in multi-label remote sensing remains largely unexplored due to the lack of standardized benchmarks and the substantial domain gap from natural images. To bridge this gap, we introduce MIRA, the first benchmark for multi-label FSCIL in open-world remote sensing. MIRA is built upon a decision-boundary-centric learning paradigm, explicitly targeting the challenges of high spectral similarity and extreme class imbalance. Specifically, we propose an Angular Margin-based Decision Boundary Refinement (AMR) mechanism, which enforces sample-adaptive angular margins to progressively enlarge inter-class separability under unreliable novel-class prototypes. Furthermore, we design an Asymmetric Gradient Rebalancing (AGR) strategy to dynamically balance positive and negative gradients in highly sparse multi-label regimes. Extensive experiments demonstrate that MIRA consistently outperforms existing methods, achieves strong cross-dataset generalization, and maintains stable performance across continual learning sequences.

Abstract:
In healthcare federated learning, clients often rely on lightweight local models for routine, low-cost inference, yet these models can become unreliable on challenging cases when incremental batches of unlabeled data arrive. External annotators could provide higher-quality labeling, but directly transmitting raw images for annotation is often undesirable and costly in practice. Motivated by this, we propose FedHARDA, a Federated Hard-case Cascading Annotation without Raw Data Release framework that selectively invokes an external annotator while avoiding direct transmission of raw patient images. FedHARDA is built on two principles: efficient annotation of difficult cases and surrogate-based external supervision without direct transmission of raw images. Specifically, each client first identifies a minimal hard subset using a disagreement-aware selection strategy that combines divergence-consensus scoring with automated knee-point detection, thereby reducing unnecessary external queries. For each selected hard image, a locally deployed diffusion model generates multiple surrogate candidates. Only candidates that pass the PAC-calibrated retrieval-risk audit and semantic-fidelity filter are released to the external annotator. Returned external labels are used directly for surrogate supervision, without any voting or aggregation. Experiments across multiple medical image classification benchmarks show that FedHARDA consistently improves annotation accuracy while substantially reducing the number of external annotation queries.

Abstract:
Audio-Visual Question Answering Continual Learning (AVQACL) requires a model to acquire new knowledge from evolving audio events, visual concepts, question semantics, and cross-modal reasoning patterns without forgetting previously learned capabilities. Unlike conventional continual learning, knowledge shifts in AVQACL are inherently heterogeneous: different tasks may perturb different functional regions of the model rather than the whole model uniformly. Existing methods, however, typically rely on globally shared regularization or monotonic expansion, making them poorly matched to localized multimodal evolution and prone to redundant growth over long task sequences. We propose Region-Aware Dynamic Capacity Regulation (RDCR), a bidirectional structural adaptation framework for AVQACL. RDCR decomposes the adaptable model space into four functional regions, namely audio perception, visual perception, question semantics, and multimodal fusion reasoning, and performs novelty-driven selective expansion only where new knowledge truly emerges. To capture reasoning changes beyond unimodal variation, RDCR further introduces a lightweight reasoning-aware fusion novelty formulation that jointly models representation-level shift and structure-level shift in the fusion region. Beyond expansion, RDCR performs redundancy-driven regional consolidation to merge overlapping adapters and suppress duplicated knowledge accumulation. We further introduce Split-Valor32k-AVQA, a new dataset that exposes AVQACL models to broader sequential knowledge shifts. Extensive experiments on standard AVQACL benchmarks and Split-Valor32k-AVQA show that RDCR achieves stronger performance with a more controlled growth trajectory than globally uniform and purely expansion-based baselines.

Abstract:
Camera-based 3D object detection has become a cost-effective solution for autonomous driving systems. Although cameras capture rich color and texture information, they lack depth data, making it challenging for detectors to predict 3D positions accurately. We observe that predictions tend to align with camera rays, with many redundant predictions clustering around the same ground truth. This suggests that models are effective at learning shifts perpendicular to the depth axis but less effective along the depth direction. Based on this observation, we propose Ray-Match, a training strategy that leverages the optical properties of cameras to guide models to learn offsets along the depth direction through one-to-many supervision without affecting inference speed. Specifically, we project predicted bounding boxes along camera rays to positions with the same radial distance as the ground truth and then calculate the IoU between them. This ray-aware IoU captures the similarity between the predicted boxes and the ground truth in the camera plane, allowing more potentially useful predictions to contribute to model learning. Our experiments, including detailed ablation studies, demonstrate the effectiveness of Ray-Match. It improves mAP by 1.7 points and NDS by 0.7 points over the state-of-the-art SparseBEV method on the nuScenes dataset.

Abstract:
Adapting video vision-language models (VLMs) is computationally expensive because video inputs produce a large number of visual tokens, making both fine-tuning and inference costly. Although visual token compression can reduce this overhead, direct adaptation on compressed inputs often causes semantic drift and noticeable performance degradation. We present Token-Budget Distillation (TBD), a parameter-efficient fine-tuning framework for adapting video VLMs under a fixed token budget. TBD freezes the pretrained backbone, updates only LoRA adapters, and integrates FlashVID-based visual token compression into the video pathway. To preserve full-token semantics under compression, TBD employs a dual-path teacher-student design, where a full-token teacher provides stable supervision and a compressed student is optimized with task loss, answer-region KL distillation, GT-anchored margin distillation, and reliability-aware KD control. This design enables the student to recover the semantic behavior of the full-token model while remaining efficient under aggressive token reduction. We evaluate TBD on three video VLM backbones, including LLaVA-Video, LLaVA-OneVision, and Qwen3-VL-8B-Instruct, across four video understanding benchmarks. TBD consistently outperforms compression-only baselines under both moderate and aggressive compression. On LLaVA-Video at retention ratio R=10%, TBD preserves 97.0% of the Vanilla model's average accuracy; on LLaVA-OneVision at R=10%, it achieves an average score of 58.4 and matches 100.0% relative accuracy.

Abstract:
While diffusion models achieve impressive image generation capabilities, they suffer from high computational costs, implausible artifacts, and limited adaptability to dynamic or private data due to static, centralized training. To address these, we propose PixARG, a novel retrieval-augmented generative framework that integrates dynamic real-world image retrieval into the diffusion pipeline to simultaneously enhance efficiency and visual realism. Specifically, PixARG constructs a pixel-level affinity map between the user's text prompt and the retrieved image to guide a heterogeneous noise initialization. By modulating noise intensity across spatial regions, this mechanism preserves semantically relevant structural priors while perturbing less relevant areas, thereby significantly accelerating convergence. To accommodate this spatially non-uniform noise and ensure synchronized generation, we introduce a pixel-wise asynchronous denoising scheduler and a spatiotemporal denoiser conditioned on the affinity map. Furthermore, we derive an affinity-adaptive training objective that elegantly balances deterministic restoration with stochastic regeneration across pixels. Experiments on LAION and COYO datasets across offline, online, and user-specific settings show PixARG enhances realism and semantic alignment while drastically reducing diffusion steps.

Abstract:
Enabling navigation agents to follow human instructions and reach target locations in complex environments is a fundamental requirement for embodied intelligence. Despite recent advances in Vision-and-Language Navigation (VLN) that leverage Vision-Language Models (VLMs) for environmental and semantic understanding, two critical limitations persist in multi-floor VLN tasks: (1) token-level representations lack explicit 3D geometric perception, and (2) 2D pre-trained VLMs fail to extract effective visual features from 3D inputs. To address these challenges, we propose ShapeNav, a shape-aware navigation agent that integrates geometric perception while directly utilizing VLMs. First, we construct a Floor-to-Floor (F2F) dataset based on the Matterport simulator for multi-floor VLN tasks and generate corresponding shape-instruction pairs. Second, we introduce a geometric patch extraction strategy to explicitly align spatial patches from input images and pointmaps, achieving cross-domain feature alignment. Specifically, the Masked Optimal Transport (MOT) module aligns shape-aware features with the image modality space, which are then fed into a ViT to derive geometry-enhanced visual features. Third, these features are aligned with shape instructions through Q-Former for geometrylanguage alignment, followed by a graph-based action selector to determine the next step. ShapeNav achieves SOTA performance on the F2F dataset (SR \uparrow 12% and SPL \uparrow 9%), demonstrating the effectiveness of incorporating 3D geometric perception into navigation tasks.

Abstract:
Flow Matching (FM) models enable efficient few-step image generation, making it important to suppress unsafe or copyrighted concepts without retraining. Yet guidance developed for diffusion-aligned settings can degrade image quality and destabilize editing when transferred to practical rectified-flow models. We explain this failure through posterior competition and late-time amplification. In competition-prone regions, branch reweighting governs the concept-conditioned posterior response, so local guidance can incur correction mismatch. The rectified affine path then amplifies this mismatch near terminal time through mean-to-velocity conversion and increasing local sensitivity. Motivated by this analysis, we propose GeoClean, a training-free framework combining Competition-Aware Lookahead Evaluation (CLE) and Amplification-Controlled Correction Smoothing (ACS). CLE stabilizes where the correction is evaluated, while ACS regulates its temporal evolution before injection. Experiments show a 45.9% average ASR reduction while preserving image fidelity, semantic alignment, and few-step efficiency.

Abstract:
EEG-to-image reconstruction aims to recover perceived visual content from non-invasive neural signals, with potential applications in brain--computer interfaces (BCIs). Existing methods typically align EEG embeddings with their corresponding image embeddings by jointly optimizing contrastive and mean squared error (MSE) losses. However, we identify two key limitations of this joint objective when used to learn representations for image reconstruction. First, standard instance-level contrastive learning treats every non-matching EEG--image pair as a negative, overlooking the nested many-to-one structure of the data and pushing semantically related samples apart. Second, our analysis indicates that reconstruction quality is more sensitive to directional alignment than to magnitude matching, whereas MSE can bias optimization toward matching embedding magnitudes. To address these limitations, we propose Concept-aware Semantic Distillation (CSD), a framework for learning EEG representations tailored to image reconstruction. CSD introduces Concept-aware Contrastive Alignment, which uses concept-balanced sampling to reduce competition from false negatives and masked neighborhood regularization to preserve neighborhood relations among non-matching samples. Semantic Direction Alignment further encourages directional agreement between paired EEG and image embeddings, thereby reducing the bias toward magnitude matching. Extensive experiments across multiple datasets demonstrate that CSD achieves competitive results on quantitative metrics while delivering superior visual fidelity. Notably, these gains are achieved without auxiliary alignment networks, highlighting the effectiveness of CSD.

Abstract:
Panoramic image generation is increasingly important for immersive applications such as virtual reality, augmented reality, and 3D content creation. Unlike perspective images, panoramic images represent a viewer-centered 360^\circ surrounding space, where directional expressions such as left, right, front, and behind play a central role in spatial understanding. However, existing text-to-panorama methods largely rely on implicit spatial reasoning and often fail to faithfully ground object-level directional descriptions in spherical panoramic scenes. A straightforward alternative is to introduce explicit layouts, but requiring manually specified spatial conditions reduces the flexibility of language-based interaction and does not directly resolve the misalignment between egocentric directional language and panoramic image space. To address this issue, we propose PanoCtrl, an object-centric framework for controllable text-to-panorama generation. Our method explicitly bridges natural language and spherical panoramic space by converting textual descriptions into structured object-level spherical conditions and integrating them into the diffusion process. Specifically, we introduce PanoParse, a text-conditioned parser that predicts object semantics and spherical bounding field-of-view (BFoV) parameters, and PanoControl, which injects object-level semantic and spatial guidance into the diffusion transformer through object-aware attention and spatial residual enhancement. To support this task, we construct PanoGround, a dataset with object-level spherical annotations and diverse directional descriptions for controllable panoramic generation. Extensive experiments demonstrate that PanoCtrl achieves state-of-the-art performance in both spatial alignment and image quality.

Abstract:
Integrating heterogeneous cancer data, such as histopathological images and genomic profiles, provides a multimodal perspective for assessing patient survival and improves predictive performance. However, existing methods often overlook the complex entanglement among multimodal shared prognostic information (consistency), modality-specific prognostic information (complementarity), and task-irrelevant information (noise), which substantially hinders progress in cancer survival prediction. To address this challenge, we propose a novel multimodal disentangled representation learning framework that effectively separates consistency, complementarity, and noise, thereby mitigating modality heterogeneity and noise interference. Specifically, we propose a task-driven disentanglement module built on a Transformer architecture to decompose multimodal features into consistency, complementarity, and noise representations, while enforcing their distance relationships via multiple regularization losses. To further capture complex intra-modal and cross-modal interactions, we design a disentangled representation graph fusion strategy that constructs consistencycomplementarity graphs and aggregates disentangled representations via graph attention. Quantitative analyses and visual evaluations on five cancer datasets consistently demonstrate the superiority of our method over existing survival prediction methods, highlighting its strong capability to exploit rich survival-related information from multimodal cancer data.

Abstract:
Vision-Language-Action (VLA) models have recently emerged as a promising paradigm for robotic manipulation, in which reliable action prediction critically depends on accurately interpreting and integrating visual observations conditioned on language instructions. Although recent works have sought to enhance the visual capabilities of VLA models, most approaches treat the LLM backbone as a black box, providing limited insight into how visual information is grounded into action generation. Therefore, we perform a systematic analysis of multiple VLA models across different action-generation paradigms and observe that sensitivity to visual tokens progressively decreases in deeper layers during action generation. Motivated by this observation, we propose DeepVision-VLA, built on a Vision-Language Mixture-of-Transformers (VL-MoT) framework. This framework enables shared attention between the vision foundation model and the VLA backbone, injecting multi-level visual features from the vision foundation model into deeper VLA layers to enhance visual representations for precise and complex manipulation. To reinforce critical visual cues without increasing computational overhead, we further introduce Action-Guided Visual Pruning (AGVP), which leverages reliable shallow-layer attention signals to prune irrelevant visual tokens while preserving task-relevant ones. Extensive experiments demonstrate that DeepVision-VLA outperforms prior state-of-the-art methods by 9.0% and 8.1% on simulated and real-world tasks, respectively, providing new insights for the design of visually enhanced VLA models. The project page is available at https://deepvision-vla.github.io.

Authors: Huibin Tan (National University of Defense Technology), Songyuan Yang (National University of Defense Technology), Jianing Yang (National University of Defense Technology), Yize Sui (National University of Defense Technology), Meng Zhang (National University of Defense Technology), Jintao Huang (National University of Defense Technology), Xiaoming Yan (National University of Defense Technology), Ligong Cao (National University of Defense Technology), Nan Hu (National University of Defense Technology), Kedi Zhang (National University of Defense Technology), Xueqiong Li (National University of Defense Technology)

Abstract:
Generating spatially-aware 3D human motions from natural language remains a fundamental challenge, as text inherently lacks precise trajectory control. While recent methods condition generation on explicit 3D scenes to enforce physical grounding, they impose a severe human-computer interaction (HCI) bottleneck, demanding specialized 3D expertise and cumbersome user inputs. To democratize spatial control, we propose a paradigm shift: an interactive, zero-3D multimodal authoring framework named MoViS. MoViS empowers users to dictate complex spatial intent simply by drawing a 2D planar sketch (Visual Prompt) over a scene image, complemented by a text instruction. However, inferring 3D physical geometry from ambiguous 2D sketches presents extreme challenges in perspective scale and long-horizon intent drift. To overcome these, MoViS introduces a Physics-Grounded Cross-Modal Alignment mechanism, uniting 3D landmark distillation and continuous 2D perspective projection to embed true spatial awareness into a discrete generative backbone. Furthermore, we propose a cognitive-inspired Sketch-Driven Compositional Decoding strategy to robustly execute complex, multi-stage tasks. To catalyze this new paradigm, we construct IndoorViS, a large-scale multimodal dataset featuring 2,015 daily tasks across 1,173 diverse scenes. Extensive experiments demonstrate that MoViS strictly maintains the physical fidelity of explicit 3D-aware models while reducing human authoring overhead by over 90%, establishing a new standard for accessible and expressive embodied agent control.

Abstract:
Multimodal learning improves performance by integrating complementary information across modalities, but complex data corruption and noise raise critical concerns about its trustworthiness, especially in safety-critical applications. Despite extensive efforts on this problem, existing methods still suffer from two limitations. They implicitly assume that all samples lie within the predefined classes, even without sufficient evidence for any class, resulting in overconfident predictions. Moreover, they suffer from modality greedy, particularly in dynamic fusion methods. To address these issues, we propose Feasibility-Based Multimodal Learning (FBML), which predicts with class-wise feasibility scores instead of logits. This allows assigning uniformly low scores to all classes for low-quality modality inputs with insufficient evidence to reject any class, thus avoiding overconfident predictions. Meanwhile, such low scores naturally down-weight low-quality modalities, enabling reliable multimodal fusion and prediction without quality-aware weighting, thus avoiding exacerbated modality greedy. Accordingly, FBML introduces low-coverage class-wise feasible subspaces to compute class-wise feasibility scores for modality inputs, whose union covers only part of the feature space. A Low-Discrimination Repulsion mechanism further adaptively pushes low-quality modality inputs with low inter-class score variation away from all subspaces, yielding uniformly low prediction scores. We further theoretically demonstrate that FBML mitigates modality greedy via a performance-independent upper bound on each modalitys gradient magnitude. Extensive experiments demonstrate that our method achieves state-of-the-art results on multiple benchmarks.

Abstract:
Spectral reconstruction from RGB images offers a low-cost alternative to hyperspectral cameras, yet remains a highly ill-posed problem. Existing methods often rely on local, semantic-agnostic cues, which can yield spectrally mixed and structurally inconsistent reconstructions around object boundaries. In order to alleviate cross-instance spectral mixing and improve region-consistent reconstruction, we propose SG-Net, a Semantic Guidance Network that injects instance-level priors to constrain the reconstruction space. Our SG-Net is designed as a single lightweight asymmetric U-Net with dual inputs (RGB image and segmentation mask). In the encoder, the Semantic-Spatial Guidance Module (SSGM) is developed for semanticspatial fusion. Specifically, a Region-Aware Feature Aggregation Module (RFAM) is built to aggregate and broadcast features within each instance, producing region-consistent guidance. In the decoder, the Spectral Reconstruction Module (SRM) introduces Semantic-Spectral Multi-head Self-Attention (SS-MSA) to inject a low-rank semantic bias into spectral attention for long-range inter-band modeling. We further adopt a segmentation-weighted loss to emphasize semantically significant instances. Experiments on NTIRE 2020 and NTIRE 2022 demonstrate state-of-the-art accuracy, while compared with the second-best method, SG-Net reduces MRAE by up to 7.20% while using 29.2% fewer parameters and 33.7% fewer FLOPs.

Abstract:
Composed image retrieval (CIR) aims to retrieve a target image given a reference image and a corresponding textual modification instruction. By combining visual and textual modalities, CIR enables fine-grained visual detail representation and flexible expression of desired changes. However, existing CIR methods typically perform cross-modal fusion in a single order, overlooking the semantic information loss caused by the inherent modality gap. To address this limitation, we propose a Dual-Order Cross-modal fusion (DOC) framework with increment-decrement guidance. DOC decomposes the textual modification instruction into increment (content to be added) and decrement (content to be removed) instructions, which explicitly capture semantic changes and guide subsequent cross-modal fusion. Furthermore, DOC performs cross-modal fusion in two complementary orders. The cross-modal first fusion branch fuses each textual instruction with the visual query before combining them, facilitating fine-grained cross-modal interaction. In contrast, the cross-modal last fusion branch first aggregates all textual instructions and then performs cross-modal fusion once, preserving holistic semantic context while reducing information loss. By modeling these complementary fusion orders, DOC effectively enhances semantic preservation and cross-modal interaction, resulting in more accurate query representations for retrieval. Extensive experiments on three benchmark datasets demonstrate that DOC outperforms existing methods.

Abstract:
Continual test-time adaptation for object detection requires detectors to adapt to evolving unlabeled target streams without sacrificing previously acquired knowledge. While parameter-efficient fine-tuning has become a preferred paradigm, existing methods often overlook critical inter-domain relationships: they either update shared parameters, triggering negative transfer under mismatched domain shifts, or deploy isolated modules that fail to exploit transferable knowledge. Furthermore, these strategies typically employ a fixed adaptation capacity, neglecting the disparate sensitivities of different feature hierarchies to environmental corruptions. To address these limitations, we propose Hierarchical Domain-aware Parameter Adaptation (H-Hub). Specifically, we design a Domain-Decoupled LoRA Hub that leverages LoRAs asymmetric structure to facilitate positive transfer while mitigating inter-domain interference. Additionally, motivated by the disparate attention responses observed across feature hierarchies under domain shifts, we adopt a hierarchical rank allocation strategy to align model capacity with the specific adaptation requirements of different feature levels. Extensive experiments on COCO-C and SHIFT benchmarks demonstrate that H-Hub achieves robust and generalizable performance in continuously changing environments.

Abstract:
Omni-modal Large Language Models (Omni-LLMs) have demonstrated remarkable capabilities in joint video-audio understanding. However, their practical deployment for efficient inference remains challenging due to the substantial computational and latency overhead. The primary difficulty lies in the massive number of tokens generated from multimodal videos, leading to the difficulty of adapting MLLMs to low-resource scenarios. To address this issue, we propose Omni \underlineDynamic \underlineToken \underlineBudget \underlineCompression (OmniDTBC), a novel, training-free framework designed for efficient audio-visual token compression. Specifically, our framework features an efficient token selection mechanism based on determinantal point processes (DPP), formulating the selection process as a maximum a posteriori (MAP) inference problem to optimally balance token importance and diversity. Furthermore, we introduce a dynamic token management strategy that adaptively allocates the token budget across different temporal segments based on informativeness, while seamlessly merging non-salient tokens to reduce redundancy and preserve overall semantic integrity. Extensive experiments on audio-video understanding benchmarks demonstrate that OmniDTBC performs better than other baselines with different token retention ratios.

Abstract:
Reconstructing 3D CT volumes from bi-planar X-ray images is a promising alternative to conventional CT scanning, as it enables volumetric imaging with significantly reduced radiation exposure and hardware cost. However, this problem is highly ill-posed due to severe depth ambiguity and insufficient cross-view constraints in bi-planar projections. Existing methods either rely on direct regression or adopt generative models with weak conditional guidance, leading to limited anatomical consistency and unreliable 3D structures. To address these challenges, we propose ConDX2CT, a multi-dimensional condition-guided diffusion framework for 3D CT reconstruction from bi-planar X-rays. Instead of directly generating full 3D volumes, we reformulate the task as slice-wise latent diffusion, where each CT slice is synthesized under spatial conditioning and then assembled into a volumetric representation. This design significantly reduces computational cost while preserving global anatomical coherence. To effectively bridge 2D observations and 3D structures, we introduce a multi-dimensional conditional modeling strategy. First, a slice-aware cross-view feature alignment module establishes explicit correspondence between X-ray features and target CT slice locations via cross-attention. Second, we propose a ray-consistent back-projection mechanism that constructs a 3D feature volume by directly replicating X-ray observations along projection rays, preserving geometric consistency without information attenuation. Based on this volume, a decoupled cross-attention fusion module is further designed to inject geometry-aware and slice-aware priors into the diffusion process in a disentangled manner. Extensive experiments demonstrate that ConDX2CT achieves competitive performance and produces structurally clearer and more anatomically consistent CT volumes.

Abstract:
Elastic precision quantization enables efficient multi-bit deployment through a single optimization process. However, extending it to Vision Transformers (ViTs) remains challenging due to high optimization costs and unstable activation distributions with extreme outliers, which severely disrupt multi-bit optimization. More importantly, existing elastic post-training quantization methods for ViTs remain largely weight-centric, leaving activation reconstruction underexplored. To address this issue, we propose PEQuant, a framework for error reconstruction at Transformer block inputs. By applying a Stacked Low-Rank Additive Prompt, PEQuant performs activation-level reconstruction without modifying quantized weights while introducing only negligible sequence overhead. To stabilize this reconstruction under low-bit quantization, we introduce a lightweight Artifact-Absorbing Extension Prompt that suppresses severe outliers and high-norm artifacts that would otherwise destabilize prompt learning. Furthermore, to support robust joint multi-bit calibration, PEQuant adopts a cascaded low-rank parameterization with Multi-Precision Activation Fusion. This design allows higher-bit branches to guide lower-bit prompt learning while preventing aggressive low-bit updates from contaminating higher-bit representations. Consequently, PEQuant naturally complements existing weight-space methods, requires no weight optimization, and transfers effectively to downstream tasks without additional fine-tuning. Extensive experiments demonstrate that PEQuant consistently improves the accuracy of various quantized ViTs.

Abstract:
Modality imbalance remains a central challenge in multimodal learning. The dominant modality suppresses weaker encoders, and the resulting model degrades sharply once a modality becomes noisy or absent. We trace the pathology to the fusion module, where two gradient defects arise, \emphMagnitude suppression and \emphDirection entanglement. Since both defects vary with network depth, per-modality rebalancing methods that apply a uniform correction across layers only partially address them. We propose Orchestrating Gradients Hierarchically (OGH), which intervenes at every encoder layer through a single adaptive scalar~\beta. OGH-M applies a geometric magnitude schedule that gives shallow layers the largest gradient amplification; OGH-D attaches lightweight bypass classifiers and replaces the contaminated gradient with a clean, encoder-only signal. An adaptive calibration procedure profiles layer-wise gradient utility periodically and sets~\beta. On seven benchmarks spanning audio-visual, sentiment, and action recognition tasks, OGH ranks first on four and is competitive on the rest, with gains up to +6.98\,pp on KineticsSounds and +5.24\,pp on CREMA-D over ten recent baselines. The per-layer balancing also improves robustness under modality degradation (e.g.\ +8.95\,pp under audio missing) without any robustness-specific training objective.

Abstract:
Multimodal recommender systems leverage heterogeneous content such as images and text to enhance user preference modeling. However, existing methods suffer from two fundamental limitations: (1) they inadequately model the compositional nature of multimodal semantics, often overemphasizing modality-shared information while under-exploring modality-specific and cross-modal synergistic signals; and (2) they remain limited in handling heterogeneous multimodal noise, which is entangled with useful information and further propagates through graph-based interaction modeling. To address these issues, we propose Disentangled and Aligned Structure-aware Graph Learning for Multimodal Recommendation (DASGRec), a unified framework that systematically integrates semantic disentanglement, structure-aware graph learning, and information-theoretic optimization. Specifically, DASGRec first decomposes multimodal representations into shared and modality-specific components, enabling structured modeling of heterogeneous semantics. It then constructs semantic view-specific interaction graphs via affinity estimation and reliability-aware filtering, allowing different semantic components to induce distinct connectivity patterns. To coordinate these structures, we introduce a dependency-based structural alignment mechanism that promotes shared consistency, shared-specific disentanglement, and specific-view non-redundancy. Furthermore, we propose a synergy-oriented modeling module to capture interaction signals that emerge only through cross-modal coupling. Finally, we incorporate an information bottleneck objective to learn compact yet sufficient representations, effectively suppressing redundant and noisy information. Extensive experiments on multiple real-world datasets demonstrate that DASGRec consistently outperforms state-of-the-art baselines.

Abstract:
\beginabstract Neural radiance fields (NeRF) and 3D Gaussian splatting (3DGS) have achieved remarkable novel-view synthesis quality, yet degrade significantly when trained on motion-blurred images. Event cameras provide blur-free temporal signals that can mitigate this degradation. However, prior event-based reconstruction methods either use events only as optimization or pose constraints, or, when constructing an explicit event-derived sharp-color prior, rely on the physics-based Event Double Integration (EDI) model alone. None combines a physics-based EDI target with a learned event-image sharp-color target. To overcome this limitation, we propose DP^2-NeRF, a dual-prior probabilistic framework that fuses EDI, geometry-faithful but noise-sensitive, with a learning-based EICA prior that suppresses noise but may introduce structural distortions. To let the model differentiate supervision quality per pixel, we extend the evidential uncertainty framework to the motion-deblurring setting, propagating uncertainties through both spatial volumetric rendering and temporal blur formation. The resulting uncertainty-aware loss adaptively down-weights unreliable observations, amplifying the benefit of dual-prior fusion. Extensive experiments on synthetic and real-world benchmarks demonstrate that, under matched evaluation conditions, DP^2-NeRF outperforms existing event-enhanced 3D reconstruction methods. \endabstract

Abstract:
Multi-object tracking (MOT) in satellite videos is particularly challenging because targets are tiny and crowded, visual contrast is low, and platform-induced motion severely degrades both appearance and motion cues. Existing approaches commonly depend on fixed association heuristics or predefined graph structures, while rarely performing uncertainty-aware fusion of temporal and spatial information. To address these limitations, we propose Temporal Heterogeneous Graph Structure Learning for Multi-Object Tracking (THG-Track), a unified framework that formulates MOT as a temporal heterogeneous graph structure learning problem. Specifically, the proposed method builds a temporal heterogeneous graph over object nodes and coarse prior nodes, and uses graph structure learning to adaptively select and weight typed intra-frame and one-step temporal edges. Based on this graph, relation-specific message passing is performed together with an entropy-gated temporal--spatial fusion mechanism, enabling more reliable association under ambiguous conditions. In addition, the learned graph decisions are fed back to the feature map to benefit detection and association. Extensive experiments on multiple public satellite-video MOT benchmarks show that THG-Track achieves state-of-the-art results and consistently improves tracking performance across diverse scenes and object categories.

Abstract:
Merging Vision-Language Models (VLMs) with Large Reasoning Models (LRMs) offers a training-efficient route to equip VLMs with advanced reasoning capabilities. However, existing methods apply fusion coefficients based solely on local, per-layer propertiesweight magnitude or cosine similarity entirely ignoring the cross-layer topological interdependencies that constitute the cross-modal feature transfer chain of multimodal models. Disrupting these pathways through indiscriminate merging causes cascading degradation of visual grounding even when individual layer similarities remain high. We propose Activation-Gradient Topology Alignment Merging (AGTAM), a novel framework that makes cross-layer topology a first-class design criterion during VLM--LRM fusion. AGTAM first constructs a \emphCross-Layer Topology Graph (CTG) by computing pairwise activation-gradient correlation coefficients across all transformer layers on a small, label-free calibration corpus. Based on the CTG, AGTAM classifies layers into cross-modal bridge layers and single-modal reasoning layers, assigning conservative fusion coefficients to the former and aggressive coefficients to the latter. A \emphTopology-Consistent Iterative Fusion (TCIF) algorithm then refines these coefficients to preserve the structural integrity of the cross-modal topology while maximising reasoning injection. The entire procedure requires no annotated reasoning data. Extensive experiments on Qwen2.5-VL (3B, 7B, 32B) and InternVL3-8B demonstrate that AGTAM achieves state-of-the-art performance across six VL reasoning benchmarks and three visual perception benchmarks, consistently outperforming existing merging methods while preserving the visual capabilities of the base VLM.

Abstract:
Auxiliary lines are essential for solving complex geometric problems, yet large vision-language models (LVLMs) still struggle to construct them effectively. Recent attempts construct auxiliary lines via code-driven rendering, a strategy that relies on accurate and executable code generation to produce visual renderings of the auxiliary lines for subsequent reasoning. However, in complex solid geometry settings, such a strong dependence on precise specifications substantially limits the robustness of this strategy. Alternatively, we turn to a simpler and more stable solution, representing auxiliary-line constructions as structured textual descriptions. To bridge the gap between textual descriptions and spatial structure, we propose a reinforcement learning framework that enhances diagram-text alignment. The core is a cross-modal reward model that evaluates how well the generated auxiliary-line description matches the ground-truth auxiliary-line diagram. The reward signal drives a GRPO-based RL stage to yield informative auxiliary-line descriptions for the reasoning. To support this study, we develop a scalable data construction pipeline and build AuxSolidMath, a dataset of 3,018 real-exam geometry problems with paired diagrams and aligned textual fields. Based on this framework, we derive GeoVLMath, an LVLM for solving complex solid geometry. Experiments on a dedicated benchmark for auxiliary-line construction as well as widely used geometry reasoning benchmarks demonstrate the effectiveness of GeoVLMath.

Abstract:
Monocular 3D human pose estimation has advanced rapidly, yet severe occlusion remains a major bottleneck. Under occlusion, the input is often not merely noisy: 2D detections can be corrupted by occluders, and monocular depth queried at projected joints can drift to visible surfaces. We propose Uncertainty-driven Evidence Reallocation Network (UER-Net), a multimodal evidence correction framework that reallocates computation toward reliable visual, language, and depth cues before 3D lifting. UER-Net combines (1) Semantic-Guided Anisotropic Sampling (SAS) for re-acquiring semantically consistent visual context, (2) Probabilistic Depth Propagation (PDP) for uncertainty-gated depth correction on the kinematic graph, and (3) Geometry-Aware Orthogonal Fusion (GOF) for depth-consistent cross-modal fusion. Experiments on Human3.6M and MPI-INF-3DHP show strong single-frame performance against competitive baselines, while additional evaluations on 3DPW and 3DPW-Occ further support the robustness of UER-Net in challenging in-the-wild and occluded scenes.

Abstract:
Document dewarping aims to restore flat and readable images from distorted document photographs to facilitate downstream Optical Character Recognition. While this task has attracted significant attention, preserving physical geometric structures remains a persistent challenge. Furthermore, although recent advances have introduced diffusion models to document dewarping, generating high-resolution and pixel-perfect deformation fields for real-world images with complex distortions remains far from straightforward. To address these challenges, we propose Cascaded Coordinate Diffusion, a novel generative model that formulates the unwarping process as a progressive task within the 2D coordinate space. Specifically, at each denoising step, our model progressively constructs the deformation field by first estimating sparse Thin-Plate Spline control points, refining them, and producing a dense coordinate grid for document unwarping.Additionally, we introduce an Intra-step Progressive Feature Alignment mechanism to dynamically warp visual conditions based on preceding predictions, providing precise geometric guidance across these cascaded stages. Furthermore, to explicitly ensure physical orthogonality, we propose a Hierarchical Axis-Alignment Loss that penalizes structural deviations to preserve the global geometric shape and rectilinear integrity. Experiments demonstrate that our method achieves state-of-the-art performance on multiple real-world benchmarks.

Abstract:
Elastic neural networks can adapt their computational footprint at inference time, enabling a single trained model to be deployed across devices with different resource budgets. However, existing approaches to elasticity focus exclusively on parameter count and, as such, address only static memory limitations. Instead, on resource-constrained embedded hardware, dynamic memory (i.e., the working RAM consumed by intermediate activations during inference) is an equally hard constraint and remains difficult to reduce with existing elasticity mechanisms. We argue that static and dynamic memory must be treated as independent optimization axes, and that practical deployment on tiny devices requires elasticity along both dimensions simultaneously. We therefore introduce GaLe (Global Approximate, Local Exact), a transformer block that decouples static and dynamic memory, enabling post-training control over each independently. Building on GaLe, we construct DUET (Dual Elastic Transformers), a family of elastic models that can adapt to a wide range of deployment targets without retraining. We also identify training instabilities specific to the small-network regime and address them through a kernel reordering strategy. We validate our approach through a comprehensive hardware benchmark of existing attention mechanisms across CPUs, GPUs, microcontrollers, and dedicated accelerators to showcase current limitations in terms of latency, performance, and compatibility. We benchmark DUET on image classification on ImageNet-1K, where it achieves a top-1 accuracy of 78.84% while requiring less than 10 MB of peak activation RAM, a 4x reduction over architectures with similar accuracy. Finally, we show that these benefits extend to image generation, where DUET achieves >80% RAM reduction relative to state-of-the-art architectures, while maintaining comparable accuracy.

Abstract:
We present the Surveillance Forgery Image Test Range (SurFITR), a dataset for surveillance-style image forgery detection and localisation, in response to recent advances in open-access image generation models that raise concerns about falsifying visual evidence. Existing forgery models, trained on datasets with full-image synthesis or large manipulated regions in object-centric images, struggle to generalise to surveillance scenarios. This is because tampering in surveillance imagery is typically localised and subtle, occurring in scenes with varied viewpoints, small or occluded subjects, and lower visual quality. To address this gap, SurFITR provides a large collection of forensically valuable imagery generated via a multimodal LLM-powered pipeline, enabling semantically aware, fine-grained editing across diverse surveillance scenes. It contains over 137k tampered images with varying resolutions and edit types, generated using multiple image editing models. Extensive experiments show that existing detectors degrade significantly on SurFITR, while training on SurFITR yields substantial improvements in both in-domain and cross-domain performance. SurFITR is publicly available on GitHub.

Abstract:
Recent advances in neural rendering have enabled high-quality novel view synthesis for single human subjects; however, existing methods remain limited in human-human interaction scenarios involving sustained physical contact and severe cross-occlusion. These conditions introduce identity ambiguity that is not adequately captured by current benchmarks. We present HHID, a human-human interaction dataset designed to evaluate identity-consistent novel view synthesis under sustained contact and cross-occlusion. The dataset consists of synchronized multi-view RGB sequences of two-person interactions captured with four calibrated cameras at 30 FPS, totaling approximately 1500 frames per camera. HHID provides rich identity-aware annotations, including foreground and per-subject instance masks, 2D keypoints, triangulated 3D joints, and temporally consistent SMPL parameters. We further define benchmark tasks and identity-aware evaluation protocols for contact-aware human interaction rendering and evaluate representative state-of-the-art methods. Experimental results demonstrate consistent performance degradation with increasing interaction intensity, revealing fundamental limitations in preserving identity and geometry under occlusion. HHID establishes a diagnostic benchmark for studying identity-consistent neural rendering in human-human interaction scenarios and enables future research on contact-aware dynamic scene understanding.

Abstract:
Existing AI-generated image quality assessment (AGIQA) methods typically either model imagetext alignment with quality-related descriptions or leverage multimodal large language models (MLLMs) for semantic guidance. However, they still face two key limitations. First, high-level semantic reasoning often interacts only weakly with fine-grained local defect perception. Second, MLLM feedback is often overly generic and lenient, which reduces the effectiveness of semantic information. To address these limitations, we propose DeVAGIQA, a deductive evidence-guided visual expert modulation framework for AGIQA. Rather than relying on free-form MLLM descriptions, DeVAGIQA constructs structured deductive quality evidence, transforming coarse semantic judgments into explicit quality priors. Guided by this evidence, four heterogeneous local experts are dynamically modulated to perform targeted local quality analysis, allowing high-level semantic reasoning to directly steer the perception and aggregation of fine-grained defects. In addition, we introduce a frequency-guided local modeling strategy to enhance sensitivity to multi-scale generation artifacts. Extensive experiments on multiple AGIQA benchmarks show that DeVAGIQA consistently surpasses state-of-the-art methods and exhibits strong cross-dataset generalization.

Abstract:
Zero-Shot Composed Image Retrieval (ZS-CIR) aims to retrieve a target image from a query composed of a reference image and a modification text, without relying on manually annotated triplets. Recent ZS-CIR research has increasingly explored explicit inference via Multimodal Large Language Models (MLLMs), where the composed query is converted into a target description for retrieval. However, they still suffer from two limitations: (1) Inadequate structured compositional reasoning, as general-purpose MLLMs struggle to internalize the stepwise decomposition of query-to-target state transitions through inference-time prompting alone. (2) Incomplete preservation of visual invariants, since compressing the composed query into a unimodal textual description inevitably loses crucial reference-specific visual cues. To address these issues, we propose Chain-of-Modification Reasoning with Proxy Ensemble (CoMPE) for ZS-CIR. First, we introduce Multimodal Chain-of-Modification Distillation (MCoMD), which formalizes the query-to-target transition as a structured Chain-of-Modification trajectory, and internalizes this structured reasoning capability into a lightweight MLLM by tuning on our automatically constructed MCoM-88K corpus. Second, we propose Dual-Proxy Ensemble Retrieval (DPER), which complements the textual description with a visual proxy derived from the reference image embedding, thereby preserving reference-specific visual invariants for more faithful retrieval. Extensive experiments on three prominent CIR benchmarks demonstrate that CoMPE consistently outperforms prior ZS-CIR methods with superior performance and competitive efficiency.

Abstract:
As the development of AI-generated contents (AIGC), multi-modal Large Language Models (LLM) struggle to identify generated visual inputs from real ones. Such shortcoming causes vulnerability against visual deceptions, where the models are deceived by generated contents, and the reliability of reasoning processes is jeopardized. Therefore, facing rapidly emerging generative models and diverse data distribution, it is of vital importance to improve LLMs' generalizable reasoning to verify the authenticity of visual inputs against potential deceptions. Inspired by human cognitive processes, we discovered that LLMs exhibit tendency of over-trusting the visual inputs, while injecting skepticism could significantly improve the models visual cognitive capability against visual deceptions. Based on this discovery, we propose Inception, a fully reasoning-based agentic reasoning framework to conduct generalizable authenticity verification by injecting skepticism, where LLMs' reasoning logic is iteratively enhanced between External Skeptic and Internal Skeptic agents. To the best of our knowledge, this is the first fully reasoning-based framework against AIGC visual deceptions. Our approach achieved a large margin of performance improvement over the strongest existing LLM baselines and SOTA performance on AEGIS benchmark.

Abstract:
The growing popularity of 360-degree videos for immersive virtual reality (VR) experiences places substantial demands on network bandwidth and computational resources. Traditional projection methods, such as Equirectangular (ERP) and Cubemap (CMP), suffer from pixel density distortions, leading to inefficient bitrate use and reduced visual quality. More advanced approaches, like Cylindrical Equal-Area (EAP) and Equi-Angular Cubemap (EAC), alleviate some issues but still face challenges with non-uniform pixel distribution. This paper introduces Spherical Fibonacci Lattice Pixelation (SFLP), which constructs pixels via the Voronoi tessellation of the Spherical Fibonacci Lattice, providing uniform pixel density for more efficient compression and higher visual fidelity. SFLP is evaluated against traditional projection methods on the SJTU and UGC360 datasets for both non-learning and learning-based compression algorithms. Rate-distortion curves and BD-rate analysis show SFLP bitrate savings of 10-40% against existing projection methods for the same visual quality. These results suggest that SFLP is a promising projection method for optimizing VR video delivery.

Abstract:
VisionLanguage Models (VLMs) have advanced video action recognition, yet most existing approaches still rely on text as a static semantic cue. In real video data, however, the primary challenge often lies in the noise (dominant backgrounds, scene biases, and temporally irrelevant motions) that obscure discriminative patterns. We build on the mixture-of-experts framework and propose TextFilter, which repositions language from a passive descriptive signal to an active mechanism for semantic-temporal filtering. Rather than solely injecting semantics, TextFilter uses text to suppress misleading visual signals. A text-guided background attenuation module first reduces semantic interference by suppressing background signals. In parallel, text-driven semantic and temporal Mixture of Cross-modal Experts (MoCE) modules leverage fine-grained LLM-generated descriptions to guide expert selection. The semantic branch focuses on object-relevant regions, while the temporal branch emphasizes action-critical motion phases. Together, these components yield cleaner representations, more reliable routing, and stronger videotext alignment. We assess TextFilter across five widely used benchmarks: Kinetics-400/600, UCF-101, HMDB-51, and SSv2. Our TextFilter delivers competitive performance compared to state-of-the-art methods in zero-shot, few-shot, and fully supervised regimes, while remaining highly parameter-efficient.

Abstract:
Layered animation is a common representation in motion graphics and interface design, but authoring high-quality motion remains a manual and iteration-heavy process. Existing multimodal generation systems often target pixel-space video synthesis or full animation-program generation, making them a limited match for workflows that start from an existing visual layer and require lightweight, editable motion control. In this paper, we study \emphtext-controlled layered animation with editable transformation tracks: given a static image layer, a text instruction, and an initial transformation state, the goal is to generate a temporally coherent transformation track that can be further edited in professional animation tools. To support empirical study of this setting, we present \emphTextAE, a benchmark built from 9,585 real animation projects and 75,520 expert-designed layer-wise transformation tracks, with a text-annotated subset of 17,774 layers. TextAE is our primary contribution and provides a foundation for studying this task. We also provide \emphTrack-FM, a flow-matching baseline that jointly predicts full transformation tracks from multimodal conditions while encouraging smooth and loopable motion. Experiments indicate that TextAE provides a useful testbed for controllable layered animation generation and that Track-FM serves as a strong first baseline for producing faithful, smooth, and editable animation effects.

Abstract:
Recent advances in VLM-based GUI agents accelerate mobile automation but also introduce security risks, as agents rely on broad system privileges and long-horizon contextual continuity. Existing attacks, such as overlay manipulation and inference-gap exploitation, target individual steps but overlook a more fundamental vulnerability: the integrity of contextual observations across extended interaction trajectories. In this work, we argue that disrupting contextual coherence is a highly effective and scalable strategy for compromising mobile agents. The key challenge is crafting plausible perturbations that continuously mislead the agent without breaking visual realism. To this end, we propose GrayUI, a systematic visual context attack framework that exploits three dimensions of contextual dependency: 1) spatial context, by embedding adversarial prompts into wallpapers and persistently occupying overlay windows to pollute the visual field; 2) temporal context, by interrupting task trajectories through splash-screen advertisements; and 3) functional context, by confusing button semantics and introducing in-app pop-ups that distort intent-to-action mapping. In this way, by exploiting system-level (e.g., wallpapers) and application-level (e.g., advertisements) visual channels, GrayUI enables comprehensive context deception from the OS background layer to the app interaction layer. Finally, we evaluate GrayUI on two public benchmarks, showing that a joint attack combining all strategies achieves an average attack success rate of 86%.

Abstract:
Multimodal Large Language Models (MLLMs) have shown strong potential in visual understanding, yet they still struggle with chart analysis, a domain that requires both fine-grained perception and effective multi-step reasoning. Existing approaches often fail to precisely capture detailed visual elements and to adapt reasoning depth to questions of varying complexity. To address these challenges, we propose ChartMaster, a holistic framework that advances chart understanding by jointly optimizing data, perception, and reasoning. Our framework is built on three key components. First, we construct ChartVerse, a large-scale synthetic dataset covering diverse chart types, rendering styles, and reasoning levels. Building on this foundation, we introduce a two-stage training paradigm. In the first stage, Multi-Negative Direct Preference Optimization (MNDPO) improves perceptual precision by training the model to distinguish correct answers from carefully constructed hard negatives, namely plausible but incorrect alternatives. In the second stage, Reinforcement Learning with Dynamic Length Reward (DLR) encourages reasoning processes that adapt to task complexity, promoting concise solutions for simple queries while supporting rigorous multi-step reasoning for more complex ones. Extensive experiments on multiple benchmarks show that ChartMaster achieves state-of-the-art performance, outperforming prior chart-specific models and rivaling proprietary systems. These results demonstrate that combining diverse data foundations with targeted optimization of perception and reasoning provides an effective path toward robust chart understanding in MLLMs.

Abstract:
Operating household appliances requires long-horizon planning that is state-dependent and robust to disturbances, yet existing large models fall short, as no sufficiently diverse, task-oriented dataset exists to support such planning. To bridge this gap, we propose MAGE, a scalable data synthesis pipeline that introduces a novel Hierarchical Appliance Graph (HAG) to automatically generate part grounding, long-horizon planning, and closed-loop recovery data from appliance manuals. With MAGE, we build UseAppliance, the first large-scale dataset for manual-grounded appliance manipulation planning, spanning 22 appliance categories with 89K+ part annotations, 53K+ manipulation tasks, and 33K+ closed-loop adjustment steps. Built on UseAppliance, we develop AppliancePlan, an end-to-end model for manual-grounded appliance manipulation planning. On RealAppliance-Bench, AppliancePlan with only 7B parameters achieves over 10 the best baseline on open-loop planning and consistently outperforms state-of-the-art models across all tasks. Real-robot experiments on six household appliances further confirm effective sim-to-real transfer, marking an important step toward general-purpose household robotics.

Abstract:
Electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS) are ideal for practical brain-computer interfaces (BCIs) due to their portability and non-invasiveness. Integrating the temporal superiority of EEG and the spatial superiority of fNIRS can effectively improve the performance of BCIs. However, three-dimensional (3D) tensor-based EEG-fNIRS decoding faces two critical challenges: (1) excessive model parameters and high computational complexity caused by 3D convolutional operations; (2) traditional fusion methods (such as addition or concatenation) struggle to fully exploit complementary information and achieve adaptive fusion. To address these issues, we propose TREF-Net, a triple-branch residual network comprising parallel EEG, fNIRS, and fusion branches that integrates feature- and decision-level fusion. First, we design an Inception-style residual block for efficient 3D brain signal tensor processing, which combines channel splitting with depthwise separable convolutions, achieving multi-scale spatio-temporal feature extraction while reducing the model parameter count by 84%. Second, we propose a mean-max dual pooling weighted fusion block that employs a parameter-free attention mechanism to enhance aligned EEG-fNIRS representations and support adaptive multi-modal fusion. Subject-specific evaluations on the public datasets demonstrate that TREF-Net achieves competitive overall performance compared with current multi-modal methods, with average accuracies of 87.90% on the word generation (WG) task, 83.91% on the motor imagery (MI) task, and 88.17% on the mental arithmetic (MA) task.

Abstract:
Object-level segmentation in dynamic 4D Gaussian scenes remains challenging due to complex motion, occlusions, and ambiguous boundaries. In this paper, we present an efficient learning-free 4D Gaussian segmentation framework that lifts video segmentation masks to 4D spaces, whose core is a two-stage iterative boundary refinement. The first stage is Iterative Gaussian Instance Tracing (IGIT) at the temporal segment level. It progressively refines Gaussian-to-instance probabilities through iterative tracing and extracts corresponding Gaussian point clouds. Thus, it can handle occlusions and preserve the completeness of object structures. The second stage is frame-wise Gaussian Rendering Range Control (RRC), which suppresses highly uncertain Gaussians near object boundaries while retaining their core contributions for more accurate boundaries. Furthermore, a temporal segmentation strategy is proposed for IGIT to balance identity consistency and dynamic awareness. Correlations within each temporal segment enforce strong multi-frame constraints for stable identities, while independence across segments allows identity changes to be captured promptly. Experiments on HyperNeRF and Neu3D datasets demonstrate that our method produces clearer segmented Gaussian point clouds with accurate boundaries and achieves higher efficiency compared to SOTA methods. Source codes will be released soon after the paper acceptance.

Abstract:
Streaming sports video commentary requires knowing when to speak, what to remember, and how to appreciate. Existing streaming methods are either too hasty to determine the appropriate timing or merely react to superficial visual changes, lacking a deep understanding of the sport. In this paper, we present LiveCue, a decoupled, event-aware streaming commentary method that determines when to comment, waits for complex actions to fully unfold before delivering an informed comment, and maintains a dynamic event-aware memory that selectively preserves highlights via adaptive temporal and event decay. Specifically, LiveCue introduces three key designs: (1) a lightweight, learnable gating model that continuously estimates visual activity intensity and detects event completion via adaptive Schmitt-trigger waveform analysis for event-aware control. (2) A reservoir that provides segmented event sequences to the large language model upon triggering while maintaining context for the gating model. (3) An event-aware round-decayed memory with importance-driven compression and dynamic score protection, retaining highlights while gracefully degrading routine moments. Notably, previous research has rarely focused on the crucial dimensions of artistic and emotional expression in artistic sports such as figure skating. To support training and evaluation, we constructed a streaming sports dataset with artistic annotations and proposed SportsCue-3K, the first streaming sports benchmark to jointly evaluate timing judgment, domain reasoning, and artistic understanding. Experiments show that LiveCue outperforms prior streaming models in both timing and commentary quality, while surpassing some existing models at the 7B/8B scale and achieving state-of-the-art performance on some standard video QA benchmarks.

Abstract:
Interactive agents need fine-grained long-video understanding. Current Video-LLMs remain brittle when relevant evidence is subtle or sparse. Fixed frame budgets force rich streams into a few sampled frames and often discard decisive cues. We introduce Video Bench, a benchmark with 2,212 real-world videos and ten tasks spanning content, semantic, and spatio-temporal perception. Its low Detail Index emphasizes sparse, localized evidence. We evaluate 33 systems under multiple-choice and open-ended settings. Models recognize appearance and text better but remain weak at temporal localization, duration, and counting; performance also falls as the required detail becomes finer. We further evaluate "Think-with-Videos" (TwV) agents that resample frames and perform multi-turn reasoning at fixed budgets. They improve fine-grained and temporal perception without more visual tokens and remain robust when frame budgets shrink.

Abstract:
Replay is mainstream in Continual Learning (CL), but strict memory constraints limit the knowledge stored in real samples. Recent dataset distillation-inspired methods synthesize dense-information replay samples via distribution matching between real and synthetic samples, which reveals two limitations: (1) Equal distance minimization between synthetic samples and all real samples neglects the dataset's tree-like hierarchy, forcing synthetic samples to indiscriminately absorb both lower-hierarchical core prototypes and higher-hierarchical noise. (2) Using a fixed CL model for feature extraction forces synthetic samples to overfit the task-specific model, causing severe forgetting during future representation drift. To address this, we propose Text-Anchor guided Riemann Angle Matching (TARAM). By matching the origin-passing geodesic angles of real/synthetic samples' Riemann centroids in hyperbolic space, this method aligns semantic directions without flattening the hierarchical depth to synthesize replay samples, inherently preserving the data's tree-like hierarchy. Since Riemann centroids naturally prioritize lower-hierarchical core prototypes, this angular alignment efficiently transfers crucial knowledge, significantly boosting the synthetic samples' signal-to-noise ratio (SNR). Furthermore, integrating sample synthesis into the CL model's dynamic optimization alongside parameter noise injection generates a diverse set of network states for feature extraction, preventing the synthetic samples from overfitting to a single, static CL model at any specific task stage. Concurrently, task-invariant text anchors provide absolute semantic coordinates, enhancing cross-task generalization of synthetic samples. Ultimately, TARAM synthesizes robust, high-SNR replay samples, effectively mitigating catastrophic forgetting. Extensive experiments on CIFAR-100 and ImageNet-R demonstrate TARAM's superiority over state-of-the-art methods.

Abstract:
Training diagnostic models with image-level labels reduces annotation cost, but the labels do not indicate where diagnostic evidence is located. We introduce Auto-Induced Multimodal Soft Evidence Distillation, which derives local soft supervision directly from class labels. For each class, a large language model produces a set of observable visual descriptors. BiomedCLIP then compares the descriptors with image patches to obtain candidate evidence maps. The maps are ranked by mean activation, high-response support, and spatial compactness, and the top-ranked maps are fused into a soft target. The student is trained jointly with class labels and soft evidence, and the learned maps guide feature aggregation for diagnosis. The language model and BiomedCLIP are used only during training; inference uses the trained student alone. We evaluate the method on BUSI and ISIC2018 in terms of classification, localization, faithfulness, robustness, and efficiency. The selected evidence-routed configuration improves lesion localization and pixel ranking over its non-routed counterpart. On BUSI, deletion tests indicate that its highest-ranked regions have a greater effect on the prediction. Descriptor controls show that class-specific semantics are more useful on fine-grained ISIC2018 than on BUSI.

Abstract:
Enabling gloss-free sign language translation (SLT) in unconstrained real-world environments is a critical step toward practical deployment. Existing progress has been driven primarily by benchmarks collected in visually homogeneous laboratory settings, where simplified scenes substantially reduce the difficulty of cross-modal alignment. As a result, a pronounced domain gap emerges when models are evaluated on newly released in-the-wild benchmarks or deployed in real-world environments. Under such a domain shift, genuine sign semantics are easily confounded by nuisance visual factors, leading models to rely on non-semantic scene cues rather than the fine-grained spatio-temporal dynamics of sign articulations. To address this challenge, we propose Domain Adaptive Contrastive Learning (DACL), a domain-adaptive training framework that promotes robust cross-modal semantic grounding in visually unconstrained environments. Specifically, DACL transfers semantic alignment priors from controlled source domains to in-the-wild target domains, while introducing signer-free pure background samples as an explicit regularizer to suppress nuisance-driven correlations and encourage the disentanglement of scene context from sign semantics. Without introducing architecture-specific modifications, DACL achieves a BLEU-4 score of 21.47 (+19.03) on the in-the-wild CE-CSL and 18.60 (+3.68) on OpenASL, substantially outperforming counterparts trained directly on the target domain.

Abstract:
Hyperspectral pansharpening aims to fuse a high-resolution panchromatic (PAN) image and a low-resolution hyperspectral image (LR-HSI) to reconstruct a high-resolution hyperspectral image (HR-HSI). Recently, pre-trained diffusion models (DMs) have emerged as a powerful backbone in this task. However, existing methods apply spatially uniform guidance strength, failing to account for the heteroscedastic uncertainty inherent in real-world hyperspectral scenes. Furthermore, existing spatial-spectral decomposition models assume spatially invariant spectral factors, failing to capture the spectral variability across pixels. To address these issues, we propose HetDiff, a novel zero-shot uncertainty-guided diffusion framework for hyperspectral pansharpening. Grounded in Bayesian MAP estimation with Jeffreys Prior, we derive a MAP-derived uncertainty loss to train the Heteroscedastic Uncertainty Zero-Shot Network (HUZSNet),which estimates a spatially varying confidence map that dynamically modulates the reverse diffusion sampling process, emphasizing guidance in reliable regions. Moreover, we propose a Variability-Aware Neural Spatial-Spectral Decomposition (VNSSD) module that explicitly addresses spectral variability that existing decomposition models neglect, thereby reducing spectral distortion in HR-HSI reconstruction. Extensive experiments on Chikusei, Pavia, and Washington DC Mall demonstrate that HetDiff consistently outperforms all competing methods across all three benchmarks.

Abstract:
Diffusion models have achieved unprecedented success in text-aligned generation, largely driven by Classifier-Free Guidance (CFG). However, standard CFG operates strictly on instantaneous gradients, omitting the intrinsic curvature of the data manifold. Recent methods like Zigzag-sampling (Z-Sampling) explicitly traverse multi-step forward-backward trajectories to probe this curvature, significantly improving semantic alignment. Yet, these explicit traversals triple the Neural Function Evaluation (NFE) cost and introduce unconstrained truncation errors from off-manifold evaluations, causing cumulative drift from the true marginal distribution. In this paper, we theoretically demonstrate that the explicit zigzag sequence is topologically reducible. We propose Implicit Z-Sampling, rigorously proving that intermediate states can be algebraically annihilated via operator dualities, physically eliminating off-manifold approximation errors. To push sampling efficiency to its theoretical lower bound, we introduce Z^2-Sampling (Zero-cost Zigzag Sampling). Exploiting the Probability Flow ODE's temporal coherence, Z^2-Sampling couples implicit algebraic collapse with a dynamically cached Temporal Semantic Surrogate. This restores the standard 2-NFE baseline without sacrificing semantic exploration. We formally prove via Backward Error Analysis that this discrete collapse inherently synthesizes a directional derivative curvature penalty. Finally, extensive evaluations demonstrate that Z^2-Sampling structurally shatters the performance-efficiency Pareto frontier. We validate its universal applicability across diverse architectures (U-Nets, DiTs) and modalities (image/video), establishing seamless orthogonality with advanced alignment frameworks (AYS, Diffusion-DPO).

Abstract:
Vision-language models exhibit remarkable zero-shot capabilities but suffer significant performance degradation under distribution shifts. While test-time adaptation (TTA) via Low-Rank Adaptation offers a parameter-efficient solution, we identify a fundamental bottleneck in current methods: the reliance on static rank configurations. Because visual inputs inherently possess varying information densities, a fixed rank forces an inevitable optimization compromise, leading to underfitting on complex scenes and overfitting on simple ones. To bridge this gap, we propose Multi-Rank Adaptation (MuRA), a novel framework that dynamically selects and fuses adaptation modules of varying capacities based on token-level visual complexity. MuRA synergizes Multi-Rank Orthogonal Decomposition to provide a superior, knowledge-preserving initialization, and Unified Component Fusion with Continuous Router Updating to sustainably learn semantic-to-rank mappings. Furthermore, we provide rigorous theoretical justifications mathematically proving the necessity and gradient stability of this adaptive mechanism. Crucially, MuRA's dynamic design uniquely thrives at the deepest visual layer, capitalizing on the shortest gradient backpropagation path. Extensive experiments demonstrate that MuRA achieves state-of-the-art accuracy across extensive domain generalization and cross-dataset benchmarks while significantly reducing both computational and memory overhead.

Abstract:
Continual learning (CL) enables a single model to incrementally learn from data streams without forgetting previously acquired knowledge. Recently, Continual Learning with Unlearning (CLU) has emerged as a promising paradigm to enforce the ``right to be forgotten'', allowing data opt-out in CL scenarios. However, existing CLU methods face three major limitations during unlearning: (1) restricted support for multi-granularity unlearning objects (sample, class, task), (2) time-consuming full model fine-tuning, and (3) reliance on costly storage of extra information, which hinder their real-world applicability. To address these issues, we propose CLgeU, a novel CLU framework that fine-tunes less but unlearns better. Specifically, a lightweight unlearning module is introduced into the model's output space to erase memory by directionally shifting the decision boundaries of target data. This is achieved by selectively optimizing the target angle factor derived from the decomposed logits, perturbing the model's classification alignment on unlearned data while keeping others unaffected. Such a design eliminates the need to replay non-target data or historical knowledge. Experiments on public benchmarks show that CLgeU achieves effective and efficient unlearning with broad compatibility across regularization-, replay-, and optimization-based CL strategies, without compromising performance on retained and future tasks.

Abstract:
Source-Free Unsupervised Domain Adaptation (SFUDA) aims to mitigate domain shifts without accessing source data, which is critical for privacy-sensitive medical applications. Existing SFUDA approaches typically rely on static, open-loop pipelines with fixed adaptation procedures. Such rigid designs lack error correction and struggle to cope with the diverse characteristics of unseen target domains. To address this limitation, we propose Search2Adapt, which reformulates SFUDA as a dynamic decision-making problem. Our framework employs four collaborative agents, Perceptor, Strategist, Executor, and Evaluator, each driven by VLMs to analyze target-domain characteristics, design adaptation strategies, execute toolchains, and assess intermediate results. At the core of Search2Adapt lies Collaborative Tree Search (CoTS), which explores a rich action space comprising VFMs and segmentation operators. To enable reliable plan search under the SFUDA settings, we introduce a VLM-driven multimodal and anatomy-aware reward that evaluates pseudo-label quality through geometric feature translation and visual evidence. Furthermore, CoT reflection allows the Strategist to calibrate the Evaluator's evaluation, forming a constructive deliberation process. By storing successful trajectories in a dynamic prior memory, Search2Adapt progressively accelerates adaptation across target samples. Extensive experiments demonstrate that our framework autonomously discovers domain-specific adaptation strategies, produces high-fidelity pseudo-labels, and achieves competitive performance in SFUDA while enabling efficient distillation of lightweight student models. Code is available at GitHub.

Abstract:
\beginabstract Spatial-spectral fusion seeks to reconstruct a high-resolution hyperspectral image from a low-resolution hyperspectral observation and a high-resolution multispectral measurement. Existing methods put the most effort on feature extraction and fusion design, typically learning a deterministic mapping, while lacking explicit constraints on the semantic ambiguity inherent in this ill-posed problem. To address this issue, this paper presents a Language Before Fusion paradigm, which argues that semantic extraction should precede fusion and may reduce ambiguity in the feasible solution space. Based on this view, a Semantic Grounded Fusion (SGF) Network is proposed to inject complementary spatial and spectral semantics into reconstruction. An Implicit Multiplicative Fusion Operator (IMFO) further lifts local interactions into a solution space, admitting a semantics-conditioned functional interpretation of fusion. For evaluation, a Vision-Language-Guided Fusion Dataset (VFD) is also elaborated. Experiments on CAVE, Harvard, Chikusei, and Pavia Center show consistent gains in both spatial fidelity and spectral preservation. Compared to the latest method in CAVE × 4, SGF improves PSNR by up to 1 dB and reduces SAM and ERGAS by up to 11.62% and 11.69%. \endabstract

Abstract:
Most multimodal sentiment analysis algorithms focus on point-wise regression. While straightforward, it uses absolute values to represent subjective sentiment and overlooks the ordinal nature of sentiment, resulting in unstable predictions and poor correlation. Inspired by the way humans perceive sentiment, pair-wise ordinal learning frameworks emerged to address this gap, capturing relative order by learning from comparisons. However, they have two limitations: First, they assign uniform importance to all pairs, failing to adaptively focus on hard-to-rank samples. Second, they employ static ranking margins, which fail to reflect the varying semantic distances between sentiment groups. To address this, we propose a Group-wise Ranking and Calibration Framework (GRCF) that introduces the Adaptive Error-weighting and Dynamic Margin mechanism, which resolves these limitations by simultaneously preserving the relative ordinal structure and ensuring absolute score calibration. Specifically, Stage 1 introduces a Group-aware Ranking Loss to build a fine-grained ordinal structure by assigning higher optimization priority to confusable pairs. Stage 2 then employs an MAE-driven objective to align the prediction magnitudes. To validate its generalizability, we extend GRCF to classification tasks, including Multimodal Humor and Sarcasm Detection. It achieves state-of-the-art performance on regression datasets while also showing strong generalizability in classification tasks.

Abstract:
Unsupervised shape abstraction is a necessary pathway towards comprehensive semantic structure analysis for arbitrary 3D shapes, yet existing approaches only focus on the visual attribute of structures with shape primitives, while ignoring the conceptual, linguistic characteristic. This paper formulate the text-promptable unsupervised shape abstraction by inter-connecting the two aspects of shape structures. As a result, our new work enables local shape abstraction based on conceptual text prompts. To bridge text-prompts with shape primitives, we leverage language-image-pretraining models, which align text with semantic-meaningful regions in multi-view shape images, and further learn the association between the image regions and shape primitives from unsupervised abstraction. In particular, we formulate the unification loss design that couples image-primitive alignment with global-local relation alignment, whose goal is to acquire image-primitive associations in an embedding space via a newly-proposed contrastive-semantic-region-structural-geometry learning (CerGe) mechanism. To combat the lack-of-data challenge, we design a graph-based data engine to generate image-primitive pairs from raw shapes and their primitive abstractions at scale, using primitive adjacency to localize coherent image regions. Based on the newly-proposed CerGe, we further propose TypingCuboids, the first framework in the new problem setting, which disentangles global abstraction and text-based progressive local matching, to effectively reveal the global-local structural relations. We conduct extensive evaluations to manifest the superiority of TypingCuboids over traditional shape abstraction methods in both usability and accuracy. Our implementation can be found in the supplementary material.

Abstract:
Multimodal Sentiment Analysis (MSA) aims to understand human intent by jointly modeling heterogeneous signals such as text and audio. While recent cross-modal Transformer-based approaches have advanced modality fusion, they inherently suffer from blindness in feature capturing when handling the heterogeneity of multimodal sequences, leading to severe noise issues. Specifically, the alignment process without explicit constraints is highly susceptible to alignment drift, causing models to blindly capture temporally irrelevant segments across modalities. Furthermore, standard fully-connected attention mechanisms enforce global traversal interactions, compelling emotion-irrelevant elements to passively absorb massive amounts of cross-modal redundant information. The substantial noise generated by this semantically indiscriminate interaction severely obscures and dilutes genuine sentiment cues. To break this bottleneck, we propose \modelname, a novel Temporal-Semantic Dual-Targeting Fine-Grained framework, which fundamentally suppresses noise through token-level precise control mechanisms. Specifically, \modelname~incorporates two orthogonal targeting innovations. First, Temporal Alignment Targeting introduces an explicit alignment supervision strategy that utilizes word-level timestamps to construct token-to-frame alignment penalties, forcing the model to precisely focus only on the corresponding temporal windows and eliminate alignment drift noise. Second, Semantic Gating Targeting introduces a query-dependent dynamic gating mechanism to selectively regulate cross-modal information flows based on the intrinsic semantics of the current token, avoiding meaningless global traversal interactions and achieving a purified fusion of sentiment cues. Extensive experiments conducted on three benchmark datasets (CMU-MOSI, CMU-MOSEI, and CH-SIMS) indicate that \modelname~consistently outperforms prior methods in the text-audio bimodal setting, exhibiting 2.13%, 1.64%, and 3.94% absolute improvements in Acc-5 over the prior state-of-the-art on each dataset, respectively, which demonstrates the effectiveness and generalization of our proposed framework.

Abstract:
\beginabstract Real-world image super-resolution (Real-ISR) aims to reconstruct high-quality (HQ) images from low-quality (LQ) inputs subject to diverse real-world degradations. Recent advances have leveraged the LQ inputs and natural image priors learned by Stable Diffusion models to achieve impressive results. However, existing methods often overlook insufficient clarity of LQ inputs inevitably induce content drift in the generated HQ images. This manifests primarily as visual detail degradation and textual semantic shift, severely compromising both fidelity and perceptual quality. To address this challenge, we propose FSP-Diff, a novel one-step diffusion model featuring a dual-pathway architecture. This architecture comprises a Detail-Conditioned Pathway for injecting structured details to recover fine structures, and a Detail-Modulated Semantic Pathway that refines semantic guidance using structured details to mitigate semantic deviations. Extensive experiments on standard Real-ISR benchmarks demonstrate that FSP-Diff surpasses existing one-step diffusion methods in both quantitative and qualitative metrics. \endabstract

Abstract:
Integrating radiology, whole-slide pathology, and structured clinical data is essential for reliable cancer prediction, yet remains challenging due to substantial cross-modal heterogeneity. Existing multimodal models often fail to establish consistent cross-modal representations and provide limited semantic interpretability, making their predictions difficult to trust in clinical settings. To address these challenges, we propose CGFusion, an interpretable multimodal framework that jointly models representation alignment and concept-level semantic grounding. First, a patient-centric contrastive module explicitly aligns CT and WSI representations at the patient level, reducing semantic inconsistency between macroscopic radiology and microscopic pathology. Second, structured clinical variables are projected to initialize the latent queries of a Perceiver-based fusion network, enabling patient-specific priors to guide cross-modal interaction from the outset. Third, an LLM-guided concept grounding module anchors fused latent tokens to clinically meaningful prototypes through temperature-scaled similarity, token-dependent abstention, and hierarchical regularization. Experiments on NLST and MMIST ccRCC demonstrate that CGFusion achieves strong overall performance on NLST and delivers more balanced predictions than competitive baselines on MMIST under severe class imbalance. These results suggest that explicitly coupling multimodal alignment with concept-level grounding is a promising direction for building reliable and interpretable multimodal medical systems.

Abstract:
Text-based localization in 3D point clouds aims to ground free-form natural-language descriptions by reasoning over object geometry and spatial relations. Existing text-to-localization methods have shown promising performance, but they remain limited by ambiguous template-based supervision and insufficient modeling of explicit geometric relations in 3D visual representations. To address these challenges, we propose ReasonLoc, a unified spatial reasoning framework that strengthens supervision, representation, and geometric language understanding for text-guided 3D localization. Specifically, we first quantify and reduce spatial ambiguity by extending directional supervision from 5 to 9 categories, leading to an updated benchmark (KITT360-9). Second, we introduce a Spatial Information Module (SIM) that constructs 5D geometric relation graphs and injects them into a query-conditioned Relation Reranker for coarse retrieval, followed by a dual-channel geometric fusion network for fine-grained coordinate regression. Third, we employ a large language model (LLM) as a constrained spatial reasoner to move beyond rigid templates, which converts free-form descriptions into structured directional and geometric cues. Finally, we establish a new free-form evaluation setting via top-candidate differencing to assess robustness under natural language variation. Extensive experiments demonstrate that ReasonLoc improves geometric-language grounding in 3D point clouds, achieving up to 31% gains in Top-1 retrieval and 33% gains in Top-1 localization over existing methods, while maintaining competitive zero-shot transfer on Toronto3D and Paris-Lille.

Abstract:
Monocular 3D hand reconstruction is intrinsically a geometric problem, yet RGB appearance features alone often struggle to resolve severe ambiguities caused by self-occlusions and hand-object interactions. While introducing depth can explicitly provide spatial cues, raw sensor-captured depth maps are extensively noisy and incomplete, limiting their usefulness for fine-grained hand reconstruction. To bridge this gap, we propose GeoHand, a novel framework that unlocks high-quality geometric priors from a frozen foundational monocular geometry estimator (MoGe2). Recognizing that these priors are oriented toward general scenes, we introduce a map-level GeoAdapter to recalibrate the spatial features, specifically adapting them for detailed hand reconstruction. Furthermore, to systematically integrate these adapted priors without overwhelming intrinsic RGB appearance cues, we employ a gated cross-modal token fusion strategy. Finally, to secure precise local articulation, we design a Keypoint-Queried Iterative Refiner (KQIR) that uses projected joint locations to query geometry-aware image features for spatial correction. By combining global geometric disambiguation with local refinement in a unified pipeline, GeoHand achieves state-of-the-art performance or highly competitive performance on FreiHAND, DexYCB, and HO3Dv3, especially under severe occlusions and hand-object interactions.

Abstract:
Existing active reconstruction methods often suffer from structural holes in 3D Gaussian Splatting (3DGS) due to the over-expansion and geometric distortion of Gaussians. To address this issue, we propose GCS-Splat, an active reconstruction framework focused on the intrinsic geometric quality of Gaussians. Our approach tightly couples intrinsic Gaussians geometric repair with active motion viewpoint planning. At the core of this framework is the Geometric Consistency Score (GCS), which quantifies Gaussian distortion to evaluate local reconstruction uncertainty. Guided by the GCS, we perform probabilistic resampling of historical frames based on geometric correlations, and direct the pruning, splitting, and cloning of Gaussians. For motion viewpoint planning, we design a two-stage coverage-repair planner: it first rapidly achieves spatial coverage through autonomous exploration of global frontiers, and subsequently conducts multi-view repair and path planning targeted at geometrically defective regions under GCS guidance. Furthermore, all generated trajectories are smoothed using B-splines to ensure superior kinematic performance. Extensive experiments on the Replica and MP3D datasets demonstrate that our method achieves favorable performance in both geometric accuracy and reconstruction quality, unifying efficient global exploration with high-fidelity reconstruction.

Abstract:
Recent advances in 3D Gaussian splatting (3DGS) have become a prominent representation for novel view synthesis. However, its application to progressive level-of-detail (LOD) modeling presents a significant challenge, due to the intrinsic irregularity of Gaussian primitives. Sparse voxel rasterization (SVRaster) alleviates this structural limitation by organizing primitives onto a regular voxel grid, offering a hierarchical parent-child relationship for progressive LOD construction. In this paper, we propose a progressive layered LOD representation learning framework, LSVRaster, on the basis of SVRaster. Specifically, we introduce a residual-based progressive LOD representation to learn an incremental and learnable residual on top of the coarser LOD levels. A view-driven progressive reconstruction scheme further assigns voxel detail according to dynamic visibility, enabling bandwidth-aware scene delivery with the minimal transition overhead. For compact appearance modeling, we propose a frequency-aware color decomposition that separates scene appearance into a view-independent component and a view-dependent high-frequency residual predicted by a lightweight MLP. A gradient-guided importance selection mechanism adaptively activates the high-frequency terms only for perceptually salient voxels, thus reducing the model size while preserving fidelity. Extensive experiments show that, by optimizing the representation of low-frequency view-independent color signals, LSVRaster achieves a rendering quality comparable to existing LOD approaches while requiring only 49% of their average model size. Furthermore, our progressive rendering mechanism enables an efficient scale switching with an average transition overhead of only 30 MB.

Abstract:
Recent advances in 4D Gaussian Splatting (4DGS) enable high-fidelity, real-time spatiotemporal rendering, but expose a fundamental trade-off between motion expressiveness and storage efficiency. While anchor-based designs achieve compactness through anchor-level parameter sharing, their rigid uniform parametrization enforces fixed Neural Gaussian counts and feature budgets per anchor. Consequently, insufficient fidelity is addressed by excessive anchor density, rather than lightweight, targeted increases in Neural Gaussian count or feature capacity, resulting in memory waste. To overcome this rigidity, we introduce an adaptive-capacity anchor-based framework that dynamically allocates the representational capacity based on local spatiotemporal demands. Adaptive Anchor Cardinality varies the number of Neural Gaussians per anchor, concentrating primitives in regions of high geometric or motion complexity while suppressing redundancy. In parallel, Adaptive Anchor Feature Masking modulates anchor-level feature channels, assigning rich features to complex regions and lightweight representations to simpler ones. Experiments on MPEG, Panoptic Sports, and N3DV datasets demonstrate substantial storage reduction without degrading visual quality. Notably, on challenging MPEG sequences with complex motion, our method achieves up to 1.5 higher compression than state-of-the-art anchor-based methods while preserving comparable quality.

Abstract:
Building Damage Assessment (BDA) aims to identify and classify the severity of structural destruction from satellite imagery. Recently, computer vision technologies have achieved promising performance in BDA with strong feature representation and comparison capabilities. However, we observe that prevailing approaches often underperform in identifying critical damage-related information by ambiguous judgment of superficial semantics. Through in-depth analysis, we identify that this stems from confusion problems (among different damage levels, and between contextual backgrounds and damage semantics), which consequently lead to reliability issues. To tackle this, we represent the first attempt to investigate data imbalance in BDA, thereby, constructing a Trusted-BDA benchmark with damage entropy metric that focuses on informative damaged regions. In addition, we propose a novel framework named Hyperbolic Semantic-Consistent Contrastive Learning (HSCCL), which contains two innovative components: Hyperbolic Hierarchical Contrastive Learning (HHCL) and Iterative Semantic-Consistent Mechanism (ISCM). Specifically, the first HHCL structures the continuous spectrum of damage severity in hyperbolic representation learning, and the second module as ISCM decouples background context from damage semantics through iterative hard negative mining. Extensive experiments verify that our proposed method achieves superior performance on the Trusted-BDA benchmark, consistently outperforming state-of-the-art imbalanced learning strategies with an improvement of up to 6.27%.

Abstract:
While GUI agents have made significant progress in web navigation and basic operating system tasks, their capabilities in professional creative workflows remain largely underexplored. To bridge this gap, we introduce Cutverse, a benchmark designed to systematically evaluate autonomous GUI agents in realistic media post-production environments. We curate expert demonstrations across 7 professional applications (e.g., Premiere Pro, Photoshop), covering 186 complex, long-horizon tasks grounded in authentic editing workflows, involving dense multimodal interfaces and tightly coupled interaction sequences. To support scalable evaluation, we develop a lightweight parser that transforms raw screen recordings and low-level interaction logs into structured, compositional GUI action trajectories with precise grounding. Extensive evaluations reveal that existing agents achieve only 36.0% task success on realistic media editing tasks, underscoring the challenges posed by complex, long-horizon media post-production workflows in our benchmark.While current models demonstrate promising spatial grounding, multimodal alignment, and coordinated action execution, they remain limited in long-horizon reliability and domain-specific planning.

Authors: Yifan Wang (Institute of Automation, Chinese Academy of Sciences), Yonghuan Li (Department of Engineering Mechanics, Zhejiang University), Suiwu Zheng (Institute of Automation, Chinese Academy of Sciences), Qianfan Zhao (Hangzhou Star Species Robotics Co., Ltd.), Youze Zhao (Institute of Automation, Chinese Academy of Sciences), Ruiqi Liu (Institute of Automation, Chinese Academy of Sciences), Boyi Sun (Institute of Automation, Chinese Academy of Sciences), Renrui Zhang (Institute of Automation, Chinese Academy of Sciences), Kai Liu (Hangzhou Star Species Robotics Co., Ltd.), Xu Yang (Institute of Automation, Chinese Academy of Sciences)

Abstract:
Autonomous task discovery is crucial for indoor cleaning robots operating in interactive, dynamic environments. A promising approach is to learn the distribution of normal scenes and identify intervention targets through anomaly detection. However, embodied anomaly perception differs from conventional static anomaly detection. Robots often need to actively explore and interact with the environment to inspect anomalies, while non-stationary scenes may also undergo legitimate changes, rendering methods that assume fixed viewpoints or visual templates unreliable. To address this challenge, we propose an embodied scene anomaly detection (SAD) framework based on a world model that captures environmental dynamics and naturally leverages temporal and spatial context. When trained exclusively on normal trajectories, the world model exhibits a strong inductive bias toward the normal manifold, and prediction discrepancies provide a principled anomaly signal. Additionally, to improve data exploitation and further strengthen normal priors, we introduce a semantic slot-oriented Factorized Object-Centric Memory that explicitly disentangles dynamics operators from static texture. Finally, we extend our framework from scene anomaly discovery to physical manipulation. Detected targets are converted into low-level language instructions by an MLLM augmented with the normal counterpart predicted by the world model, which a fine-tuned VLA policy will execute. Extensive experiments demonstrate the effectiveness of our approach on a newly built simulation benchmark and real-world cleaning tasks.

Abstract:
Many online applications require users to upload ID document images for remote verification. With the rapid advancement of image editing models, digital ID images can now be manipulated with minimal effort, making forged IDs a serious security threat. However, the lack of diverse and high-fidelity forged ID datasets has limited the development and fair evaluation of detection methods. To address this gap, we introduce MFID (Multi-type Forged Identity Document), a comprehensive benchmark covering diverse tampering types and supporting generalization settings with unseen document layouts and manipulation types. Based on this resource, we propose IDForenLLM, a hierarchical forensic framework that mimics coarse-to-fine human inspection. The model first performs macro-level analysis to generate structured region hypotheses with explicit evidence or verification proposals, and then routes suspicious regions to specialized micro-level models for collaborative analysis to capture fine-grained tampering artifacts. The final predictions are obtained by aggregating macro-level detections and refined micro-level results. The combined evidence enables precise forgery localization. Extensive experiments on MFID and additional benchmarks demonstrate consistent improvements.

Abstract:
Video super-resolution (VSR) aims to recover high-fidelity high-resolution videos from low-resolution inputs and is central to applications ranging from mobile capture to streaming and archival restoration. Existing approaches trade off among local-detail fidelity, long-range spatio-temporal modeling, perceptual realism and efficiency: convolutional alignment techniques preserve local structure but suffer when motion is large or degradations are complex; transformer-based methods capture long-range dependencies yet require architectural or algorithmic adaptations to be computationally feasible; and recent latent or diffusion-based generators synthesize rich texture but demand specialized temporal constraints to maintain coherence. In this work, we present MotionCraft, a controllable VSR framework that formulates restoration as motion-aware latent state prediction inspired by world models, and integrates adaptive sparse attention with an explicit, user-accessible control interface. MotionCraft combines robust motion fusion, a Latent World Transformer that balances locality and targeted non-local interactions, and a compact conditional decoder to deliver temporally consistent, high-quality reconstructions under streaming constraints. Empirical evaluations demonstrate that our design attains strong reconstruction and perceptual performance while enabling predictable trade-offs between temporal smoothness and fidelity.

Abstract:
Professional photo finishing relies on both global adjustments and region-specific local edits guided by semantic masks, yet current automated methods handle this workflow only partially. We present PrismGPT, a Vision--Language Model (VLM) framework that produces structured, region-aware editing plans from a single input image without relying on commercial black-box tools. Training a VLM to simultaneously diagnose aesthetic deficiencies at both global and local levels while predicting precise editing parameters is challenging due to the vast combinatorial decision space. We address this through \emphproxy-guided learning: two simpler proxy tasks---operation decomposition and region-aware aesthetic ranking---teach the foundational skills the model needs, while a \emphcompetence-based dynamic scheduler automatically rebalances the multi-task training ratio, progressively shifting emphasis from the proxy tasks to the primary editing task as each skill is mastered. Crucially, all reasoning traces used for supervised fine-tuning are \emphself-synthesized by the same base model, eliminating the need for a stronger external teacher. Experiments on MIT-Adobe FiveK and SPIRE, a new professionally retouched benchmark we introduce, show that PrismGPT achieves state-of-the-art results while using only ～6% of the training data compared to the previous best method.

Abstract:
Instruction-based image editing (IIE) aims to modify an image according to a natural language instruction. Despite recent advances in diffusion transformers, existing methods often introduce unintended changes to regions unrelated to the requested edit. We attribute this limitation to the absence of an explicit mechanism for edit localization. Different editing operations (e.g., subject addition, removal, and replacement) induce distinct spatial patterns, yet existing IIE models typically perform localization in a task-agnostic manner. To address this limitation, we propose a training-free, task-aware edit localization framework that exploits the intrinsic source and target image streams of IIE models. For each image stream, we construct feature centroids from attention-based edit cues, and then partition tokens into edit and non-edit regions based on feature similarity. Observing that effective localization is inherently task-dependent, we introduce a unified mask construction strategy that selectively leverages the source and target streams according to the editing task. We also provide a systematic analysis of our underlying insights and design choices. Extensive experiments on EdiVal-Bench demonstrate that our framework consistently improves content consistency in non-edit regions while maintaining strong instruction-following performance on top of Step1X-Edit and Qwen-Image-Edit.

Abstract:
Editing compressed images at a semantic level typically requires three heavyweight steps: decoding the bitstream to obtain a reconstructed image, applying computer vision operations such as detection, segmentation, or inpainting in the pixel domain, and re-encoding the result to produce an updated bitstream. This pipeline makes semantic editing computationally expensive and impractical at the decoder side. To address this limitation, we present the first INR-based image compression framework that produces a semantic-embedded bitstream, enabling semantic-level editing directly in the bitstream through only lightweight operations. Object attributes, including color, position, and occlusion, can be modified at will without any decoding, re-encoding, or pixel-domain processing. Our encoder extracts and explicitly embeds structured semantic information into the bitstream at predefined locations, enabling direct access and modification. Both the original and any edited bitstream are decoded through a single, uniform decoding pipeline. Furthermore, we introduce a Latent Modulation Transform conditioned on color embeddings that enhances overfitting capability and editing flexibility. Experiments demonstrate that our method achieves competitive rate--distortion performance compared to state-of-the-art codecs while providing unique bitstream-level editing functionality. The framework is also naturally compatible with downstream machine-vision tasks, as object-level semantics are efficiently encoded in the bitstream.

Abstract:
Image forensics has become increasingly challenging in the era of heterogeneous manipulations, where editing, AI-generated content, and deepfake synthesis expose highly diverse forensic cues across tracks and samples. Existing methods predominantly follow a static one-shot paradigm, in which all inputs are processed with the same computation graph and evidence pathway. Such a design is increasingly restrictive for generalizable forensics, where different samples and forgery tracks often require different types and amounts of evidence. In this work, we propose SAFER (Sequential Adaptive Forensic Evidence Reasoner), a prior-guided forensic agent that reformulates forgery analysis as a budget-aware sequential evidence reasoning problem. Given an input image, SAFER first derives an initial forensic prior using a lightweight multi-scale front-end. Conditioned on this prior, a policy network then adaptively queries one or more track-aware experts from a heterogeneous expert pool under a limited computation budget. The queried evidence is subsequently integrated by a reliability-aware decision aggregation module to produce the final forgery prediction, tampering mask, and explainable report. To make sequential decision-making trainable, we further introduce a progressive policy learning strategy that combines imitation learning and reinforcement learning. Extensive experiments show that SAFER achieves robust performance across diverse scenarios, while providing a more favorable accuracy--efficiency trade-off than conventional static pipelines. Our results suggest that adaptive sequential evidence reasoning offers a practical new paradigm for real-world multimedia forensics.

Abstract:
With the rapid advancement of generative models, manipulated images and text have become increasingly prevalent, posing serious threats to information security. Existing media manipulation detection methods exhibit two key limitations: (1) they fail to capture modality-specific forgery trace. (2) they struggle to model the cross-modal interactions between tampered visual regions and manipulated text fragments. These result in degraded performance in predicting forgery types and grounding manipulated regions. To address these challenges, we propose Fiord, a fine-grained visual and text forgery distillation framework for multimodal media manipulation detection. Built upon a pretrained multimodal encoder, Fiord introduces a visual regional forgery distillation module and a text fragment forgery distillation module to capture fine-grained forgery traces. The former constructs a cross-domain discriminative space and leverages multi-scale spatial and frequency trace to capture subtle forgery patterns, while the latter identifies tampered fragments via token-level reasoning and cross-modal grounding. A unified multi-task learning framework further enables joint optimization of manipulation localization, forgery type prediction, and overall media authenticity. Extensive experiments demonstrate the effectiveness of each module, and our method outperforms the state-of-the-art methods on large-scale benchmark datasets. We will release all the source codes.

Abstract:
In image editing, it is essential to incorporate a context image to convey the user's precise requirements, such as subject appearance or image style. Existing training-based visual context-aware editing methods incur data collection effort and training cost. On the other hand, the training-free alternatives are typically established on diffusion inversion, which struggles with consistency and flexibility. In this work, we propose VicoEdit, a training-free and inversion-free method to inject the visual context into the pretrained text-prompted editing model. More specifically, VicoEdit directly transforms the source image into the target one based on the visual context, thereby eliminating the need for inversion that can lead to deviated trajectories. Moreover, we design a posterior sampling approach guided by concept alignment to enhance the editing consistency. Empirical results demonstrate that our training-free method achieves even better editing performance than the state-of-the-art open-source training-based models.

Abstract:
Multi-view clustering is often challenged by incomplete observations arising from modality-wise information asymmetry. Existing incomplete multi-view clustering methods typically rely on simple imputation strategies or basic completion techniques, which struggle to capture the distributional heterogeneity and preserve the intrinsic sample relational structures from recovered representations, ultimately leading to suboptimal clustering performance. To address these limitations, we propose DFL (Disentangled Factor Learning with Normalizing Flows), a novel framework that explicitly models the latent generative process for incomplete multi-view clustering. Specifically, we disentangle latent embeddings into class-aware semantic factors and view specific auxiliary factors via conditional normalizing flow, and further imposes mutual information regularization terms to ensure cross-view semantic alignment. Missing views are then reconstructed through invertible flow transformations, which transfer both shared semantics and view-private residuals from observed views to maintain distribution consistency. Furthermore, we introduce a multi-level structural alignment mechanism, comprising prototype-level alignment to synchronize cluster geometry across spaces and pairwise similarity alignment to preserve global inter-sample relational structures. We iteratively refine clustering assignments and representation learning via an expectation-maximization strategy. Extensive experiments on benchmark datasets verify the effectiveness and robustness of the proposed framework for incomplete multi-view clustering.

Abstract:
Incorporating items' information shared among multiple modalities into a fused representation, multi-modal recommendation (MR) has demonstrated documented success than canonical unimodal recommendation. Although several attempts have been made to extract the discriminative information unique in each modality, existing methods suffer from a core limitation: the joint learning of modal-homogeneity discriminative information (HOI) and modal-heterogeneity discriminative information (HEI) tends to weaken their individual effectiveness. To remedy this deficiency, we propose a novel method, dubbed Disentangle and Distillation-based Dynamic Ensemble for multi-modal Recommendation (D3ER). We introduce gradient boosting into MR for the first time to formalize the optimization objective for alternately learning HOI and HEI. This design enables models dedicated to each type of information to focus on their proficient samples, thereby promoting specialized optimization. Furthermore, to mitigate the inherent high storage cost and risk of local optima in gradient boosting, we enhance our framework with knowledge distillation and a global correction regularization. Experiments on prevalent real-world datasets confirm the superiority of our proposed method on MR.

Authors: Qiang Li (XPeng Inc.), YingWenQi Jiang (ShanghaiTech University), Tuoxi Li (XPeng Inc.), Duyu Chen (XPeng Inc.), Xiang Feng (ShanghaiTech University), Yucheng Ao (University of Science and Technology of China), Shangyue Liu (XPeng Inc.), Xingchen Yu (University of Science and Technology of China), Youcheng Cai (University of Science and Technology of China), Yumeng Liu (University of Science and Technology of China), Yuexin Ma (ShanghaiTech University), Xin Hu (XPeng Inc.), Li Liu (XPeng Inc.), Yu Zhang (XPeng Inc.), Linkun Xu (XPeng Inc.), Gao Bingtao (XPeng Inc.), Xueyuan Wang (XPeng Inc.), Shuchang Zhou (XPeng Inc.), Xianming Liu (XPeng Inc.), Ligang Liu (University of Science and Technology of China)

Abstract:
Realistic and controllable simulation is critical for advancing end-to-end autonomous driving, yet existing approaches often struggle to support novel view synthesis under large viewpoint changes or to ensure geometric consistency. We introduce HybridWorldSim, a hybrid simulation framework that integrates multi-traversal neural reconstruction for static backgrounds with generative modeling for dynamic agents. This unified design addresses key limitations of previous methods, enabling the creation of diverse and high-fidelity driving scenarios with reliable visual and spatial consistency. To facilitate robust benchmarking, we further release a new multi-traversal dataset MIRROR that captures a wide range of routes and environmental conditions across different cities. Extensive experiments demonstrate that HybridWorldSim surpasses previous state-of-the-art methods, providing a practical and scalable solution for high-fidelity simulation and a valuable resource for research and development in autonomous driving.

Abstract:
Imitation learning based visuomotor policies have achieved strong performance in robotic manipulation, yet they often remain sensitive to egocentric viewpoint shifts. Unlike third-person viewpoint changes that only move the camera, egocentric shifts simultaneously alter both the camera pose and the robot action coordinate frame, making it necessary to jointly transfer action trajectories and synthesize corresponding observations under novel egocentric viewpoints. To address this challenge, we present EgoDemoGen, a framework that generates paired observationaction demonstrations under novel egocentric viewpoints through two key components: 1) EgoTrajTransfer, which transfers robot trajectories to the novel egocentric coordinate frame through motion-skill segmentation, geometry-aware transformation, and inverse kinematics filtering; and 2) EgoViewTransfer, a conditional video generation model that fuses a novel-viewpoint reprojected scene video and a robot motion video rendered from the transferred trajectory to synthesize photorealistic observations, trained with a self-supervised double reprojection strategy without requiring multi-viewpoint data. Experiments in simulation and real-world settings show that EgoDemoGen consistently improves policy success rates under both standard and novel egocentric viewpoints, with absolute gains of +24.6% and +16.9% in simulation and +16.0% and +23.0% on the real robot. Moreover, EgoViewTransfer achieves superior video generation quality for novel egocentric observations.

Abstract:
Compositional Zero-Shot Learning (CZSL) aims to recognize novel compositions of known primitives. Recent advances show that Compositional Scoring with Probabilistic Decomposition (CSPD) improves performance across models. While these gains are commonly linked to alleviating long-tailed distributions, we show that these gains are more closely linked to Primitive Distribution Misalignment (PDM), which describes the imbalance in the distribution of primitives within compositional space and the distribution of compositions in data space. We quantify PDM, offering a more precise explanation for CSPD's effectiveness. Additionally, we observe that CSPD decomposes predictions into multiple terms, each associated with an independent decoding path, but without enough constraints to ensure spatial consistency across these branches. To address these issues, we propose Primitive Distribution Alignment (PAL), a unified framework that aligns primitive and compositional learning across multiple paths. PAL enhances interactions between primitive and compositional representations through semantic expansion and a multi-level graph structure. Furthermore, we introduce graph Laplacian regularization in the cross-attention decoder, ensuring different paths focus on consistent regions. Extensive experiments under both closed-world and open-world settings demonstrate that PAL achieves state-of-the-art performance.

Abstract:
Multimodal crack segmentation has recently emerged as a more promising solution compared to RGB-only methods. However, prior multimodal paradigms require separate models for specific modality pairs. To enable flexible task adaptation, we propose UniCrack, an all-in-one crack segmentation framework that handles both single- and multi-modal inputs (e.g., RGB, depth, thermal, and polarization) using a shared set of parameters. To this end, we employ the off-the-shelf vision foundation model SAM, which inherently struggles to preserve slender, fine?grained topologies. We thus introduce: (1) a Spatio-Frequency Collaboration Adapter that learns cross-modal commonalities and specificities; (2) a Frequency-Domain Feature Calibration module that filters modality noise via spectral modulation; and (3) a Detail?Guided Progressive Refinement strategy that recovers thin, topological structures of cracks lost during feature extraction. Experiments demonstrate that UniCrack not only rivals but often surpasses task?specific specialists across diverse single- and multi-modal crack segmentation benchmarks, offering a robust and generalized solution for automated structural inspection.

Abstract:
Retrieval-augmented generation (RAG) is an important technique for enabling LLMs to utilize proprietary multi-modal datasets. Due to limited local computational resources, RAG services are commonly outsourced to third-party providers. However, the black-box nature of these services provides data owners with little transparency into how their data are accessed and utilized. As a result, it becomes challenging to verify whether the entrusted data are improperly reused in unauthorized systems or fully incorporated into the target service, posing significant risks to both the intellectual property and the service quality of valuable data assets. In this paper, we propose a unified framework based on propagating dataset watermarking (PDW) to audit data ownership and utilization integrity in fully black-box multi-modal RAG services. PDW embeds imperceptible signals into the database and activates them through query probes, allowing their effects to propagate through retrieval and generation to verify actual data usage. By modeling probe responses probabilistically, our framework enables hypothesis testing with theoretical guarantees for data auditing. Experiments on two public datasets, four retrieval models, and four LLMs demonstrate that, even under adversarial settings, our framework can reliably identify partial and unauthorized database usage while incurring negligible impact on overall performance.

Abstract:
Creating engaging videos traditionally requires professional editors and high-performance equipment, making the process challenging and inaccessible for general users. With the growing demand for high-quality video content and the widespread use of resource-constrained devices, there is an urgent need for lightweight video editing algorithms that can assist non-expert creators. In this work, we propose a Lightweight Shot Feature Extraction Network (LSFENet), tailored for video editing tasks. LSFENet employs a decomposed 3D convolution strategy to significantly reduce model scale while preserving the integrity of spatial and temporal information. Furthermore, we introduce Logit Dynamic Adjustment (LDA) to improve model prediction robustness for alleviating long-tail problem. We validate the effectiveness and generalization of LSFENet and LDA on two video editing datasets, AVE and MovieShots. For the shot attribute analysis task, LSFENet achieves per-class accuracies of 69.78%, 57.32%, and 53.96% for shot size, angle, and motion, respectively, while using only about 2.8M parameters and 2.2G FLOPs. In the next shot selection task, LSFENet uses 4.60M parameters and 15.70G FLOPs, respectively. Compared with AVE method with 35.60M and 15.70G, LSFENet improves prediction accuracy by around 1.2%. These results demonstrate that LSFENet effectively balances efficiency and performance, providing a promising solution for scalable and practical automatic video editing.

Abstract:
Robust and efficient point cloud registration under cross-modal and low-overlap conditions remains a critical challenge in transcranial navigation, where preoperative medical image data must be aligned with noisy and partial RGB-D observations. Existing methods relying on local descriptors or global learning frameworks often degrade significantly under noise, sparsity, and modality discrepancies, due to insufficient geometric information and a lack of semantic constraints. In this paper, we propose a facial landmarkguided cross-modal registration framework that integrates semantic cues, parametric geometric representation, and global structural consistency. In the preoperative stage, we reconstruct patients 3D head model from 3D medical images, and introduce a Quadric Surface with Trigonometric Polynomial (QSTP) description, which parameterizes local geometry around keypoints as a continuous surface within a local reference frame, enabling robust feature encoding under challenging conditions. In the intraoperative stage, we extract reliable 3D biomarkers from detected facial landmarks and formulate biomarker-keypoint correspondence estimation as a neighborhood-constrained graph matching (NCGM) problem, jointly optimizing description similarity and pairwise geometric consistency to achieve reliable matching in the presence of significant outliers. Extensive experiments on public datasets and real-world transcranial navigation scenarios demonstrate that the proposed method achieves superior accuracy and efficiency compared to advanced approaches, particularly under severe noise and partial observations. The proposed framework also enables stable real-time tracking and augmented-reality-guided transcranial examination, highlighting its practical applicability in clinical settings.

Abstract:
Group emotion recognition (GER) aims to infer the collective emotional states from images containing multiple individuals, addressing the fundamental challenge of accurately localizing key persons and emotion-relevant objects within the scene. Current graph-based approaches are often hindered by background noise, whereas methods based on the Most Important Person (MIP) paradigm tend to neglect significant non-facial cues. To address these shortcomings, we propose a Semantic-Prior-Guided Mask Attention Framework (SPMAF), which is founded on three principal innovations. First, we extend the MIP concept to MIPO, a novel representation that simultaneously encodes full-body information and emotion-relevant objects, thereby enabling a more holistic understanding of the scene. Second, we develop an innovative semantic localization strategy that combines Vision-Language Models (VLMs) with the Segment Anything Model (SAM) to generate high-fidelity, pixel-level masks of critical regions. Third, we introduce a lightweight mask-guided attention mechanism that selectively enhances emotion-relevant features within the global context through an additive enhancement process. Experimental results demonstrate that SPMAF attains accuracies of 95.44% and 95.13% on the GroupEmoW and SiteGroEmo datasets, respectively, outperforming the baseline model with the same backbone by 5.54 and 8.67 percentage points. These findings substantiate the efficacy of our semantic guidance and context-preserving attention mechanisms.

Abstract:
Real-world multimodal datasets suffer from noisy correspondence, which degrades model performance. The existing methods rely on global similarity to measure mismatch between samples. However, they overlook a critical phenomenon, locally semantically inconsistent false positives, where high CLIP similarity scores are incorrectly assigned to semantically mismatched samples. Consequently, such false positive pairs are used in model training without being filtered out, which undermines the model's performance. To address this issue, we propose a novel Localized Consistency Mining (LoCM) method to enhance the reliability of vision-language models during fine-tuning on noisy correspondence data. The LoCM framework comprises two modules: (1) A Local Noise Discrimination module that identifies fine-grained inconsistency by complementing CLIP's global matching with local correspondence analysis, enabling further noise filtering. (2) A Local Mismatch Refinement module realigns filtered samples with the integration of local consistency discrimination, mitigating local inconsistencies in the refined samples of existing methods. Experiments on the Flickr30K, MS-COCO and real-world noisy CC152K datasets demonstrate that LoCM significantly outperforms state-of-the-art approaches. Specifically, at a 60% noise rate, it improves Rsum by 14.7 on Flickr30K and 7.5 on MS-COCO. LoCM also obtains a 3.1 improvement on CC152K.

Abstract:
Breast ultrasound (BUS) video segmentation is important for tumor analysis, but it often fails to achieve promising performance due to artifacts and non-ridge tumor deformation in video sequences. Recently, space-time memory networks with linear state-space memory have achieved significant success in video segmentation. However, since memory update rule cannot stably preserve the associated historical memory, the state-space memory in previous methods will cause serious degradation on their segmentation performance, particularly when the video frames vary significantly. To address this problem, this paper presents Null Space Memory Network (NSMNet), a novel method that improves memory stability via the proposed Null Space Memory Update (NSMU) rule. NSMU controls the memory updates to the null space of the historical key subspace, which can compress the information from new frames without disrupting established key-value matching. To solve the over restrictive problem in NSMU, we further propose Approximate Null Space Projection (ANSP), which adaptively adjusts the number of latent semantic directions according to the intrinsic rank of the historical covariance matrix. This strategy preserves critical historical context information of the video frames and retains sufficient representational capacity to learn the emerging visual patterns. Extensive experiments on BUS video datasets demonstrate that NSMNet significantly enhances the segmentation results and keeps the linear computational complexity.

Abstract:
Structured-light (SL) cameras power depth sensing in millions of devices, and recent neural SL decoding methods have substantially improved their depth quality. Such reliable geometry can directly enable stable tracking and faithful reconstruction in SLAM. We present \ours, a practical SLAM system tailored for high-fidelity structured-light depth. Inspired by neural structured light (NSL)~\citeNSL, we incorporate strong monocular depth priors into iterative SL stereo decoding, reducing depth RMSE by 35% on Replica-SL compared to NSL. We then build a depth-centric SLAM pipeline that keeps dense metric geometry as the primary tracking signal, adding sparse visual correspondences for degenerate cases and lightweight bundle adjustment for long-range drift. On the synthetic Replica-SL benchmark, \ours achieves the best tracking accuracy and improves reconstruction F-score by 1.6 points over the strongest baseline under a shared-depth protocol. On 8 challenging real scenes, it is the only method that avoids catastrophic failure on every sequence while achieving 43.3% lower trajectory deviation than DROID-SLAM. The SLAM system runs online at 20.9 FPS, showing that stronger structured-light depth and depth-centric design together enable practical, robust SLAM.

Abstract:
The scaling of Large Multimodal Models (LMMs) is constrained by the quality-quantity trade-off inherent in synthetic data. Previous approaches, such as LLM-as-a-Judge, have proven their effectiveness in addressing this but suffer from prohibitive computational costs and lack of interpretability. To address this, we propose One-Step-Train (OST), a framework that reformulates data selection as an incremental optimization utility ranking problem. Instead of relying on semantic heuristics, OST measures the exact value of each sample by running a quick, single-step simulation on a small model. Extensive experiments on multimodal mathematical reasoning benchmarks demonstrate that OST consistently delivers high accuracy with minimal computational cost. By selecting the top-50% subset, OST reduces training costs by 43% (and total time consumption by 17%) while surpassing the strong LLM-as-a-Judge baseline by 1.8 points. Furthermore, under a fixed compute budget, our method using only the top-20% subset achieves a 5.6 point gain over LLM-as-a-Judge, improves upon heuristic scoring baselines like DEITA, and outperforms the Full-SFT baseline by 8.8 points. Notably, while Full-SFT suffers from performance degradation due to noise, our optimization-grounded approach effectively identifies toxic samples, successfully reversing the negative transfer frequently observed in complex reasoning tasks.

Abstract:
With the popularity of the large language models (LLMs), text steganography has achieved remarkable performance. However, existing methods still have some issues: (1) For the white-box paradigm, this steganography behavior is prone to exposure due to sharing the off-the-shelf language model between Alice and Bob. (2) For the black-box paradigm, these methods lack flexibility and practicality since Alice and Bob should share the fixed codebook while sharing a specific extraction prompt for each steganographic sentence. In order to improve the security and practicality, we introduce a black-box text steganography with a dynamic codebook and multimodal large language model. Specifically, we first construct a dynamic codebook via some shared session configuration and a multimodal large language model. Then an encrypted steganographic mapping is designed to embed secret messages during the steganographic text generation. Furthermore, we introduce a feedback optimization mechanism based on reject sampling to ensure accurate extraction of secret messages. Experimental results show that the proposed method outperforms existing white-box text steganography methods in terms of embedding capacity and text quality. Meanwhile, the proposed method has achieved better practicality and flexibility than the existing black-box paradigm in some popular online social networks.

Abstract:
In sensitive scenarios such as security surveillance and emergency search and rescue, Unmanned Aerial Vehicle (UAV) face recognition faces simultaneous challenges from severe physical degradation, scarcity of annotated samples, and privacy compliance constraints. Existing research mainly focuses on either low-quality face recognition under plaintext conditions or template protection and secure inference under privacy-preserving conditions. However, privacy-preserving target-domain adaptation for UAV face recognition remains underexplored. This paper proposes Pri-UAVFace, a Minimal Secure Multi-Party Computation (MPC) adaptation framework for privacy-preserving UAV face recognition. Rather than migrating the entire training workflow into MPC, it learns general identity representations and degradation robustness from public plaintext data, while confining residual target-domain adaptation to the ciphertext domain. To this end, it combines physical simulation and consistency learning for public-domain pre-learning, followed by MPC-based few-shot adaptation using Parameter-Efficient Fine-Tuning. Furthermore, an encryption-friendly discriminative head, named MPCCosineHead, is designed to reduce the protocol interaction overhead caused by non-linear operators. Evaluated on the real-world DroneFace dataset, Pri-UAVFace achieves an AUC of 0.9943 and a TAR@FAR=10^-3 of 0.9308 under a 5-shot setting. Through joint method and protocol optimization, it reduces the number of MPC communication rounds from 864 to 152 and shortens single-step latency by 75%. These results demonstrate effective UAV target-domain adaptation with minimal MPC utilization, providing a feasible path for privacy-sensitive edge multimedia sensing.

Abstract:
The rapid advancement of generative models has spurred the critical need to evaluate the worst-case robustness of deepfake detectors. In this paper, we reveal a fundamental blind spot in current forensic paradigms: while existing detectors excel at capturing digital synthesis artifacts, their effectiveness drops drastically when AI-generated content is cloaked in authentic physical imaging characteristics. We posit that genuine photographs inherently possess hardware-intrinsic statistical signatures, which are imperceptible footprints imprinted by optical sensors and Image Signal Processing (ISP) pipelines, and are fundamentally absent in purely data-driven generative models. Driven by this insight, we propose ISPCloak, a novel optimization-free adversarial attack framework that explicitly weaponizes the ISP pipeline to mislead the judgment of deepfake detectors. Rather than relying on computationally expensive gradient perturbations, our method first employs an Invertible ISP network to project images into the RAW domain. Then, we seamlessly imprint the complex statistical priors of real cameras onto AI-generated images by injecting realistic Poisson-Gaussian sensor noise and conducting forward ISP reconstruction. Synergized with generative artifact suppression and adaptive masking, this streamlined physical simulation enables ultra-fast generation of adversarial examples. Extensive experiments show that embedding authentic physical perturbations fundamentally disrupts a broad range of current detection mechanisms, yielding universally evasive adversarial examples with imperceptible visual alterations.

Abstract:
Text-to-image diffusion models have achieved remarkable generative capabilities, yet accurately aligning complex textual prompts with synthesized layouts remains an ongoing challenge. In these models, the initial Gaussian noise acts as a critical structural seed dictating the macroscopic layout. Recent online optimization and search methods attempt to refine this noise to enhance text-image alignment. However, relying on unconstrained Euclidean gradient ascent mathematically inflates the latent norm and destroys the standard Gaussian prior, causing severe visual artifacts like color over-saturation. Furthermore, these methods suffer from inefficient semantic routing and easily fall into the ``reward hacking'' trap of external proxy models. To address these intertwined bottlenecks, we propose Oracle Noise, a zero-shot framework reframing noise initialization as semantic-driven optimization strictly confined to a Riemannian hypersphere. Instead of relying on complex external parsers, we directly identify the most impactful structural words in the prompt to efficiently route optimization energy. By updating the noise strictly along a spherical path, we mathematically preserve the original Gaussian distribution. This geometric constraint eliminates norm inflation and unlocks aggressive step sizes for rapid convergence. Extensive experiments demonstrate that Oracle Noise significantly accelerates semantic alignment and achieves superior aesthetics without black-box models. It completely mitigates Euclidean-induced degradation, establishing state-of-the-art performance across human preference metrics (e.g., HPSv2, ImageReward), semantic alignment (CLIP Score), and sample diversity, all within a strict 2-second optimization budget.

Abstract:
As generative artificial intelligence evolves, deepfake attacks have escalated from single-modality manipulations to complex, multimodal threats. Existing forensic techniques face a generalization bottleneck: by relying excessively on superficial, modality-specific artifacts, they neglect the shared latent forgery knowledge hidden beneath variable physical appearances. Consequently, these models suffer catastrophic performance degradation when confronted with unseen "dark modalities." To break this limitation, this paper introduces a paradigm shift that redefines multimodal forensics from conventional "feature fusion" to "modality generalization." We propose the first modality-agnostic forgery (MAF) detection framework. By explicitly decoupling modality-specific styles, MAF extracts the essential, cross-modal latent forgery knowledge. Furthermore, we define two progressive dimensions to quantify model generalization: transferability toward semantically correlated modalities (Weak MAF), and robustness against completely isolated signals of "dark modality" (Strong MAF). To rigorously assess these generalization limits, we introduce the DeepModal-Bench benchmark, which integrates diverse multimodal forgery detection algorithms and adapts state-of-the-art generalized learning methods. This study validates universal forgery traces and enables robust detection of unseen modalities through MAF, offering a promising path toward universal multimodal defense.

Abstract:
Multimedia workflows increasingly depend on metadata completion, provenance capture, approval logging, packaging checks, and delivery verification, yet research has weakly specified how such governance demands should enter ongoing work. We formulate this as a multimedia systems problem of governance mediation: deciding whether, when, how, and to whom governance should be introduced during live production. To ground this agenda, we instantiate it in Kairos, an agent-based mediation prototype that separates risk appraisal, workflow-state assessment, responsibility routing, and intervention mediation while keeping the visible workspace fixed. We evaluate Kairos through a bounded evidence chain: semi-structured interviews, a risk-injected workflow study, a controlled proof-of-concept comparison, and a repeated-use feasibility probe. In the controlled study, UI+Mediation reduced governance failure at handoff, shortened resumption lag, and lowered NASA-TLX relative to UI-only, without evidence of increased task completion time. In the repeated-use probe, adaptive mediation improved acceptance and reduced repeated dismissal under continued use. These results do not constitute deployment-scale validation. Rather, they support a narrower claim: governance entry conditions in multimedia workflows are structured enough to be treated as a first-class multimedia systems problem and can be operationalized in a prototype mediation policy. More broadly, the paper reframes governance-as-interruption as a new research agenda for SIGMM.

Abstract:
Industrial anomaly detection is commonly studied under two simplifying assumptions: training data are clean and each model is built for a single product category. In practical manufacturing, however, multiple product types are often pooled for joint training, while the nominally normal set may still contain a small number of unlabeled defects or outlier samples. This leads to a more realistic yet underexplored setting, which we term label-free unified noisy-normal anomaly detection. In this regime, the difficulty is not only increased data heterogeneity, but also a deeper mismatch between the problem and existing memory-based pipelines: once multiple normal modes and contaminated samples are entangled in a single feature space, global ranking, filtering, and memory construction can no longer reliably separate natural inter-category variation from true anomalous deviation. We present SUMNAD, a semantic-routed subspace memory framework that approaches this setting from a different perspective: mixed noisy features should be decomposed before purification, rather than purified directly in a global mixed space. SUMNAD instantiates this idea through latent subspace routing, subspace-consistent denoising, and budgeted memory retrieval with continuous fallback. Rather than positioning these components as an isolated engineering combination, we view them as a concrete proof-of-concept for subspace-consistent purification in unified noisy anomaly detection. Experiments on MVTec AD and VisA under multiple contamination protocols show that this perspective is not only conceptually well-motivated, but also practically effective, yielding consistent improvements in both image-level detection and pixel-level localization.

Abstract:
How much do Vision-Language Models (VLMs) actually depend on their visual input? We introduce the Visual Reliance Score (VRS) and a five-condition behavioral testing framework to rigorously answer this question. By evaluating each query under standard, blank-image, random-image, contradictory-image, and image-only conditions across three image-essential benchmarks (DocVQA, ChartQA, TextVQA), we produce a behavioral fingerprint revealing when and how models engage with, or neglect, visual information. Across 24,000 inferences spanning two model families (Qwen2.5-VL, InternVL2) and two scales (7B and 72B), we find: (1) visual reliance is high on image-essential benchmarks (VRS 0.85-0.98) but varies significantly across tasks; (2) modality neglect persists even on these benchmarks, with up to 24% of answers unchanged when the image is removed, and most neglected answers being correct, revealing shortcut reliance on language priors; (3) scaling reduces neglect: the 72B model shows 1.0% average neglect vs. 15.6% for the 7B model, indicating that larger models are more visually grounded; and (4) a forced grounding intervention reduces neglect by 47% on ChartQA, though the effect does not transfer to less-structured scene-text tasks. We propose VRS as a mandatory companion metric for VLM evaluation and provide a taxonomy of neglect types (benign, harmful, and catastrophic) with implications for both benchmark design and training methodology.

Abstract:
Under skeletal topology constraints, the temporal evolution of human motion exhibits non-uniform local dependencies. Existing methods typically rely on globally uniform feature mixing, failing to capture these complex dynamics. This limitation leads to two core issues: local insensitivity, where globally uniform feature mixing tends to over-smooth highly nonlinear and abrupt local joint variations, causing subtle motion details to be diluted by the global context; and global instability, where prediction errors accumulate as the forecasting horizon extends, undermining stable long-term extrapolation. To address these issues, we propose the Local-Sensitive and Globally-Stable Network (LSGSNet), which performs collaborative modeling of human motion prediction from the perspectives of local sensitivity and global stability. Specifically, we design a Local Sensitivity Encoder (LSE) that employs spline basis functions for nonlinear representation learning, leveraging their local support property to enhance sensitivity to localized joint variations and capture fine-grained local dynamics. Furthermore, we introduce a Global Stability Propagation (GSP) module, which constructs a robust long-range temporal operator by dynamically gating short and long convolutions. This preserves local temporal details while stably propagating historical dependencies, thereby mitigating error accumulation during future extrapolation. The experimental results demonstrate that our method achieves the new state-of-the-art, outperforming existing methods by 26.3% on Human3.6M.

Abstract:
Video anomaly detection (VAD) is challenging as anomaly is inherently context-dependent and requires hierarchical reasoning over scene context, object dynamics, and their consistency with environment-specific norms. Although Vision-Language Models (VLMs) provide strong priors, they are mainly optimized for generic tasks rather than anomaly reasoning. Existing VLM-based VAD methods typically rely on parameter tuning, auxiliary modules, or prompt rephrasing, but do not explicitly improve the hierarchical reasoning process required for anomaly understanding. We propose PRIME (Prompt Refinement via Iterative Module-wise Evolution), a framework that adapts VLMs to VAD via rule-guided, module-wise optimization of hierarchical prompts. PRIME decomposes the prompt into Scene, Object, and Reasoning modules for contextual grounding, event-level evidence extraction, and norm-aware decision making. This design replaces holistic prompt rewriting with constrained module-wise refinement, enabling clearer error attribution and more stable optimization. Building on this structure, PRIME first distills execution traces and diagnostic feedback on training data into a compact rule memory, and then uses the induced rules to guide targeted updates over a population of candidate prompts. To preserve complementary strengths across prompts, PRIME retains non-dominated candidates via Pareto-based selection. After training, PRIME yields a single deployable prompt together with the learned rules, enabling efficient inference without model tuning. Experiments on UCF-Crime and XD-Violence demonstrate its strong explainable VAD performance, with 89.14% AUC and 82.41% AP, and promising cross-dataset generalization.

Abstract:
Listener facial expression generation in natural dyadic interaction requires responses that are not only realistic, but also consistent with social norms and emotional expectations, which are often reflected in human preferences. However, human-preference alignment remains largely underexplored in listener facial expression generation. In this paper, we propose a human-preference aligned listener facial expression generation method by leveraging human feedback to produce emotionally and socially appropriate expressions for natural dyadic interaction. A key idea of our method is to frame the generation of listener facial expressions as an action learning problem, allowing human feedback to assess the quality of responses without being confounded by the listeners appearance. Specifically, we first train a vision-language-action model via supervised fine-tuning to map the speakers multimodal cues to low-dimensional facial actions. We then construct a human preference dataset by collecting annotations over sampled listener responses, and further introduce a human-feedback reinforcement learning strategy to optimize the model via direct preference optimization (DPO). Experiments on the L2L-Trevor and RealTalk benchmarks show that our method improves emotional alignment and social appropriateness while maintaining competitive motion quality.

Abstract:
Audio-driven talking head synthesis has achieved remarkable visual fidelity, yet existing frameworks struggle to capture the fine-grained affective nuances and biological vitality essential for realistic human interaction. Most methods rely on coarse categorical emotion labels and neglect the physiological priors inherent to human speech, producing ``animated waxworks'' that lack lifelike dynamism. To address these challenges, we propose EMO-PhysioFlow, a novel multimodal framework that integrates large language model (LLM) reasoning with physiologically-conditioned latent diffusion at three complementary levels. At the representation level, a 4-bit quantized LLaVA-v1.6 performs clip-level affective reasoning, decoupling facial dynamics into canonical phonetic motions and continuous emotional residuals via a learnable Lie group gating mechanism. At the generation level, a physiological oscillator modulates the diffusion denoising process through hierarchical Physio-AdaLN layers, injecting affective-conditioned simulated vitality priors that mimic involuntary bio-rhythmic dynamics, to bridge the uncanny valley. At the supervision level, a preference alignment stage employing diffusion-DPO with action unit (AU) rewards steers the generative manifold toward perceptually accurate muscle activations. Extensive experiments on MEAD, CelebV-HQ and our self-collected dataset demonstrate that EMO-PhysioFlow consistently achieves higher quality with emotional accuracy, biological realism, and video quality. Qualitative comparisons and user studies further validate that it generates biologically vivid, identity-preserving portraits with fine-grained, physiologically coherent emotional expressions that generalize robustly to unseen identities and extreme affective conditions.

Abstract:
Recent few-shot anomaly detection methods have achieved strong performance on visual surface defects. However, performance remains limited for logical anomalies determined by multiple components within products, such as incorrect spatial connection or component quantity. Existing methods attempt introducing multimodal large language models to describe product logic rules. Nevertheless, the textual descriptions of component relationships are difficult to align with visual features, leading to suboptimal performance. To this end, we propose a new perspective for logical anomaly detection by modeling a product as a graph composed of multiple components and their relationships. Based on this perspective, we introduce GLOG, a few?shot anomaly detection method. Specifically, we formulate part-segmented components and containers as graph nodes, utilizing spatial and containment edges to explicitly encode their positional and inclusive relationships. Over this graph, we devise an edge-specific message passing strategy to propagate pairwise relations across multi-node associations, holistically capturing both global spatial layouts and quantitative information. Besides, a visual branch is incorporated to jointly detect visual surface anomalies. Extensive experiments on multiple benchmark datasets demonstrate the effectiveness of our proposed method.

Abstract:
Industrial anomaly generation provides an effective way to alleviate the scarcity of defective samples for visual inspection, yet existing methods often struggle to simultaneously achieve realistic defect appearance, accurate mask alignment, and semantic consistency under limited anomaly supervision. In this paper, we propose AnomalyBrush, a dual-branch conditional diffusion framework for few-shot anomaly image generation. Specifically, a prior branch extracts multi-scale structural and semantic defect cues from a small set of real anomaly exemplars, while a synthesis branch generates anomalies on normal images within a anomaly mask through adaptive cross-attention fusion. To further improve mask-aligned generation, we introduce Topology-aware Structural Attention Recalibration (TSAR), which combines SSIM-based structural discrepancy with a distance-transform-based geometric prior to emphasize under-synthesized regions, especially along boundaries and thin structures. In addition, we propose Prototype-augmented Semantic Alignment (PSA), which enhances category-faithful synthesis by matching generated anomaly features to a real anomaly prototype bank and by anchoring object-level semantics in a shared vision--language embedding space. Extensive experiments on MVTec AD and Real-IAD demonstrate that AnomalyBrush produces more realistic and diverse anomaly samples with better spatial correspondence to the target masks, and consistently improves downstream anomaly localization performance. These results highlight the effectiveness of our framework as a practical data generation solution for industrial anomaly inspection.

Abstract:
Text-to-Motion generation aims to synthesize 3D human motion from natural language. While diffusion models naturally capture motion continuity, they often underperform discrete token-based approaches. Emerging applications such as motion editing and multimodal control, however, require fine-grained control, compositionality, and diverse sampling, which are better supported by continuous frameworks. Motivated by this gap, we conduct a systematic analysis of existing continuous diffusion methods and identify two key challenges. Prior continuous approaches fall short in both motion representation and generation, exhibiting limited discriminative quality in latent spaces and suboptimal performance in reconstruction and synthesis. To address these challenges, we propose COME, a continuous diffusion framework that jointly improves motion representation and generation. It integrates a high-quality motion tokenizer, MoCMAE, which produces more discriminative and structured features, with a diffusion generator, ccDIT, that stabilizes training and inference while preserving fine-grained semantic alignment. Extensive experiments show that COME achieves state-of-the-art generation quality and efficiency, outperforming both prior continuous and discrete approaches.

Abstract:
Vision-and-Language Navigation (VLN) aims to enable an embodied agent to follow natural-language instructions and navigate to a target location in unseen 3D environments. Recent VLN methods increasingly build on general vision-language models (VLMs) and directly learn an end-to-end mapping from instructions and egocentric observations to low-level actions. While effective, this paradigm still suffers from a key limitation: VLMs tend to learn priors that favor semantic understanding, making direct observation-to-action learning prone to learn shallow pattern matching rather than the dynamic spatial awareness that understands how and why the observation transition occurs over time. We argue that adapting VLMs to VLN requires endowing them with two complementary capabilities for acquiring such awareness, namely backward action reasoning (why) and forward transition prediction~(how). Based on this insight, we propose SpaAct, a simple yet effective training framework that activates the dynamic spatial awareness in VLMs. Specifically, SpaAct introduces two spatial activation tasks: Action Retrospection, which asks the model to infer the executed action sequence from visual transitions, and Future Frame Selection, which forces the model to predict the visual transitions conditioned on history and action. These two objectives provide lightweight supervision on both backward action reasoning and forward transition prediction, encouraging the model to build dynamic spatial awareness in a VLM-friendly way. To further stabilize adaptation, we design TriPA, a Tri-factor Progressive Adaptive curriculum learning method that organizes training samples from easy to hard, allowing the model to gradually acquire navigation skills from basic locomotion to long-horizon reasoning. Experiments on standard VLN-CE benchmarks show that SpaAct consistently improves VLM-based navigation and achieves state-of-the-art performance. We will release the code and models to support future research.

Abstract:
Video thumbnails serve as the primary gateway to content discovery on web platforms, yet creating effective thumbnails demands professional design expertise that remains inaccessible to most creators. Existing automated approaches rely on keyframe selection, heuristic region extraction, or unconstrained generation, and fail to jointly optimize semantic fidelity, aesthetic quality, and compositional coherence. We present AgenticMontage, a multi-agent framework that synthesizes perceptually grounded composite thumbnails from video content through coordinated intelligent agents. Our framework decomposes thumbnail creation into a structured pipeline of strategic planning, aesthetic-driven keyframe curation, hierarchical clustering, and staged diffusion synthesis. The synthesis agent generates thumbnails through a multi-phase diffusion process incorporating Prompt-to-Prompt attention control for semantic consistency, Chain-of-Thought layout reasoning for principled composition, object-aware guidance for content preservation, and critic-driven iterative refinement for quality assurance. By grounding every synthesis decision in perceptual dimensions of semantic relevance, visual appeal, and narrative significance, AgenticMontage produces thumbnails that approach the quality of professionally designed references. Extensive experiments including quantitative evaluation and a user study demonstrate that our method outperforms existing automated baselines across multiple dimensions while narrowing the gap to human-crafted thumbnails.

Abstract:
Multi-view clustering (MVC) aims to discover cluster structure from multiple views by exploiting shared semantics while preserving complementary information. Recent mixture-of-experts (MoE) based MVC methods have demonstrated the value of shared sparse experts for collaborative representation learning, yet they still leave sparse expert learning weakly coupled to clustering. In particular, when clustering is performed in a post-hoc manner on learned embeddings, the clustering module acts only as a readout and therefore cannot provide in-training guidance to routed features. To address this limitation, we propose GSMVC, a framework that integrates guided sparse expert learning, shared hyperspherical prototype clustering, and LLM-based regime calibration. Specifically, we first employ a shared sparse MoE backbone to learn collaborative multi-view representations. We then introduce a shared hyperspherical prototype bank that produces per-view prototype logits and unified cluster assignments, enabling unified prototype supervision to participate in routed representation learning through prototype-space contrastive learning. Building on this backbone, we further use a large language model as a high-level controller that maps structured summaries of routing and hyperspherical prototype states to expert-prior calibration signals, instead of directly making per-sample clustering decisions. Experiments on MVC benchmarks show that GSMVC achieves better clustering performance than existing deep MVC methods.

Abstract:
Recent advances in Multimodal Large Language Models (MLLMs) have created new opportunities for facial expression recognition (FER), moving it beyond pure label prediction toward reasoning-based affect understanding. However, existing MLLM-based FER methods still follow a passive paradigm: they rely on externally prepared facial inputs and perform single-pass reasoning over fixed visual evidence, without the capability for active facial perception. To address this limitation, we propose ActFER, an agentic framework that reformulates FER as active visual evidence acquisition followed by multimodal reasoning. Specifically, ActFER dynamically invokes tools for face detection and alignment, selectively zooms into informative local regions, and reasons over facial Action Units (AUs) and emotions through a visual Chain-of-Thought. To realize such behavior, we further develop Utility-Calibrated GRPO (UC-GRPO), a reinforcement learning algorithm tailored to agentic FER. UC-GRPO uses AU-grounded multi-level verifiable rewards to densify supervision, query-conditional contrastive utility estimation to enable sample-aware dynamic credit assignment for local inspection, and emotion-aware EMA calibration to reduce noisy utility estimates while capturing emotion-wise inspection tendencies. This algorithm enables ActFER to learn both when local inspection is beneficial and how to reason over the acquired evidence. Comprehensive experiments show that ActFER trained with UC-GRPO consistently outperforms passive MLLM-based FER baselines and substantially improves AU prediction accuracy.

Abstract:
Composing music that responds meaningfully to narrative videos remains an open challenge. Most existing systems bridge video and music through emotion, typically valence and arousal, but these representations discard the temporal and causal structure of narrative; they capture how a scene feels, not what is happening or why. We propose narrative suspense and musical tension as an alternative bridge: suspense is a structural property of narrative computed from the state of story threads, and musical tension is a perceptual property quantified through psychoacoustically grounded features. We present a system that extracts narrative structure from video using a multimodal language model, computes a per-event suspense curve, and composes music whose musical tension is aligned with the suspense curve through multi-objective evolutionary optimization. A within-subjects listening study with 83 participants compared the full system against an ablation that removes the tension alignment objective while matching musical quality. The results show that tension-aligned music is rated significantly higher on music-narrative fit, with no difference in musical pleasantness. The effect is strongest for scenes with clear dramatic structure and absent for narratively ambiguous content, suggesting that the approach is most effective when viewers share a common expectation of a scene's emotional trajectory.

Abstract:
In event-driven financial forecasting, textual news acts as an exogenous causal intervention on market states. Unlike traditional 1D numerical sequences, 2D K-line charts provide the spatial topology needed to localize such interventions and encode gestalt market patterns. However, existing multimodal models treat text and images symmetrically, conflating the causal impact of news with spurious correlations from unobserved market inertia. To address this issue, we propose Causal Market-State Generation (CMSG), which formalizes financial news as a do-calculus intervention. Specifically, CMSG uses vision models' spatial priors to identify news-sensitive chart regions and applies a Vector Quantization-based Causal Tokenizer to orthogonally decompose them into endogenous market inertia and an exogenous intervention residual. This decomposition serves as a structural proxy for confounder adjustment, supported by a contrastive objective that separates factual and counterfactual attention maps for identifiable Individual Treatment Effect estimation. We evaluate CMSG on Quant-MM, a benchmark of 2,500 annotated macro-news events paired with K-line charts from November 2024 to March 2026. Results show that CMSG outperforms 12 baselines, achieving 46.85% Macro-F1 (4.75% above the strongest time-series SOTA) and a 4.8522 Information Ratio, thereby setting a new benchmark for risk-adjusted returns.

Abstract:
Multimodal sarcasm detection has emerged as a critical research direction, and recent multimodal large language models (MLLMs) have opened new possibilities toward deeper affective reasoning in conversational video scenarios. However, existing approaches still suffer from two key challenges: data-level annotation deficiency and architecture-level context entanglement, where the former provides only utterance-level binary labels while neglecting fine-grained reasoning annotations, and the latter entangles contextual history with target delivery, preventing explicit identification of multimodal conflicts. To address these issues, we first introduce Multimodal Understanding and Sarcasm Explanation DataSet (MUSE-Set), a bilingual video dataset with four-dimensional explanatory annotations decomposing sarcasm reasoning into character descriptions, situational contexts, incongruity analysis, and intended meanings. Building on this, we propose Memory-Enhanced Incongruity Reasoner (MEIR), a novel framework for multimodal sarcasm detection and explanation. MEIR employs a cognitively inspired dual-phase architecture: an episodic memory generator first compresses context videos into structured character-tracked memories, which an incongruity reasoner then contrasts with target utterance cues to explicitly identify conflicts and generate fine-grained explanations. To the best of our knowledge, MEIR is the first unified framework to jointly perform multimodal sarcasm detection and natural language explanation generation. Extensive experiments demonstrate that MEIR achieves best performance in both tasks, consistently outperforming strong MLLMs including Gemini-3-Flash and GPT-4o.

Abstract:
Deepfake detection under cross-dataset settings remains highly challenging. Although adapting CLIP has shown strong potential for generalizable face forgery detection, existing adapter-based solutions may still be distracted by irrelevant visual context and overly rely on dominant but non-robust artifacts, which weakens generalization to unseen or high-quality forgeries. To address this issue, we propose the F^2-Adapter, a Focus-and-Forget framework built on a frozen CLIP image encoder. Our method combines an artifact focusing mechanism, which introduces content-aware sparse routing over visual tokens to reduce interference from irrelevant regions, and a shortcut forgetting mechanism, which progressively suppresses highly activated feature locations during training through a curriculum-based masking strategy. These two components are complementary: the former encourages the model to concentrate on forensic-relevant areas, while the latter discourages over-reliance on easy cues and promotes the learning of more generalizable forensic patterns. Extensive cross-dataset experiments show that the proposed method achieves state-of-the-art performance, reaching an average AUC of 90.4% and consistently outperforming strong baselines on unseen benchmarks.

Abstract:
Financial chart summarization requires numerical grounding and domain-specific reasoning, yet existing data and general-purpose VLMs provide limited support for this task. We introduce FinSumm-8B, an analyst-oriented VLM trained by supervised fine-tuning and fact-calibrated preference optimization (FCPO), which uses source-data verification to calibrate heterogeneous preference pairs. We also construct FinChartSum, a 1K-sample benchmark covering real and synthesized charts. On a balanced 200-sample cross-model evaluation, FCPO improves the base Qwen3-VL-8B LLM-Score from 0.675 to 0.762 and exceeds the 30B model (0.726), while GPT-5.4 (0.858) and Gemini-3.1-Pro (0.844) remain stronger under judge-based quality evaluation. These results support domain-specific post-training for efficient financial chart summarization while exposing remaining gaps on source-free real charts.

Abstract:
Diffusion Transformers with Mixture-of-Experts (DiT-MoE) improve model capacity under sparse activation, but diffusion inference is still bottlenecked by redundant computation across timesteps. Existing caching methods mainly operate at the token level, which becomes suboptimal in DiT-MoE because each token update is internally decomposed into multiple routed expert branches. Our analysis shows that cross-timestep redundancy in DiT-MoE is better characterized at the expert-branch level than at the whole-token level. Based on this observation, we propose MoECa, a fine-grained caching framework that performs branch-level feature reuse across timesteps. MoECa further introduces expert-aware adaptive control and synchronized cache updates across MoE and attention paths to maintain stable intermediate states. Experiments on multiple DiT-MoE models show a favorable speed--quality trade-off, with up to 2.93× speedups while preserving generation quality.

Abstract:
Scanpath prediction aims to generate realistic fixation sequences conditioned on visual scenes and tasks, capturing human visual attention. Recent studies model inherent gaze variability using generative techniques, particularly diffusion models. While effective, existing methods treat scanpath generation as modeling of holistic sequence distributions, in which contextual information is used as a conditioning signal rather than explicitly embedded in fixation sequences. This formulation overlooks that scanpaths are inherently composed of spatial-temporal, context-dependent fixation transitions, which may limit fine-grained coherence. To address this limitation, we propose a fixation-level generative paradigm for scanpath prediction. Unlike existing methods that capture coarse global structures, we model scanpaths as sequential fixation transitions, enabling fine-grained dependency and dynamic modeling across fixations, thereby naturally balancing coherence and variability. Specifically, we introduce a new framework, ScanFlow, with three key components. First, we integrate spatial-temporal and contextual semantics to construct scanpath representations that capture transition dependencies to preserve coherence. Second, we use flow matching with semantics-aligned contrastive regularization to model stochastic transition dynamics, thereby capturing structured and context-aware variability. Third, we adopt a flexible autoregressive strategy that balances deterministic and stochastic components to sequentially generate diverse yet context-coherent fixation sequences. Extensive experiments across five datasets with various viewing scenarios demonstrate the effectiveness of ScanFlow for high-quality and realistic scanpath generation.

Abstract:
Dataset Distillation (DD) aims to compress a large training set into a tiny yet effective distilled dataset, thereby substantially reducing training and storage costs. Recently, the strong capabilities of diffusion models have driven rapid progress in DD, showing clear promise for large-scale and high-resolution settings. However, most existing methods still rely on diffusion models pre-trained on or further adapted to the full target dataset, incurring substantial additional cost and undermining the motivation of DD. On the other hand, although some methods turn to general text-to-image models to avoid target-specific training, they suffer from distribution mismatch, as generic generative priors fail to faithfully capture target-specific semantics, leading to suboptimal performance. To address these challenges, we propose Structure-Anchored Refinement (SAR), a dataset distillation framework built upon text-to-image diffusion priors that reformulates DD as a structure-anchored semantic refinement process. SAR first identifies representative samples from the real training data that capture the core distribution and uses them as structural anchors. It then leverages diffusion priors to refine their semantics while preserving spatial layouts and intra-class variations, thereby increasing the density of class-relevant information. Within this refinement process, we introduce an adaptive patch reweighting mechanism to suppress background regions and other irrelevant factors during generation, enabling more stable alignment with the target distribution. Extensive experiments show that SAR achieves performance on par with or better than that of existing approaches across multiple benchmarks, including ImageNet-1K and its subsets.

Authors: Joy Zhang (Researcher at Beijing Innovation Center of Humanoid Robotics), Che Liu (Imperial College London), Xiancong Ren (Beijing Innovation Center of Humanoid Robotics), Hanchu Ni (Peking University), Yingji Zhang (University of Manchester), Shuai Zhang (Westlake University), Zeyuan Ding (Beijing Innovation Center of Humanoid Robotics), Jiayu Hu (Beijing Innovation Center of Humanoid Robotics), Haozhe Shan (Fudan University), Junbo Qi (Waseda University), Yan Bai (NVIDIA), Li Dengjie (Meituan Inc.), Jiachen Luo (Queen Mary, University of London), Yidong Wang (Peking University), Yong Dai (Beijing Innovation Center of Humanoid Robotics), Zenglin Xu (Shanghai Academy of AI for Science), Bin Shen (Celonis AI), Qifan Wang (Meta AI), Jian Tang (Beijing Innovation Center of Humanoid Robotics), Xiaozhu Ju (Beijing Innovation Center of Humanoid Robotics)

Abstract:
Developing a versatile embodied brain is fundamentally challenged by severe task heterogeneity, where discrete reasoning and continuous numerical perception tasks coexist with radically different data densities, creating imbalanced training distributions that resist straightforward scaling solutions. Conventional post-training pipelines that apply SFT sequentially before RL are not architected for this regime: their uniform treatment of task distributions suppresses underrepresented capabilities and exacerbates catastrophic forgetting in data-sparse tasks. In response, we introduce Deliberate Practice Policy Optimization (DPPO), an iterative post-training framework tailored for the heterogeneous, data-sparse embodied setting. Rather than applying SFT and RL in sequence, DPPO repurposes RL rollouts as a difficulty-aware triage mechanism: rollouts that expose unresolved weaknesses route the corresponding samples to SFT for targeted knowledge injection, while already-mastered data is filtered out to avoid redundant supervision. Each SFT round in turn consolidates a stronger baseline for the next RL round to probe, forming a self-reinforcing cycle that compounds gains across iterations, mitigates catastrophic forgetting, and maximizes utility from finite embodied data. Conceptually, DPPO is grounded in the complementary roles of SFT and RL: SFT consolidates knowledge from positive expert exemplars, while RL expands capability boundaries through comparative rollout trajectories, and their iterative interplay converts diagnosed weaknesses into structured supervision signals for continual policy reinforcement. Empirically, our DPPO-trained model, Pelican-VL 72B, yields a 20.3% performance improvement over its base model, outperforms open-source models at the 100B scale by 10.6%, and remains competitive with substantially larger proprietary models, providing a self-improving framework for the community.

Abstract:
In face recognition systems, facial templates are widely adopted for identity authentication due to their compliance with the data minimization principle. However, facial template inversion technologies have posed a severe privacy leakage risk by enabling face reconstruction from templates. This paper proposes a Layer-Based Facial Template Inversion (LBFTI) method to reconstruct identity-preserving fine-grained face images. Our scheme decomposes face images into three layers: foreground layers (including eyebrows, eyes, nose, and mouth), midground layers (skin), and background layers (other parts). LBFTI leverages dedicated generators to produce these layers, adopting a rigorous three-stage training strategy: (1) independent refined generation of foreground and midground layers, (2) fusion of foreground and midground layers with template secondary injection to produce complete panoramic face images with background layers, and (3) joint fine-tuning of all modules to optimize inter-layer coordination and identity consistency. Experiments demonstrate that our LBFTI not only outperforms state-of-the-art methods in machine authentication performance, with a 25.3% improvement in TAR, but also achieves better similarity in human perception, as validated by both quantitative metrics and a questionnaire survey.

Abstract:
In real-world multimodal web scenarios, graph-structured data often arrives in a streaming manner, making graph continual learning a crucial paradigm for continuously modeling such evolving structures. However, existing graph continual learning methods still face two fundamental challenges. 1) semantic-structural separation, where the graph-based methods excel at modeling topological relationships but neglect deep semantics. 2) imbalanced knowledge transfer, where existing models fail to effectively leverage general knowledge gained from early tasks to benefit subsequent new tasks. To address above issues, we propose a novel framework, UNleash Large Language Models PotentIal for Graph ConTinual Learning (UNIT). By fine-tuning large language model only on the first task, we bridge the distributional gap between the pre-trained LLM corpus and the target task dataset to enhance the adaptability of LLMs for graph-structured tasks. Meanwhile, we propose an uncertain-aware anchor generation mechanism to effectively preserve representative knowledge across tasks, avoiding the neglect of universal knowledge learned from previous tasks. Additionally, we introduce structural confluence modeling to explicitly integrates graph topology information into semantic information, enhancing the collaborative capabilities between semantic understanding and structural modeling. Extensive experiments demonstrate that our proposed method achieves state-of-the-art performance in the graph continual learning task.

Abstract:
Existing computer-aided design (CAD) reverse engineering methods predominantly rely on a single input modality. Even when certain multimodal models can accommodate multiple inputs, they are typically restricted to selecting only a single modality during inference, failing to fully exploit the complementary potential of diverse information sources. To this end, we propose UniFuse-CAD: a unified multimodal CAD program synthesis framework built upon a Vision-Language Model (VLM), supporting any combination of text, image, and point cloud inputs to generate executable CadQuery programs. Experimental findings reveal significant complementarity among visual modalities, but naively concatenating text with visual features disrupts visual representations and induces negative transfer. To address this issue, we design the Gated Text-to-Visual Injection (GTVI) module, which selectively injects text information into visual tokens via gated cross-attention, combined with a Do-No-Harm training constraint and a text-vision alignment loss, unleashing the performance gains of text on hard samples without harming the visual pathway. Furthermore, to address the limited coverage of advanced design features in existing CAD datasets, we construct ForgeCAD: a large-scale multimodal dataset generated through a Large Language Model (LLM)-driven semantic augmentation pipeline, significantly improving data scale and the diversity of advanced features. Experiments demonstrate that UniFuse-CAD outperforms existing methods across various modality combinations, exhibiting superior reconstruction accuracy and robustness.

Abstract:
With the widespread development of diffusion models, 3D point cloud generation has achieved remarkable advancements, facilitating the efficient synthesis of complex geometric shapes. However, existing methods often overlook high-dimensional geometric priors by directly using conventional isotropic frameworks. Consequently, the 3D generation process lacks robustness in the structural fidelity, particularly during few-step inference. To address the issue, we propose HeatOL, a heat equation-driven generative model for high-quality point cloud generation under an operator learning paradigm. In the forward process, we utilize the heat equation to construct a geometric evolution that progressively smooths local geometric details. The heat dissipation is governed by a rigorously parameterized anisotropic heat kernel to effectively capture both global context and local gradients. Meanwhile, to tackle the ill-posed nature of solving the inverse heat equation, we design 3D frequency-aware Diffusion Transformer (DiT) modules for the reverse generation process. Specifically, we tailor the 3D window attention mechanism with varying directional sizes for frequency decomposition and incorporate an adaptive spectral filter to stabilize the inverse evolution through Fourier regularization. Building upon the spectral architecture, we develop an efficient solution map for approximating the heat flow trajectories via Fourier Neural Operators (FNO). Extensive experimental results show that HeatOL outperforms state-of-the-art models in unconditional generation and conditional completion. With further validations in the few-step regime, HeatOL yields significant gains in the structural fidelity of point cloud diffusion models while maintaining robustness for few-step sampling.

Abstract:
All-in-one image restoration is a unified low-level vision task that aims to effectively recover high-quality images from inputs degraded by various types and levels of corruption using a single model. Recent works have achieved remarkable progress by learning degradation-adaptive prompts or network architectures. However, these methods typically apply a uniform restoration strategy across the entire image, neglecting the fact that different regions may suffer from distinct degradation types and varying degrees of severity. In contrast, we propose to perform restoration at the pixel level, thereby enabling more fine-grained and precise control over the restoration process. Specifically, we present MGN-AIR, a novel pixel-level restoration framework for all-in-one image restoration. Our approach first learns to estimate a pixel-level visual prompt. Then, it leverages both textual and visual prompts to provide global and local degradation cues, guiding the model on where to look and how to restore at each pixel. We conduct extensive experiments on multiple all-in-one image restoration benchmarks, covering a wide range of tasks including denoising, deraining, deblurring, dehazing, desnowing, and low-light enhancement. Experimental results demonstrate that our proposed method consistently and significantly outperforms existing approaches.

Abstract:
Aerial object detection is severely challenged by drastic scale variations, ego-motion blur, and complex illumination. While integrating event cameras with RGB sensors is promising, existing spatial-domain fusion methods still struggle to suppress low-frequency background features and are constrained by rigid same-stage fusion requirements, thereby neglecting intra- and inter-modal cross-stage interactions. To address this, we propose AeroFuse, a novel Joint Spatial-Scale and Frequency-Domain RGB-Event Fusion Network. AeroFuse operates via two core modules: First, the Multi-Scale Aggregation (MSA) module employs non-destructive spatial reshaping and residual soft-thresholding to break rigid same-stage fusion, enabling robust cross-scale feature guidance while suppressing background noise. Second, the Frequency-based Fusion (FF) module shifts the fusion paradigm to the Fourier domain. By leveraging a High-Frequency Structural Attention (HFSA) mechanism and explicit phase-amplitude disentanglement, the FF module effectively enhances high-frequency edge details and aligns geometric structures (phase) while preserving complementary textures (amplitude). To address the scarcity of paired aerial RGB-Event data, we additionally contribute a large-scale synthetic dataset derived from the VisDrone2019 and UAVDT benchmarks using the V2E simulator. Extensive experiments demonstrate that AeroFuse achieves state-of-the-art performance.

Abstract:
Decoding dynamic visual stimuli from neural activity is vital for understanding human cognition. Current fMRI-to-video reconstruction methods have achieved promising visual fidelity. However, cross-subject and few-shot adaptation remains still face severe limitations. Existing approaches often project the diverse neural activities of new subjects into a static pre-trained bottleneck, which inherently triggers manifold collapse and restricts adaptation to unseen physiological distributions. To address this issue, we propose a novel cross-subject framework named Mind-Imagery to achieve fast new-subject adaptation by actively synthesizing paired fMRI samples. First, we introduce the Consensus Decoupler (CD) to explicitly disentangle multi-subject brain activity into semantic consensus, individual variance, and physiological noise. Second, the Imagery Generator (IG) is proposed to reconstitute these latent components and obtain semantic-to-fMRI synthesis under hierarchical cycle-consistency supervision. Furthermore, by leveraging the rich semantic priors of LLMs, Mind-Imagery enables open-vocabulary fMRI synthesis from diverse text variants, establishing a highly scalable paradigm for few-shot cross-subject generalization. Extensive experiments demonstrate that Mind-Imagery achieves state-of-the-art performance in both cross-subject and few-shot scenarios, and retains 98.3% of its full-session performance (0.847 vs. 0.862) using merely 10% of calibration scans.

Abstract:
As multimodal large language models (MLLMs) are increasingly applied in scenarios with long textual contexts and high-resolution images, KV cache memory footprint during inference has grown substantially. To optimize inference efficiency, existing compression methods mostly rely on scoring-based strategies, which estimate token importance from localized attention weights and perform retention independently at each head. As a result, the distribution of retained tokens may exhibit irregular sparsity across layers. Drawing inspiration from structured pruning in model compression, we investigate how unstructured token retention affects model inference performance. Our observations reveal that the irregular sparsity introduced by unstructured retention breaks the vertical integrity of critical token refinement, hindering the accumulation of token information across layers and leading to severe performance degradation. Motivated by these findings, we propose VertiKV, a structured KV cache compression strategy that: (1) enforces vertical token consistency via Vertical-Aligned Structural Eviction (VASE), which partitions adjacent layers into minimal structural units (i.e., chunks) with adjustable depths and executes a token-level union within each chunk to align with the inherent vertical refinement process of token representations; and (2) enhances semantic consistency via Top-Down Information Enhancement (TDIE), which exploits the aggregated semantics of evicted tokens encoded in retained tokens and performs similarity-guided merging from upper aligned retained tokens within each chunk to fully utilize residual context. Extensive experiments on representative MLLMs across challenging multimodal long-context benchmarks demonstrate that our method achieves 1.21.9 decoding speedup and reduces KV cache memory by 8095% while maintaining task performance.

Abstract:
Recommender systems often suffer from over-concentrated results when optimized for accuracy, limiting recommendation diversity. While diversity has become an important objective, existing methods still face fundamental limitations from both semantic and structural perspectives. On the one hand, multimodal recommender systems provide rich semantic signals, yet are often treated as auxiliary to ID embeddings, leading to ineffective semantic utilization for diversity. On the other hand, the accuracy--diversity trade-off persists during training, where accuracy gains often come at the cost of diversity. We show that this phenomenon is fundamentally governed by both the structural properties of learned representations and the dynamics of optimization. To address these challenges, we propose MOSS, a unified framework for multimodal decoupling through semantic and structural shaping. Specifically, we first introduce an MLLM-based Decoupled Feature Extraction module to disentangle multimodal content into informative representations. To enhance semantic utilization, we design a Category-aware Embedding Selection module that extracts compact, category-aware semantic features. To address structural limitations, we propose two complementary modules: Alternating Dual-Graph Propagation, which alleviates representation collapse by integrating collaborative and multimodal propagation, and Entropy-based Diversity Regularization, which mitigates optimization bias by promoting balanced category exposure. Extensive experiments on real-world datasets demonstrate that MOSS consistently achieves a superior accuracy--diversity trade-off over strong baselines.

Abstract:
Continual post-training is essential for adapting Multimodal Large Language Models (MLLMs) to evolving tasks, but it often leads to catastrophic forgetting. The core challenge is to achieve rapid adaptation to new task formats and semantics while preventing policy drift from accumulating into forgetting during sequential training. Existing paradigms address this trade-off only partially: Supervised Fine-Tuning (SFT) enables rapid adaptation but often disrupts prior knowledge, whereas Reinforcement Learning (RL) better preserves existing capabilities through on-policy updates but typically converges slowly and struggles on open-ended tasks. We propose OP-CL, a two-stage on-policy guided framework for continual post-training in MLLMs. In Stage 1, On-Policy Corrective Supervision transforms model rollouts into corrective supervision, providing efficient learning signals for acquiring new-task behaviors, especially on open-ended tasks. Since such rapid adaptation may still introduce policy deviation, Stage 2, Reward-KL Stabilized Reinforcement further stabilizes continual adaptation by adaptively regulating KL regularization to limit excessive policy updates. The two stages therefore address two complementary requirements of continual post-training: efficient acquisition and stable retention. Experiments show that OP-CL consistently achieves a better balance between new task adaptation and old task retention than representative SFT- and RL-based baselines.

Abstract:
Class-Guided Camouflaged Object Detection (CGCOD) leverages category-level semantic priors to improve model perception in complex camouflage scenes. However, when applying vision-language models like CLIP to pixel-level segmentation, the deep layers of its Vision Transformer generate numerous Semantic Occupancy Tokens (SOT), degrading the spatial structure of visual features. This further exacerbates the optimization conflict between semantic discrimination and pixel-wise localization, becoming a critical bottleneck that limits fine-grained segmentation performance. To address this, we propose CT-CLIP, a novel framework for class-guided camouflaged object detection that calibrates SOT to restore spatial structure. At its core lies the Multi-stage Attention Calibration Module, which sequentially mitigates SOT-induced spatial degradation through three synergistic operations: SOT Reallocation redistributes attention from SOT to object regions; Hierarchical Attention Fusion adaptively aggregates rectified multi-layer attention maps to enhance last-layer spatial perception; and DINO-Guided Augmentation injects self-supervised attention from DINO as fine-grained structural priors, achieving high-fidelity attention reconstruction. Building on the restored features, a Semantic Query Decoder is devised to reformulate segmentation as a semantic-guided visual querying process, where text embeddings serve as queries, and the concatenation of CLIP visual features with DINO-enhanced features serves as keys and values in cross-attention, thereby progressively aligning semantics with spatial structures. Experimental results on the CGCOD benchmarks demonstrate that CT-CLIP significantly outperforms existing methods.

Abstract:
Most existing text-video retrieval methods project both text and video into a shared latent semantic space for direct similarity calculation. However, these approaches often overlook the rich multi-modal information in videos and fail to fully capture the textual semantic content. To address these limitations, we propose a novel coarse-to-fine Multi-Modal Diffusion (MMD) Model for query-aware text-video retrieval. Our model introduces three key components: query-aware text-video cross-modal matching, fine-grained multi-modal diffusion matching and caption-assisted matching driven by multi-modal large model. First, MMD enhances text-video matching by focusing on audio-augmented multi-modal information in the video and utilizing multi-modal large model to generate video captions. Second, a fine-grained multi-modal diffusion model with cross-modal attention mechanism for adaptive multi-modal fusion is proposed to reason about fine-grained multi-modal text-video matching. The proposed multi-modal diffusion model models cross-modal retrieval tasks as a process of gradually generating joint distributions from noise through a framework based on multi-modal propagation. Third, we design a two-stage inference method which elegantly balances retrieval accuracy and speed. During inference, query-aware text-video cross-modal matching and caption-assisted matching are used to obtain the top K candidate samples quickly. Fine-grained multi-modal diffusion matching is applied to rerank only the top K samples. Extensive experiments on three benchmark datasets demonstrate the superior effectiveness of the proposed model compared to state-of-the-art approaches.

Abstract:
Collaborative perception extends the sensing range of a single vehicle by fusing observations from nearby agents, which improves the robustness of autonomous driving. In realistic deployments, however, the received collaborator messages are often affected by both communication delay and relative-pose noise, which jointly cause stale observations, spatial misalignment, and unstable feature fusion. Existing methods usually address these issues from either the spatial or temporal side, but handling them jointly in a unified and efficient manner remains challenging. In this paper, we propose CoAnchor, an anchor-centric spatio-temporal alignment framework for asynchronous collaborative perception. Instead of directly reasoning on dense BEV features, CoAnchor builds sparse object-level spatio-temporal anchors as a shared interface for pose correction and tightly connects spatial refinement, temporal propagation, and current-time verification within one unified loop, while keeping the overall correction process lightweight. Extensive experiments on both simulated and real-world datasets illustrate that CoAnchor remains competitive under clean settings and improves the robustness under joint delay and pose perturbations with a favorable practical accuracy-efficiency trade-off.

Abstract:
Are current Vision Language Models (VLMs) ready to comprehend and reason about complex embodied interactions in 3D environments? We introduce Embodied3DBench, a robot-centric benchmark targeting low-level spatial intelligence in embodied 3D environments. To systematically evaluate these foundational perceptual capabilities, the benchmark includes 6 task categories divided into two core groups: Spatial Structural Understanding (Grounding, Spatial Relation Prediction, and Multi-view Correspondence) and Interaction-Oriented Perception (Affordance Prediction, Grasp Point Prediction, and Trajectory Prediction). The benchmark spans 12 subcategories and contains over 21k high-quality question-answer pairs. We evaluate 13 state-of-the-art models, and the results show that while current models exhibit relatively strong high-level spatial reasoning, such as understanding object-to-object positional relations, they remain fragile in interaction-oriented perception, highlighting a significant lack of robust 3D-aware interaction priors. To actively bridge this capability gap revealed by our benchmark, we further synthesize a large-scale training dataset comprising 1.3M QA pairs. Notably, fine-tuning on this dataset yields significant improvements in low-level spatial intelligence. Ultimately, Embodied3DBench fills a critical gap by providing both a systematic evaluation framework and a scalable data solution, setting a clear target for the development of interaction-aware multimodal systems.

Abstract:
Despite the progress of multimodal large language models (MLLMs), they continue to exhibit deficiencies in visual perception. Following visual instruction tuning, internal MLLM representations rapidly deviate from their original semantic states during inference, causing severe information degradation. While existing methods attempt to leverage external vision foundation models (VFMs) to align internal representations, we find that direct alignment with VFMs enhances visual semantics but fails to mitigate representation deviation. To address this, we propose Spatial-Spectral Visual Anchor Learning (SSVAL). The core of SSVAL is Visual Anchor Prompt Injection (VAPI), which introduces prompts that absorb rich knowledge from external VFMs during training, enabling them to serve as stable visual anchors that mitigate representation deviation during inference. Additionally, we incorporate auxiliary spatial and frequency-domain representation alignment losses to provide complementary vision-specific supervision at intermediate LLM layers. Extensive experiments demonstrate that SSVAL significantly outperforms existing methods.

Abstract:
Multimodal Sentiment Analysis with Missing Modalities (MSA-MM) infers sentiments when multimodal inputs suffer from random token missingness. Such missingness, arising from noise, sensor failures, or privacy constraints, fundamentally introduces semantic ambiguity, hindering accurate token disambiguation and multimodal fusion. Existing approaches to MSA-MM usually rely on single-point embeddings to represent missing tokens. They fail to explicitly capture and resolve fine-grained semantic ambiguity, limiting multimodal fusion and downstream performance. To address this, we propose Uncertainty-Aware Token Disambiguation (UATD), a unified probabilistic framework that models, disambiguates, and leverages semantic uncertainty for robust MSA-MM. Specifically, UATD models each missing token as a Gaussian mixture distribution to preserve multiple plausible semantic hypotheses instead of collapsing it into a single embedding. It then determines mixture weights based on component priorities, suppressing ambiguous hypotheses while emphasizing reliable ones for effective disambiguation. Furthermore, it calibrates both intra- and cross-modal representations to ensure semantic consistency, and injects this calibrated uncertainty into cross-modal attention for robust multimodal fusion. Extensive experiments on three challenging datasets across various missing rates demonstrate that UATD consistently achieves state-of-the-art performance. Moreover, qualitative visualizations show that UATD learns more discriminative representations.

Abstract:
Aligning large language models with fine-grained human emotions remains difficult because high-quality affective data are scarce, synthetic self-training can reduce distributional diversity, and standard autoregressive generation lacks explicit mechanisms for response verification and multi-turn planning. We present DPE-Align, a dual-process framework that separates response generation from iterative emotional verification and retrieval-augmented strategic planning. DPE-Align combines an external Critic with a Strategic Action Graph that supports conditional branching, rollback, and dynamic re-planning. We also introduce \textscFineEmo-12k, a research-oriented corpus of 12,350 multi-turn dialogues covering 35 fine-grained emotion sub-categories. In a blind realism study, experts identified the filtered synthetic dialogues as synthetic in 55.5% of judgments, which was not significantly different from chance (p=0.12, binomial test). On the held-out \textscFineEmo-12k test set, DPE-Align, using an 8B Generator and a locally deployed Qwen2.5-72B-Instruct Inference Critic, achieves higher task-specific fine-grained accuracy than zero-shot GPT-5 (84.2% vs.\ 81.5%), while GPT-5 retains a higher general response-level Win-Rate (68.5% vs.\ 66.8%). DPE-Align also maintains consistent improvements across three external benchmarks. Ablations attribute 60.3% of the improvement over the strongest domain-specialized baseline to the System~2 architecture, while 97.9% of the maximum accuracy improvement is retained when GPT-5 is replaced by Qwen2.5-72B-Instruct as the Inference Critic.

Abstract:
Recent 3D shape understanding studies commonly transfer semantics from 2D vision-language models through global contrastive alignment. While effective, this paradigm remains brittle for sparse, noisy, or semantically ambiguous point clouds. A natural remedy is to supplement parametric alignment with external semantic evidence. However, naive retrieval of such evidence often returns noisy or weakly relevant examples, which further degrades recognition. To address these issues, we present KAR3D, a knowledge-augmented retrieval framework that equips 3D recognition with an external multimodal memory, enabling the model to retrieve, rerank, and fuse trustworthy priors. KAR3D first performs hierarchical cross-modal retrieval to narrow candidates with text semantics and point-cloud geometry. It then refines them via reward-driven semantic reranking and injects the selected priors through confidence-aware fusion, allowing the model to fall back to geometric cues when retrieval is unreliable. In this way, KAR3D enables robust retrieval-augmented 3D representations that exploit external semantic evidence without being dominated by noisy priors. Experiments on ModelNet40 and ScanObjectNN show that KAR3D brings consistent improvements in zero-shot, few-shot, and fine-tuned 3D classification, while qualitative cross-modal retrieval examples further illustrate its ability to establish semantic matches across modalities. These results point to retrieval-augmented, reference-based reasoning as a promising paradigm for scalable 3D shape understanding and demonstrate significant improvements over existing baselines.

Abstract:
Multimodal large language models (MLLMs) have achieved impressive performance in multimodal emotion recognition (MER) tasks and lifted MER to a new level that is complex emotion understanding with advanced video understanding abilities and natural language description. However, existing MLLM-based methods often use a fixed prompt to perceive the emotions, ignoring the dynamicity and complexity of the emotion source in the multimodal inputs. To address these issues, we propose a novel Reinforcement Learning-based Dynamic Agent Specialization framework (EmoAgent-R1) to optimize the emotion recognition, reasoning, and generalization abilities of an MLLM with dynamic agent specialization based on reinforcement learning. Specifically, we first adopt a cold start strategy to endow an MLLM with preliminary emotion recognition, reasoning, and agent routing ability by training with synthetic answer-conditioned chain-of-thought data and agent routing data. Then, we further train the MLLM with reinforcement learning to perceive emotions in a two-step agentic workflow with agent selection and agent specialization. To effectively train EmoAgent-R1, we propose a novel Progressive Group-Relative Policy Optimization (P-GRPO) to combine group-based relative advantages with a PMI-inspired progressive token-level modulation to transform sparse rewards into fine-grained learning signals, mitigating the coarse-grained uniform credit assignment issue in GRPO. Extensive experiments on MER benchmarks demonstrate the superiority of our EmoAgent-R1 in stronger emotion reasoning performance and improved optimization stability.

Abstract:
\beginabstract With the rapid development of generative models such as diffusion models, the quality of image generation has improved significantly, posing significant challenges to the distinguishability of forged images while also bringing concerns about misuse. Based on the perspective from signal theory that phase carries structural information, this study proposes a phase feature-dominated generated image detection framework: Reconstruction Error Phase Structure Response (REPSR). The framework employs an encoder to reconstruct images, thereby amplifying phase structural differences, and performs phase frequency band analysis on both the original image and reconstruction residuals to construct multi-dimensional phase features for lightweight classifier discrimination. Experimental results demonstrate that the proposed method possesses clear physical meaning and interpretable visualization pathways, exhibits stable cross-domain generator generalization, and features a lightweight backend detection module with only 0.56M online parameters (the VAE reconstructor can be precomputed offline), making the online detection stage suitable for deployment in resource-constrained scenarios. \endabstract

Abstract:
Continuous Spatio-Temporal Video Super-Resolution (C-STVSR) aims to simultaneously enhance the spatial resolution and frame rate of videos by arbitrary scale factors, offering greater flexibility than fixed-scale methods that are constrained by predefined upsampling ratios. In recent years, methods based on Implicit Neural Representations (INR) have made significant progress in C-STVSR by learning continuous mappings from spatio-temporal coordinates to pixel values. However, these methods fundamentally rely on dense pixel-wise grid queries, causing computational cost to scale linearly with the number of interpolated frames and severely limiting inference efficiency. We propose GS-STVSR, an ultra-efficient C-STVSR framework based on 2D Gaussian Splatting (2D-GS) that drives the spatiotemporal evolution of Gaussian kernels through continuous motion modeling, bypassing dense grid queries entirely. We exploit the strong temporal stability of covariance parameters for lightweight intermediate fitting, design an optical flow-guided motion module to derive Gaussian position and color at arbitrary time steps, introduce a Covariance resampling alignment module to prevent covariance drift, and propose an adaptive offset window for large-scale motion. Extensive experiments on Vid4, GoPro, and Adobe240 show that GS-STVSR achieves state-of-the-art quality across all benchmarks. Moreover, its inference time remains nearly constant at conventional temporal scales (× 2--× 8) and delivers over 3× speedup at extreme scales (× 32), demonstrating strong practical applicability.

Abstract:
\beginabstract Underwater instance segmentation has made significant progress; however, segmenting novel categories from extremely few reference samples remains challenging. Traditional fine-tuning methods and foundation models such as the Segment Anything Model (SAM) still suffer from feature misalignment, instance adhesion, and localization failures under underwater visual degradation. To address these issues, we propose UW-ProSAM, which re-examines this task from the perspective of prototype learning. We transform the challenge of one-shot novel-category segmentation into a cross-domain style alignment problem coupled with the dynamic extraction of degradation-robust ``semantic invariants.'' Specifically, the Style-Aligned Prototype Refinement (SAPR) module performs pair-specific frequency-domain style calibration and reverse-verified prototype refinement to obtain robust foreground and background representations. The Uncertainty-Guided Instance Separation (UGIS) module models spatial heteroscedastic uncertainty and reweights clustering distances to separate adhered instances into reliable point prompts for the frozen SAM decoder. Under the identical 1-shot setting on UIIS10K Fold~0 and USIS10K Fold~0, UW-ProSAM achieves 16.7% and 16.2% mAP, respectively, outperforming matched 1-shot baselines. It also remains competitive with strong 5-shot and 10-shot fine-tuning methods while requiring only one annotated support image. \endabstract

Abstract:
Diffusion models have become a core paradigm for multimedia generation, offering powerful concept-driven controllability for personalization, semantic editing, and selective unlearning. However, as semantic control extends beyond natural-language prompts to learned embeddings and intervention pipelines, the safety and governance of these systems become increasingly difficult to evaluate in a unified manner, especially for safety-sensitive, identity-linked, and other privacy-relevant concepts. Existing studies mainly rely on heuristic audits, adversarial probing, or task-specific erasure benchmarks, and therefore provide limited support for systematic comparison across models, conditioning channels, and deployment conditions. We present a concept-level probabilistic audit and reporting framework for diffusion models. We formalize governance-relevant concept behaviors as Bernoulli semantic events induced by stochastic generation, and define a Concept Risk Operator that maps model-channel configurations to structured risk profiles, enabling comparison across prompting interfaces, learned embedding channels, models, and recorded conditions. We apply sample-level post-hoc calibration and configuration-level risk aggregation, and show that probability error can change thresholded actions near policy boundaries. Experiments on SD1.5, SD2.1, and SDXL reveal consistent yet non-uniform operational risk patterns across concept families, channels, recorded conditions, and shifted protocols. In particular, embedding-based access and obfuscated prompts expose risks often understated by standard-prompt evaluation. A pooled multi-protocol calibrator improves held-out probability reliability, but we do not claim transfer from a standard-only calibrator. CLRC provides a common audit schema for probabilistic and decision-aware governance of multimedia generation systems.

Abstract:
Fake news increasingly relies on cross-modal image-text forgeries, making transparent and verifiable reasoning chains an urgent need for Detecting and Grounding Multi-Modal Media Manipulation (DGM4). Existing methods produce black-box detection results without any decision rationale, limiting their reliability in forensic practice. Multi-modal Large Language Models (MLLMs) offer a natural path toward explainability, but applying them to DGM4 raises two difficulties. First, models tend to generate explanations disconnected from predicted evidence locations, producing unverified attribution. Second, enforcing evidence-conclusion consistency requires active optimization, yet uniform training signals fail to distinguish localization tokens from classification tokens, making multi-head joint training unreliable. We propose a multi-modal manipulation detector based on an Evidence-Grounded Forensic Reasoning (EFR) framework. EFR introduces an Anchor-and-Verify reasoning chain that enforces modality-isolated perception before cross-modal comparison, with conclusion coordinates as explicit anchors to which downstream evidence must spatially correspond. A verifiable reward system then enforces evidence-conclusion consistency during training, while a Modality-Decoupled Advantage (MDA) routing mechanism mitigats credit misassignment across prediction tasks. Experiments show that EFR achieves state-of-the-art performance while producing structured forensic reasoning records that explicitly bind explanations to evidence.

Abstract:
Industrial anomaly analysis requires more than anomaly presence prediction: the model must localize subtle defects and explain their causes and implications. Existing MLLM-based solutions often rely on an external anomaly detector to provide visual cues, yet this expert-aid paradigm is brittle because the cues are threshold-sensitive, potentially noisy, and can suppress broader contextual reasoning. We propose OmniAD, a end-to-end framework that intrinsically couple anomaly detection and anomaly understanding within a single multimodal reasoning process. OmniAD first performs visual reasoning through Semantic Anomaly Encoding, which converts anomaly regions into a compact textual representation and removes the need for thresholded heatmaps. It then performs Visual Guided Textual Reasoning, where the internally generated anomaly evidence guides high-level analysis without introducing an external attention bias. To improve learning under limited industrial data, we adopt joint post-training with Supervised Fine-Tuning and Group Relative Policy Optimization using format, detection, and answer rewards. OmniAD achieves 79.9% average accuracy on MMAD, outperforming strong open-source and proprietary MLLM baselines, and also delivers robust threshold-free performance on four industrial anomaly detection benchmarks. These results suggest that strengthening intrinsic perception within MLLMs is a practical route toward accurate and interpretable industrial anomaly reasoning.

Abstract:
Token pruning has emerged as an effective approach to reduce the substantial computational overhead of Large Vision-Language Models (LVLMs) by discarding less informative visual tokens while preserving performance. However, existing methods typically rely on individual attention sources from different LVLM components, resulting in incomplete and suboptimal pruning decisions due to biased attention distributions. To address this, we propose DeSAP, a novel Decoupled Similarity-Aware Pruning method for precise, task-aware visual token pruning. Specifically, DeSAP introduces a decoupled similarity to capture fine-grained cross-modal relevance between visual and text tokens, providing explicit task-related guidance for pruning. By integrating this similarity with visual saliency signals derived from visual attention, DeSAP performs token pruning under the guidance of both task-related and visual cues, enabling robust pruning even under aggressive pruning ratios. Extensive experiments across diverse benchmarks show that DeSAP consistently outperforms existing methods in both accuracy and efficiency.

Abstract:
Federated cross-modal retrieval is particularly challenging in heterogeneous settings, where clients may contain only images, only text, or paired image-text data. Such modality heterogeneity leads to incomplete local supervision and weakens the preservation of the shared cross-modal similarity structure that reliable retrieval depends on. Existing federated optimization methods therefore often suffer from representation drift across clients and fail to maintain coherent retrieval semantics at the global level. To address this issue, we propose FedCoPA, a federated framework for modality-heterogeneous cross-modal retrieval. FedCoPA leverages a frozen vision-language model as a shared teacher and transfers its semantic structure to distributed clients through local teacher-space transfer and server-side prototype anchoring. Specifically, on the client side, FedCoPA performs modality-specific alignment for all clients and further incorporates relational distillation and contrastive learning on multimodal clients to preserve retrieval-relevant local structure. On the server side, it maintains modality-specific prototype memories in the teacher space, which serve as persistent global anchors beyond isolated local batches. Experiments on four datasets demonstrate that FedCoPA consistently improves cross-modal retrieval performance under heterogeneous federated settings.

Abstract:
VIS-NIR image fusion aims to produce RGB outputs that preserve visible appearance while incorporating complementary structure from near-infrared imagery. Existing RGB-output methods typically rely on fixed color decompositions, applying the same transfer geometry to all regions and materials. This uniform treatment cannot adapt to local VIS-NIR relationships and often leads to color distortions. We propose Adaptive Color Space Learning (ACS), which learns a local color basis at each pixel and fuses only a structure-aligned transfer coordinate while preserving the complementary appearance coordinates for reconstruction. The method keeps the RGB reconstruction path explicit and uses projection-guided multi-scale fusion to increase the NIR contribution where the visible transfer coordinate is less reliable. Experiments and controlled analyses on RGB-NIR Scene and NAID show that ACS gives better appearance fidelity and structural transfer than the compared methods while keeping the trainable part compact.

Abstract:
Learning-based remote photoplethysmography (rPPG) effectively suppresses motion artifacts in non-contact physiological measurements but suffers from performance degradation due to domain shifts, demographic diversity, and dynamic illumination. These challenges are further complicated by the fact that diverse populations and acquisition devices require different optimal penetration wavelengths. Conventional feature engineering cannot fully overcome these gaps. To address these issues, we propose PhysFVN, a collaborative optimization framework integrating the vision-language model (VLM) with dual-modal visible and near-infrared rPPG technology. Specifically, we first leverage the VLM to extract demographic features like skin tone, age, and ethnicity, along with environmental context features such as scene types, lighting conditions, participant states, and facial occlusion from the visible video. These features are combined with hemodynamic data from the backbone network and projected into the semantic space of a large language model (LLM), which captures long-range temporal dependencies and enables cross-modal alignment. Simultaneously, we incorporate textual descriptions alongside the extracted features when feeding into the LLM, further enhancing task comprehension. Building upon these, we design a wavelength-adaptive gating mechanism that dynamically fuses visual and near-infrared contributions based on the physiological features, allowing automatic selection of the optimal wavelength for different demographics and scenarios. Additionally, a contrastive learning strategy enhances quasi-periodic signal mining by aligning multi-spectral channels within each video based on attribute matrices. Experiments show that PhysFVN achieves superior robustness under cross-subject, cross-scenario, and lighting variations.

Authors: Peiyuan Jiang (University of Electronic Science and Technology of China), Yao Liu (University of Electronic Science and Technology of China), Yanglei Gan (University of Electronic Science and Technology of China), Jiaye Yang (University of Electronic Science and Technology of China), Khwaja Mutahir Ahmad (Southwest Jiaotong University), Zhenlong Liao (Civil Aviation Flight University of China), Lu Liu (University of Electronic Science and Technology of China), Xuefeng Peng (University of Electronic Science and Technology of China), Yuewei Xue (University of Electronic Science and Technology of China), Daibing Yao (Yizhou Prison), Qiao Liu (University of Electronic Science and Technology of China)

Abstract:
Non-contact deception detection remains challenging because visual and auditory deception cues often lack stable cross-subject patterns. To improve non-contact deception detection, we explore cross-modal knowledge distillation (CMKD) from galvanic skin response (GSR), which provides more reliable physiological evidence of deception, to guide visual and auditory representation learning. To enable this study, we introduce MuDD, a large-scale multimodal deception detection dataset collected under the Guilty Knowledge Test (GKT) paradigm, with recordings from 130 participants spanning 690 minutes across six modalities. Building on MuDD, we propose GSR-guided Progressive Distillation (GPD), a cross-modal distillation framework that models knowledge transfer as a dynamic process. Unlike existing CMKD methods that typically rely on predefined distillation schemes, GPD adaptively regulates both the transfer pathway and the transferred knowledge as the modality gap evolves. Specifically, GPD uses gap-aware dynamic routing to select distillation configurations based on the evolving representational gap between teacher and student, and progressively adjusts the relative importance of feature-level and logit-level knowledge during training for stable cross-modal transfer. Extensive experiments and visualizations show that GPD outperforms existing methods and achieves state-of-the-art performance on both deception detection and concealed-digit identification.

Abstract:
Autonomous teams increasingly pair cameras, LiDAR, and microphones, yet existing protocols offer no formal guarantee of how much communication coordination requires. We introduce the Multimodal Coordination Game and prove a worst-case lower bound for two-agent teams: at sufficiently fine precision, every protocol achieving near-optimal coordination must exchange bits growing linearly in the intrinsic task dimension and logarithmically in the inverse target precision. The constant is geometric rather than architectural, governed by the fraction of state coordinates the agents do not share. Under the coordinate-observation model this makes the requirement sensitive to sensor type: spatial--temporal pairings face a strictly stronger bound than same-domain pairings of equal dimension, verified empirically for camera--audio against camera--LiDAR. When a constant fraction of observation dimensions are non-shared, a matching upper bound recovers the same scaling up to constants and additive dimension- and smoothness-dependent terms. The price of modality heterogeneity (PMH) shows when diversity pays: under complementarity and a single-modality homogeneous reference, diverse teams outperform uniform ones above a bandwidth threshold. Guided by these results, the Equilibrium-Guided Communication Protocol allocates bandwidth in proportion to each agent's task relevance, weighted by its modality's marginal information contribution, cutting communication 38--45% on CARLA and AVE-Multi while improving coordination on all four.

Abstract:
Large-scale adaptive bitrate streaming uses bitrate ladders---discrete sets of transcoding levels that determine how platforms serve diverse users under dynamic networks. Yet ladders are typically static and tuned for average workloads, making them poorly suited to live-streaming dynamics. Because content complexity and audience populations vary substantially across channels and over time, static ladders either over-provision resources or under-serve viewers. At platform scale, even small inefficiencies create significant cost and user impact. Real-time ladder optimization is therefore challenging: it must search a large encoding space while jointly optimizing QoE and infrastructure cost within milliseconds. We present ATL, a real-time Adaptive Transcoding Ladder system for large-scale live streaming. ATL formulates ladder selection as a submodular maximization problem that captures diminishing quality returns and over-provisioning penalties. As a closed-loop system, ATL monitors content and audience dynamics to adapt ladders without disrupting in-flight sessions. A production deployment on one of the largest live streaming platforms demonstrates gains over the production baseline: 5.71% QoE improvement and 11.66% CDN cost reduction. Compared with state-of-the-art methods, ATL achieves 6.0% higher QoE at 7.3% lower bitrate, while reducing decision latency from 2.2 s to 130 ms.

Abstract:
Enhancing the robustness of ``encoder--noise layer--decoder'' architectures remains a key challenge in deep learning--based watermarking. Existing methods typically rely on joint training with combined noise layers, which implicitly forces the model to compromise among conflicting optimization objectives induced by different distortions. In this paper, we reformulate multi-distortion robustness as a \emphtask composition problem and propose LEGO, a modular robustness enhancement framework based on hierarchical task vector composition. Specifically, we introduce a mechanism-driven distortion taxonomy that partitions distortions into five domains according to their degradation mechanisms, ensuring that distortions within each domain share compatible optimization directions. With the watermark embedder fixed, we optimize only the extractor and design a hierarchical fine-tuning strategy to learn domain-level task vectors capturing shared robustness patterns and distortion-specific task vectors for fine-grained adaptation. Benefiting from this divide-and-conquer task vector composition design, the watermark extractor achieves improved robustness to multiple distortions while supporting seamless and incremental extension to new distortion types. Extensive experiments demonstrate that LEGO is efficient and flexible, achieving superior or comparable robustness to state-of-the-art methods across diverse distortion settings.

Abstract:
Achieving precise audio-visual synchronization remains an important challenge in joint audio-video generation. Despite advances in generative models, current diffusion-based frameworks persistently struggle to maintain strict spatiotemporal alignment between generated speech and lip movements. Our analysis attributes this misalignment to two primary factors. First, prevalent frameworks typically rely on serial cross-attention mechanisms to fuse multimodal conditions, sequentially fusing text and cross-modal conditions inevitably leads to bidirectional feature entanglement, making it difficult to decouple fine-grained synchronized local dynamics from global semantics and to effectively utilize the prior knowledge of pre-trained model. Second, existing methods lack joint synchronization constraints across both time and space, typically optimizing for either temporal or spatial alignment in isolation, thereby failing to achieve precise spatiotemporal lip-sync. To address these issues, we propose SyncDiT with Parallel Decoupled Cross-Attention mechanism. By decoupling text-driven semantic conditions and cross-modal dynamic conditions into independent parallel streams, our model achieves decoupling control and protects the prior space of the model. To enforce precise synchronization, we introduce a Spatiotemporal Synchronization Mechanism. Temporally, our proposed chunk-wise attention restricts feature interactions to localized windows. Spatially, we introduce a noise-aware latent sync guidance mechanism, which guides the synchronization of facial regions and audio through a trained sync estimator. Finally, to mitigate the unstable convergence and mutual interference inherent in joint generation, we implement a Multi-Task Progressive Training Strategy, anchoring the model with unidirectional priors before joint fine-tuning. Extensive experiments demonstrate that our approach significantly outperforms existing methods across visual fidelity, speech quality, and spatiotemporal synchronization.

Abstract:
Despite the remarkable success of multimodal joint training, severe modality imbalance often leads fast-converging dominant modalities to dictate the global optimization trajectory, leaving lagging modalities in a chronically under-optimized state. Existing debiasing methods typically rely on zero-sum static suppression strategies, which are highly prone to disrupting optimization continuity and sacrificing discriminative features within the dominant branches. To overcome this limitation, inspired by the interaction mechanism between the leader and followers in the Stackelberg game model, we propose a dynamic gradient routing framework that transforms traditional inter-modal competition into an asymmetric leader-follower synergy. Modeling the fusion backbone as the Leader and individual unimodal branches as Followers, our method introduces a state-aware factor to characterize the relative optimization state of each branch, coupled with a temporal alignment constraint to limit excessive directional deviation during updates. By dynamically computing branch weights for gradient reconstruction, the framework routes globally informed gradients to lagging modalities while avoiding direct suppression of dominant ones. Under the unified BalanceBenchmark protocol, extensive experiments conducted across four benchmark datasets (Kinetics-Sounds, CREMA-D, CMU-MOSEI, and VGGSound) demonstrate that our method consistently outperforms existing mainstream debiasing baselines, establishing a new paradigm that shifts multimodal optimization from passive suppression to active synergistic routing.

Abstract:
\beginabstract Embodied AI agents rely on multi-modal perception for situational awareness in complex urban environments. While visual simulation has achieved photorealism, generating physics-grounded acoustic datasets remains a critical bottleneck. Existing Geometric Acoustics (GA) renderers, though efficient for high frequencies, often yield null signals in Non-Line-of-Sight (NLoS) blind spots because they cannot resolve low-frequency wave diffraction. To bridge this gap, we propose a parallel, voxel-based acoustic benchmark generation engine powered by a mesoscopic Lattice Boltzmann kernel. Scalable to the hundred-million-voxel level on supercomputers, the framework captures 3D wave dynamics and continuous multipath healing within an explicitly bounded low-frequency band (up to approximately 100 Hz in the present benchmark) and under rigid structural-boundary assumptions, while circumventing the meshing bottleneck of traditional solvers. We validate the engine through a multi-resolution ablation study (grid spacing from 1.5 m to 0.5 m) in a realistic urban intersection. Leveraging these fluid-dynamic data, we establish a diagnostic benchmark spanning GCC-PHAT, MUSIC, and a pre-trained Deep Learning architecture (DL-CRNN), quantifying the out-of-distribution shift between geometric assumptions and wave-based transients. Millisecond-level analysis exposes spatial-aliasing failure modes in these representative localization paradigms. Accordingly, the simulated fields serve as a validated low-frequency hard-boundary wave reference, rather than full-band real-world acoustic ground truth, for diagnosing NLoS perception algorithms and developing wave-informed spatial-audio models. \endabstract

Abstract:
Data-free model merging offers an attractive way to consolidate multiple Multimodal Large Language Models (MLLMs) into a single system without retraining, reducing deployment cost while potentially combining complementary multimodal capabilities. However, existing heterogeneous MLLM merging methods primarily focus on mergeability under architectural heterogeneity while overlooking a more fundamental issue: preserving multimodal alignment. Here, multimodal alignment refers to the coordination between visual encoding and language reasoning encoded in language-side parameters. As a result, source information may be transferred along incompatible directions, silently disrupting vision--language coordination even when local parameter matching appears reasonable. To address this challenge, we propose APT-Merging, a strictly data-free alignment-preserving knowledge transfer framework for heterogeneous MLLM merging. APT-Merging comprises three key components: (1) a subspace extraction strategy based on truncated SVD that exposes the alignment-critical directional structure of the original model; (2) a dynamic mapping mechanism that aligns source components with compatible directions in the original model while filtering directionally incompatible information; and (3) a protective injection scheme that transfers only reliability-gated positive surplus along singular directions supported by the original model, thereby preventing off-subspace perturbations and reducing negative transfer. Extensive experiments on two heterogeneous MLLM settings and eight vision-language benchmarks demonstrate that APT-Merging achieves the strongest overall performance among compared data-free baselines. Further ablations show that alignment preservation is the most influential factor in achieving robust heterogeneous MLLM merging.

Abstract:
When labels are corrupted by noise, cross-modal hashing often fails to maintain reliable semantic neighborhoods and clear decision boundaries in the Hamming space. Existing robust hashing methods mainly improve the reliability of samples or pairwise supervision, while the class-level semantics contained in multi-label annotations are not explicitly retained in conventional binary relations. We propose Semantic Guidance Boundary Aware Hashing (SGBH), which exploits label information at both class and relation levels for robust hash learning. Semantic Compass constructs class-level semantic anchors from active category names and transfers category-relevant cues to multimodal hash representations through gated cross-attention and one-way guidance. On the resulting representations, Adaptive Boundary Formation calibrates mini-batch-specific positive and negative boundaries to emphasize hard pairs, while Noise-Tolerant Semantic Alignment reduces the influence of corrupted supervision through cross-modal neighborhood reliability and graded label-overlap relations. Experiments on MIRFLICKR-25K, NUS-WIDE, and MS COCO demonstrate competitive retrieval performance across different code lengths and consistent robustness under varying levels of label noise.

Abstract:
Video-based misinformationassembled from authentic footage, fabricated captions, and manipulated speechspreads rapidly across platforms, yet existing multimodal detectors typically produce only a binary verdict or a post-hoc rationale without determining which specific element across video, speech, and text is responsible for making the narrative false, or to what degree. Current approaches treat misinformation analysis as a single-step classification augmented with explanations, collapsing the structured, multi-step reasoning that human fact-checkers perform into a single inference pass. We formalize LLM-Driven Evidence-Grounded Multimedia Misinformation Attribution (L-EMMA), a task that jointly requires veracity classification, fine-grained falsity grounding, evidence-backed provenance chains, and quantified interventional responsibility scores. To address L-EMMA, we propose ProgV-Trace, a framework in which a large language model generates an executable verification programa typed operator sequence in a domain-specific languagethat decomposes claims, orchestrates evidence retrieval, dynamically expands a temporal-multimodal provenance graph, and plans minimal counterfactual interventions to isolate each element's contribution to the misinformation verdict. A program-to-student distillation stage transfers the LLM's structured reasoning into a compact verifier for deployment without LLM inference. We further introduce a Program-Hard Challenge Protocol with fresh, semantically coordinated, and multi-step-verification samples that stress-test generalization beyond standard benchmarks. Extensive experiments on GroundLie360, FakeVV, and FakeSV demonstrate that ProgV-Trace achieves 87.2% F1 and 71.4% Resp-NDCG, surpassing the strongest baseline by +4.5 F1 and +24.2 Resp-NDCG pointsof which a same-model ablation confirms +21.8 points stem from the interventional attribution method itself. The distilled student retains over 96% attribution quality at 196 ms per sample, and responsibility scores agree with human judgments 84.9% on average across five manipulation types.

Abstract:
Interactive text-based person retrieval aims to resolve ambiguity in an underspecified initial query through iterative clarification. However, existing methods largely rely on free-form dialogue, which makes interaction depend on unstructured language generation rather than directly selecting the most discriminative visual evidence among competing candidates. As a result, acquired evidence is only implicitly encoded in dialogue history, remaining uncertainty is not explicitly represented, and generated questions are often weakly grounded in the visual differences that determine ranking. This limitation is further amplified by the long-tailed distribution of visual attributes: frequent attributes are easy to ask about but quickly become weakly discriminative, whereas long-tail attributes are often more informative for disambiguation yet underutilized. To address this, we reformulate interactive retrieval as a retrieval-state-conditioned evidence acquisition problem. We introduce a grounded discrete concept space composed of reusable concept atoms, which provides a shared representation for the initial query, candidate-side evidence, and user feedback. At each interaction round, the system selects the next concept atom conditioned on the current candidate set using a redundancy-aware utility that jointly favors ambiguity reduction, candidate discrimination, and high-quality partitioning of the candidate pool. In our framework, natural language serves only as an interaction interface, while decision-making is performed over structured concept-level evidence. This formulation replaces unconstrained dialogue generation with structured, utility-driven clarification, enabling targeted and interpretable interaction without training additional question-generation or interaction-policy modules. Experiments on Interactive-PEDES under a unified multi-round protocol show that our method consistently outperforms language-centric baselines, delivering faster early-stage refinement and larger cumulative improvements across interaction rounds, with R@1 improving by 31.01% over five rounds.

Abstract:
Data-free knowledge distillation transfers the knowledge of a pre-trained teacher to a compact student without accessing the original training data. While this paradigm has been extensively studied for image classification, extending it to semantic segmentation remains particularly challenging, as dense prediction depends not only on object presence but also on spatial continuity, boundary structure, and scene-level context. Existing data-free methods, largely designed around image-level recognition, often provide proxy samples that are poorly aligned with these structural demands, resulting in fragmented predictions and severe performance degradation. Moreover, confidence-based sampling can be misleading in this setting, since out-of-distribution images may still elicit overconfident teacher responses through superficial local cues. In this paper, we present DFSS, a distribution-guided framework tailored for Data-Free Semantic Segmentation. Instead of relying on potentially unreliable output confidence, DFSS exploits the source-domain traces preserved in the teacher's Batch Normalization layers. Specifically, we propose Approximate Distribution Sampling, which ranks open-world images by their multi-layer feature-statistic discrepancy to the teacher and selects proxy samples that better approximate the latent source distribution. Building on this source-consistency order, we further introduce Weighted Distribution Progressive Distillation, a curriculum-style distillation strategy that emphasizes highly aligned samples at the early stage and gradually restores the contribution of harder, more diverse instances. Extensive experiments on NYUv2, CamVid, and Cityscapes show that DFSS consistently improves over competitive data-free distillation baselines, achieving stronger segmentation accuracy with fewer selected proxy samples.

Abstract:
Image harmonization aims to reduce illumination inconsistencies between foreground and background in composite image. Despite the success of diffusion models in image generation and editing, their applications to harmonization remains challenging for two issues: (1) existing datasets are predominantly synthetic and fail to capture complex real-world illumination, especially for portraits; (2) adjusting foreground illumination (low-frequency information) often degrades structural details and identity-related features (high-frequency information). To address these challenges, we construct a real-scene portrait harmonization dataset, termed RPHarmony, and propose a conditional diffusion-based method for portrait illumination harmonization. The proposed method disentangles low-frequency illumination and high-frequency structural information in the frequency domain and enables adaptive multi-frequency interaction via a frequency-aware fusion module. Besides, a mask-guided illumination cross-attention mechanism is introduced to model foregroundbackground relationships, while a self-supervised identity encoder is employed to preserve identity consistency. Extensive experiments on RPHarmony, IH-SG and iHarmony4 demonstrate that our method outperforms existing approaches in both quantitative metrics and visual quality, improving illumination consistency while preserving fine-grained details and identity.

Abstract:
Graphical User Interface (GUI) grounding is a fundamental perception task for multimodal agents, enabling them to interpret natural language instructions and interact with digital interfaces. Existing methods face a fundamental trade-off between accuracy and efficiency: direct full-image inference often fails to capture small or visually similar UI elements, while multi-crop strategies improve localization at the cost of multiple expensive Vision-Language Model (VLM) calls per query. To address this challenge, we propose RankGround, a two-stage framework that achieves accurate GUI grounding with a single VLM call per query. Central to our approach is GroundRanker, a lightweight multimodal reranker that identifies the most promising crop from a dense candidate set. Because no off-the-shelf ranking dataset is available, we construct ranking supervision data from existing grounding datasets. A strict containment criterion and boundary-aware positive augmentation improve alignment and spatial coverage in cluttered layouts. GroundRanker is then trained with a two-stage curriculum: a pointwise objective first learns coarse containment, and a listwise objective refines subtle semantic and spatial distinctions among visually similar crops. Experimental results show that RankGround consistently outperforms strong baselines while reducing computational cost. It achieves 1.4 faster inference and improves localization accuracy by 5.5% on average over the second-best method across all backbones and screen scales, establishing a new state of the art in both efficiency and precision for GUI grounding.

Abstract:
Audio-visual speech recognition (AVSR) relies on effective fusion of audio and visual modalities, yet existing approaches treat cross-modal interaction as a single-step operation without structured iterative refinement. We present DoubleHelix, a multimodal fusion framework that reformulates fusion as an iterative cross-modal interaction process with adaptive degradation-aware enhancement. The framework comprises three components including ReverseParallelHelix for multi-turn structured interaction with learned alignment constraints, QualitySensor for learning degradation-aware gating signals, and HelixReplication for consistency-guided conditional feature enhancement. Experiments on LRS3 demonstrate that DoubleHelix achieves 0.68% WER on clean audio, outperforming previous best results by 5.6% relative improvement under matched backbone settings. Comprehensive ablation studies validate each component contribution, including targeted analysis of design choices such as asymmetric pathway weighting. The framework shows improved robustness under evaluated babble-noise conditions, achieving 11.6% WER at SNR -5dB.

Abstract:
4D effects, often referred to as multisensory effects, have been widely integrated with audiovisual content to improve immersion and realism. Generating 4D effects in movies has been studied for over a decade, along with the rise of 4D cinema. Prior efforts mainly focused on matching haptic motion effects to audiovisual features, yet they do not account for individual differences and perception. To help this situation, we developed TaME4D to provide the most plausible 4D effects to individual users, an automated system that generates user-tailored 4D effects based on language-based preferences and descriptions. TaME4D exploited an orchestration of multiple AI models, such as computer vision and audio feature extraction to detect audiovisual features, and Multimodal Large Language Models (MLLMs) to infer the context and metaphoric meaning of the given scene. In addition, to reflect the individual preferences, the system retrieves text inputs from users, used to find the best-fit 4D effects. To show the effectiveness of TaME4D, we built a hardware setup comprising four multisensory rendering mechanisms of vibration, airflow, bass shaking, and thermal feedback in a comfort chair, and conducted a user study. The study results indicated that our user-tailored 4D effects were evaluated as more immersive and contextually appropriate.

Abstract:
3D scene graph generation requires understanding both object semantics and metric spatial relations. However, existing open-vocabulary methods often struggle with spatial predicates because language similarity reflects semantic co-occurrence rather than explicit geometry. We propose GHOST, a \underlineGeometry-constrained \underlineHybrid diffusion framework for \underlineOpen-vocabulary 3D \underlineScene graph genera\underlineTion. GHOST jointly models continuous node geometry and discrete edge predicates, following a geometry-to-constraint-to-predicate denoising process to ensure geometrically consistent relation prediction. It introduces language-grounded constraint induction, which maps predicate embeddings to differentiable geometric constraints that guide both layout refinement and predicate reweighting. A Geometry-Topology Collaborative Denoiser further couples node and edge reasoning through noise-adaptive interaction and orientation-aware modeling. Experiments on 3DSSG demonstrate state-of-the-art performance in both closed-set and open-vocabulary settings, including over 17-point gains over RelationField on spatial predicates.

Abstract:
Multimodal Large Language Models (MLLMs) are emerging as a foundation for automated front-end engineering because they enable joint reasoning over screenshots and natural-language instructions. However, most front-end assistants treat this workflow as a set of isolated code-generation tasks, directly decoding framework-specific syntax for each request. This direct-decoding paradigm leaves three limitations unresolved: (i) it provides no transferable prediction target across tasks and frameworks, (ii) it offers no stable representation for localized edits and repairs, and (iii) it provides no auditable path from multimodal input to executable outcomes. In this paper, we propose DeltaUI, a unified framework that recasts generation, editing, and repair as predictions of transitions in a shared, framework-normalized UI state space, modeling UI change rather than directly decoding target code. DeltaUI begins by parsing heterogeneous front-end artifacts into a typed UI state graph that separates shared UI semantics from framework-specific surface syntax. It then uses a multimodal transition model to predict structured state operations by encoding the source state and multimodal evidence (e.g., screenshots, instructions, and issues), preserving explicit source-to-update correspondence. A deterministic framework-conditioned realizer converts the resulting state into target-framework code. Finally, an execution-aware critic drives a bounded refinement loop using compiler and render feedback, ensuring the full path remains auditable. Extensive experiments on DesignBench and Design2Code demonstrate that modeling UI-state transitions yields consistent gains over direct code-decoding baselines across multiple MLLM backbones while retaining effective transfer to real-world webpages. The code and implementation of DeltaUI are provided in the supplementary material.

Abstract:
While Parameter-Efficient Fine-Tuning (PEFT) with vision-language foundation models (VLMs) achieves strong in-domain performance in federated learning (FL), generalizing VLM-based FL methods to unseen domains remains challenging. To address this, this paper focuses on FL in base-to-new settings, examining two key trade-offs: (1) the performance trade-off between seen and unseen classes and (2) the performance trade-off among different clients. This paper explores selective mask-based PEFT, a technique rarely used for VLM-based FL, to address these two challenges, leveraging its ability to effectively utilize the prior knowledge of VLMs and its flexibility for client-specific adaptation. Specifically, we propose a personalized Federated Mask-based Selective PEFT framework (pFedMSF). To enhance generalization to unseen classes, pFedMSF employs binary masks to select a small subset of tunable weights, preserving the prior knowledge of the pre-trained VLM. To address the performance trade-off among clients, besides global masks shared by all clients, pFedMSF further introduces local mask for personalized model weight update to individual client. Experimental results show that pFedMSF significantly outperforms the existing federated PEFT method on both seen and unseen classes, and achieves balanced performance on all clients.

Abstract:
Adapting CLIP for zero-shot sketch-based image retrieval (ZS-SBIR) via prompt learning faces a fundamental tension: the model must bridge the sketch-photo domain gap through task-specific adap- tation, yet the added flexibility risks overfitting to seen training categories and eroding CLIPs zero-shot generalization. We present SeCo-SBIR, a semantically consistent prompt learning framework that resolves this tension from both sides. First, a text-guided multi- modal prompting strategy routes learnable prompt vectors through CLIPs text encoder and projects the resulting intermediate rep- resentations into the visual encoder at every layer via learnable coupling functions. Because the text encoder has already learned robust, abstract category-level semantics from large-scale language supervision, this mechanism injects transferable semantic knowl- edge directly into the visual pathway - adapting the model to the sketchphoto domain while inherently favoring generalization to unseen classes. Second, a perturbation-based consistency constraint addresses the residual overfitting risk from the learnable coupling functions by aligning the adapted model with a frozen CLIP refer- ence branch using an asymmetric InfoNCE objective - augmented inputs feed the frozen branch while clean inputs feed the train- able branch - anchoring the learned representations to CLIPs gen- eralizable feature space. Together with lightweight adapters and a multi-objective loss combining triplet, NT-Xent, and classifica- tion terms, SeCo-SBIR achieves state-of-the-art results on all three standard ZS-SBIR benchmarks across categorical, generalized, and across-dataset settings

Abstract:
The crucial challenge in 4-dimensional micro-expression recognition (4DMER) lies in modeling the subtle motion between point cloud frames. Existing methods primarily transform 3D point cloud sequences into 2D representations to adapt to well-established 2D learning paradigms, thereby resulting in the loss of 3D information. Thus, this paper explores a novel point cloud-based 4DMER paradigm that represents micro-expression (ME) motions from 3D point cloud sequences, and proposes a local-to-global Point Micro-Expression Dynamic Network (PMEDNet). PMEDNet models local point clouds based on facial anatomy priors of ME-related regions and establishes correlations between local point cloud features to represent frame-level spatial features. Furthermore, existing ME dynamic modeling methods adopt an adaptive learning way guided by emotion-level supervision, which has limited interpretability and makes it difficult to accurately perceive micro-motions from 3D sequences. Therefore, based on the low-rank characteristics of facial muscle micro-motions during ME evolution, PMEDNet introduces a novel Low-rank Motion Descriptor (LMD) that estimates the salient deformation directions through low-rank analysis and models micro-motions in an interpretable way. Extensive experiments demonstrate that the proposed method can effectively perceive the subtle deformations between ME point cloud sequences and achieve competitive performance.

Abstract:
Existing Domain Generalization (DG) approaches ignore the impact of deep model architectures across diverse domains, which may lead to sub-optimal performance. Although neural architecture search (NAS) can be employed to automatically discover architectures with stronger generalization, existing methods primarily focus on maximizing in-distribution performance on the training set, thereby hindering Out-of-Distribution (OoD) generalization towards unseen domains. In this paper, we propose Domain Generalized Neural Architecture Search (DGNAS), a novel NAS paradigm for OoD generalization towards unseen domains by exploiting large-scale vision language models (VLMs). Specifically, we reformulate NAS into an ''augment first, search and fine-tune'' paradigm. We formulate VLMs as an efficient proxy to simulate arbitrary domain shifts in a joint vision-language space, thereby synthesizing OoD domains only with verbal descriptions. Subnets are iteratively searched and fine-tuned on those augmented domains to facilitate unbiased domain alignment and improved OoD generalization. Furthermore, an auxiliary search agent empowered by large language models (LLMs) could harness inherent domain knowledge to refine subnet candidates and quickly navigate the search process. Experiments show DGNAS achieves superior performance on four domain generalization tasks, with significant improvements over baselines and more lightweight architectures. DGNAS outperforms its counterpart AutoFormer by 11.75% while reducing parameters by 4.30% and search time by 35.47%. DGNAS enables data-free domain augmentation, precise OoD adaptation and LLM-driven search navigation, which pioneers a new direction for generalizable architecture search.

Abstract:
Open-Vocabulary Audio-Visual Event Localization (OV-AVEL) requires models to localize audio-visual events while predicting explicit categories for events unseen during training. Current baselines adopt a text-centered pairwise modality decision strategy. However, it lacks unified modeling of multimodal semantic consistency, rendering models more prone to errors when modalities are corrupted by noise. Meanwhile, existing methods suffer from severe open space risk, where models tend to misclassify unseen categories as seen ones during training. To mitigate these issues, we propose the Spherical Consistency and Test-time Adaptation (SCTA) framework. Unlike existing decision strategies, SCTA introduces the Circumscribed Hypersphere Alignment Representation Measure (CHARM) for multimodal consistency modeling. It uses the volume of the circumscribed hypersphere formed by modality vectors in high-dimensional space as a consistency measure, enabling a unified trimodal collaborative decision mechanism and improving noise robustness. To alleviate open space risk, we propose a forward decoupling and backward adaptation strategy. During forward propagation, we design a Subspace Decoupling Module (SDM) to enhance the discriminability of event semantics. In the backward phase, we pioneer integrating Test-Time Adaptation (TTA) into OV-AVEL by introducing Event Consistency TTA (ECTTA), which adjusts decision boundaries for each sample based on the test data. Extensive experiments on OV-AVEBench show that SCTA improves model capabilities in localization and generalization to unseen categories, significantly outperforming SOTA methods.

Abstract:
Reference Audio-Visual Segmentation (Ref-AVS) aims to segment objects in audible videos based on multimodal cues in reference expressions. Previous methods overlook the explicit recognition of expression difficulty and dominant modality in multimodal cues, over-rely on the quality of the instruction-tuning dataset for object reasoning, and lack reflective validation of segmentation results, leading to erroneous mask predictions. To address these issues, in this paper, we propose a novel training-free Multi-Agent Recognition, Reasoning, and Reflection framework to achieve high-quality Reference Audio-Visual Segmentation, termed MAR3. Incorporating the sociological Delphi theory to achieve robust analysis, a Consensus Multimodal Recognition mechanism is proposed that enables LLM agents to explicitly recognize the difficulty of reference expressions and the dominant modality of multimodal cues. Based on our modality-dominant difficulty rule, we propose an adaptive Collaborative Object Reasoning strategy to reliably reason about the referred object. To further ensure precise mask prediction, we develop a Reflective Learning Segmentation mechanism, in which a check agent examines intermediate segmentation results and iteratively corrects the object text prompt of the segment agent. Experiments demonstrate that MAR3 achieves superior performance (69.2% in J&F) on the Ref-AVSBench dataset, outperforming SOTA by 3.4% absolutely.

Abstract:
Instruction-based image editing has recently attracted substantial interest for generative multimedia. While significant advances have been achieved, existing works mainly rely on the explicit visibility of target objects, posing significant challenges of geometric inconsistencies and texture artifacts when editing partially occluded objects. We demonstrate that such an assumption is often incorrect, as object interactions and occlusions commonly exist in scenarios. To tackle this challenge, we propose AmodalEdit, a new framework for instruction-based image editing via amodal scene layout perception. In particular, given an input image with text instructions, we first introduce an instruction graph construction module that maps the input modalities into a structured representation with object editing and occlusion attributes. We then adopt an amodal layout perception module to infer complete object masks and underlying relationships. Finally, we propose an instruction amodal adapter module to enforce global scene coherence by guiding target objects with instructions, while preserving the physical integrity of non-target objects. Experimental results demonstrate that AmodalEdit consistently surpasses state-of-the-art methods in structural integrity and physical plausibility. Furthermore, we showcase the superiority of AmodalEdit in multi-turn image editing by maintaining the amodal scene layout across editing sequences.

Abstract:
Generative retrieval has demonstrated significant success by unifying representation learning and search into a single sequence-to-sequence generation task. However, extending this paradigm to cross-modal retrieval reveals a critical challenge arising from the inherent information asymmetry across different modalities, such as the gap between concise text queries and dense visual candidates. This structural mismatch causes the autoregressive decoder to suffer from forced hallucination when generating identifiers via standard trie-constrained beam search, where the model is severely penalized for failing to guess fine-grained details absent from the query, allowing irrelevant candidates to hijack top rankings. To address this issue, we propose Wildcard Inference with Dynamic Expansion (WIDE). WIDE employs Adaptive Entropy Thresholding (AET) to calibrate layer-specific uncertainty boundaries offline. During the decoding generation phase, Asymmetry-aware Wildcard Decoding (AWD) detects semantic blind spots and emits wildcards instead of forced deterministic identifiers, dynamically expanding the search space without incurring log-probability penalties. Finally, Blind-Spot Re-ranking (BSR) evaluates the expanded candidate pool using a hybrid scoring mechanism that combines discrete generation confidence with continuous semantic similarity. Extensive experiments on the M-BEIR benchmark demonstrate that WIDE outperforms state-of-the-art generative retrieval methods, effectively suppressing forced hallucination while maintaining compact index structures.

Abstract:
Spatio-Temporal Video Grounding requires jointly localizing target objects across both temporal and spatial dimensions based on natural language queries, posing fundamental challenges for existing Multimodal Large Language Models (MLLMs). We identify two core challenges: entangled spatio-temporal alignment, arising from coupling two heterogeneous sub-tasks within the same autoregressive output space, and dual-domain visual token redundancy, where target objects exhibit simultaneous temporal and spatial sparsity, rendering the overwhelming majority of visual tokens irrelevant to the grounding query. To address these, we propose Bridge-STG, an end-to-end framework that decouples temporal and spatial localization while maintaining semantic coherence. While decoupling is the natural solution to this entanglement, it risks creating a semantic gap between the temporal MLLM and the spatial decoder. Bridge-STG resolves this through two pivotal designs: the Spatio-Temporal Semantic Bridging (STSB) mechanism with Explicit Temporal Alignment (ETA) distills the MLLM's temporal reasoning context into enriched bridging queries as a robust semantic interface; and the Query-Guided Spatial Localization (QGSL) module leverages these queries to drive a purpose-built spatial decoder with multi-layer interactive queries and positive/negative frame sampling, jointly eliminating dual-domain visual token redundancy. Extensive experiments across multiple benchmarks demonstrate that Bridge-STG achieves state-of-the-art performance among MLLM-based methods. Bridge-STG improves average m\_vIoU from 26.4 to 34.3 on VidSTG and demonstrates strong cross-task transfer across various fine-grained video understanding tasks under a unified multi-task training regime.

Abstract:
Contrastive learning has become a dominant paradigm for multi-view clustering due to its ability to align representations across views while preserving inter-sample discriminability. However, existing contrastive multi-view clustering methods generally treat all views as equally reliable and enforce uniform cross-view alignment. In real-world applications, where views often exhibit heterogeneous distributions and imbalanced quality, such forced alignment can undermine representation learning by propagating noise from low-quality views to high-quality ones. To address this challenge, we propose a novel framework called Robust Hierarchical Contrastive Learning for Multi-view Clustering with Imbalanced Views (RCHMC-IV). Specifically, RCHMC-IV uses a weighted cross-view learning mechanism that explicitly accounts for both feature-level discrepancies and cluster-level distribution shifts across views. By adaptively weighting features and clusters, RCHMC-IV facilitates discriminative cross-view representation learning and mitigates the risk of degrading high-quality views through forced alignment with inferior ones. Extensive experiments across twelve datasets demonstrate that RCHMC-IV significantly outperforms existing state-of-the-art approaches.

Abstract:
Generalizable cross-view geo-localization aims to match the same location across views in unseen regions and conditions without GPS supervision. Its core difficulty lies in severe semantic inconsistency caused by viewpoint variation and poor generalization under domain shift. Existing methods mainly rely on 2D correspondence, but they are easily distracted by redundant shared information across views, leading to less transferable representations. To address this, we propose GeoLink, a 3D-aware semantic-consistent framework for Generalizable cross-view geo-localization. Specifically, we offline reconstruct scene point clouds from multi-view drone images using VGGT, providing stable structural priors. Based on these 3D anchors, we improve 2D representation learning in two complementary ways. A Geometric-aware Semantic Refinement module mitigates potentially redundant and view-biased dependencies in 2D features under 3D guidance. In addition, a Unified View Relation Distillation module transfers 3D structural relations to 2D features, improving cross-view alignment while preserving a 2D-only inference pipeline. Extensive experiments on multiple benchmarks show that GeoLink consistently outperforms state-of-the-art methods and achieves superior generalization across unseen domains and diverse weather environments.

Abstract:
Attributed hypergraph clustering (AHC) is crucial for uncovering meaningful communities in complex relational data that exhibit both multi-way interactions and rich node attributes. While recent learning-based AHC methods have achieved notable advances, they suffer from three major limitations: (1) most methods decouple representation learning and clustering into separate stages, resulting in suboptimal non-cluster-related embeddings; (2) recent methods, such as CACH, introduce clustering-related losses but rely on heuristic alignment between soft cluster assignments and hard pseudo-labels, failing to optimize the underlying structure-attribute clustering objective; and (3) many approaches depend on computationally intensive contrastive learning and attention mechanisms, hindering scalability. To address these challenges, we propose LNAHC, a lightweight, end-to-end neural framework for AHC that comprises two key components: a novel attribute-aware hypergraph random walk strategy for constructing a structure-attribute proximity matrix that preserves both local and global semantic relationships, and a differentiable multi-hop attributed hypergraph normalized cut module that directly optimizes a relaxed cut-like loss, thereby regularizing the embedding space to produce high-quality clusters without post-processing or heavy architectural components. Extensive experiments on seven datasets demonstrate that the proposed LNAHC achieves superior performance compared to nine baselines.

Abstract:
Video Reasoning Segmentation (VRS) is currently bottlenecked by a single-event reasoning paradigm, which reduces video temporal dynamics into isolated, independent occurrence matching. Consequently, existing methods suffer from severe semantic misalignment when target identification relies on a sequence of temporally conditional events, thereby exposing a critical Multi-Hop Reasoning gap. To bridge this gap, we introduce ChronoVRS benchmark, comprising 2.5K videos and 30K queries explicitly designed for multi-hop reasoning, shifting the focus from static world knowledge to sequential temporal prerequisites. To resolve complex spatiotemporal dependencies, we propose ChronoChain, a novel multi-agent framework utilizing a backward-chaining mechanism. Mirroring goal-directed human cognition, ChronoChain dynamically constructs a relational video object graph from isolated, task-relevant entities. Operating upon this structured foundation, the framework traverses temporal, identity and semantic edges in reverse to decode multi-hop dependencies. Anchoring on the final query objective and regressively verifying prerequisite conditions, ChronoChain effectively overcomes the limitations of isolated visual matching, achieving state-of-the-art performance across referring, reasoning and spatiotemporal pixel-level retrieval benchmarks.

Abstract:
Synthesizing target-view projection from a single panoramic X-ray (PX) is a promising direction for low-dose and low-cost dental imaging enhancement. However, diffusion-based dental projection synthesis faces two key challenges: ambiguous cross-view geometry caused by the lack of explicit 3D priors, and degradation of diagnostically important high-frequency anatomical details during iterative denoising. To address these challenges, we propose DentalDiff, an angle-controllable latent diffusion framework for CBCT projection synthesis from a single PX. First, we introduce a Cross-View Fusion (CVF) module that enhances geometry-aware structural alignment during denoising by explicitly modeling semantic, visual, and target-view features through multi-scale cross-view attention. Second, we design a High-Frequency Perception Loss (HPL) that combines perceptual consistency with frequency-aware focal reconstruction to preserve fine dental structures, including enamel boundaries, root contours, and local anatomical textures. This complementary design enables DentalDiff to jointly improve cross-view geometric consistency and high-frequency anatomical fidelity in synthesized projection. Extensive experiments on 499 internal paired PX-CBCT cases and 153 external public CBCT cases from the ToothFairy Challenge Dataset demonstrate that DentalDiff consistently outperforms existing baselines in both perceptual quality and geometric accuracy, producing more anatomically faithful and view-consistent target-view projection.

Abstract:
Practical anti-UAV surveillance often records visible and thermal infrared streams that are spatially misaligned and have different resolutions, making cross-view association unreliable and naive fusion prone to tracking drift, especially for tiny UAVs. However, most existing anti-UAV tracking methods still rely on a single modality. To this end, this paper revisits unaligned RGBT anti-UAV tracking and proposes U2Track, a novel alignment-then-fusion approach for robust anti-UAV tracking. U2Track is an object-centric architecture for reliable alignment and cross-modal interaction under uncertain correspondence. Specifically, the alignment network learns box mappings between RGB and TIR to establish cross-view correspondence without dense image registration. Moreover, we introduce a deformable cross-modal token fusion module. It combines a shallow fusion prior with bidirectional deformable cross-attention to handle residual offsets and a reliability-aware gating to suppress unreliable cross-modal information. Overall, U2Track offers a practical solution for robust RGBT anti-UAV tracking in realistic heterogeneous sensor settings. Experiments on the unaligned Anti-UAV benchmark demonstrate state-of-the-art performance and strong robustness under unaligned views, and further demonstrate strong generalization to other unaligned RGBT tracking scenarios.

Abstract:
Talking face generation aims to synthesize realistic speaking portraits from a single image, yet existing methods either rely on computationally expensive end-to-end video generation or use explicit optical flow and local warping that struggle with large poses, occlusions, and non-rigid articulations, often causing identity drift. We present ~\emphIMTalker, a novel framework that achieves efficient and high-fidelity talking face generation through implicit motion transfer. The core idea is to replace traditional flow-based warping with a cross-attention mechanism that implicitly models motion discrepancy and identity alignment within a unified latent space, enabling robust global motion rendering. To further preserve speaker identity during cross-identity reenactment, we introduce an identity-adaptive module that projects motion latents into personalized spaces, ensuring clear disentanglement between motion and identity. In addition, a lightweight flow-matching motion generator produces vivid and controllable implicit motion vectors from audio, pose, and gaze cues. Extensive experiments demonstrate that IMTalker surpasses prior methods in motion accuracy, identity preservation, and audiolip synchronization, achieving state-of-the-art quality with superior efficiency, operating at 40 FPS for video-driven and 42 FPS for audio-driven generation on an RTX 4090 GPU.

Abstract:
Dynamic Facial Expression Recognition (DFER) is a critical task for affective signal understanding. However, existing methods still struggle to capture discriminative expression cues due to subtle facial variations, complex scene interference, and ambiguous expression semantics. Although CLIP provides powerful vision-language priors, existing CLIP-based DFER methods mostly adapt the frozen backbone in a uniform manner, without explicitly modeling the functional differences among structural, dynamic, and semantic cues. Consequently, they cannot simultaneously support key-region perception, critical dynamic modeling, and fine-grained affective semantic alignment. To address this issue, we propose HS-CLIP, a hierarchy-aligned stage-wise CLIP adaptation framework for DFER. The core idea is to assign heterogeneous cues to the representation stages that best match their functional roles. Specifically, we design a structure-guided visual adaptation mechanism that enhances expression-relevant facial regions in the early stage and emphasizes critical dynamic changes in the intermediate stage, thereby reducing redundancy from irrelevant frames. We further introduce a dynamic-aware cross-modal semantic adaptation mechanism that regulates vision-language interaction at the high-level stage to achieve finer-grained affective alignment. Extensive experiments on multiple challenging in-the-wild DFER benchmarks demonstrate that HS-CLIP achieves superior overall performance with high parameter efficiency. These results demonstrate that hierarchy-aligned stage-wise adaptation offers an effective and potentially generalizable paradigm for adapting frozen vision-language models to DFER.

Abstract:
The empirical success of attention mechanism in Multimodal Large Language Models (MLLMs) often obscures its inherent, subtle flaws. Specifically, MLLMs consistently exhibit disproportionate attention toward certain semantically uninformative visual tokens, a phenomenon termed register or Visual Attention Sinks. While existing inference intervention methods attempt to identify these sink tokens and redistribute their attention weights, such approaches typically treat these tokens in isolation and suffer from computational inefficiency. Instead, we reframe this phenomenon as a generalized textual bias exerted over visual features that extends beyond isolated sink tokens. From this perspective, a pervasive structural bias leads to the dilution of the semantic visual signal, precipitating multimodal hallucinations as the model prioritizes linguistic priors over valid visual evidence. To address this limitation, we introduce Saliency-guided Purification and Adaptive Redistribution (SPAR), a training-free, plug-and-play intervention. SPAR mitigates this generalized textual bias by purifying structural noise and subsequently redistributing the reclaimed attention budget to the most informative visual regions. Comprehensive evaluations across a diverse spectrum of hallucination benchmarks demonstrate that SPAR effectively restores authentic visual grounding with negligible computational overhead.

Abstract:
Despite their strong performance in embodied tasks, recent VisionLanguageAction (VLA) models remain highly fragile under multimodal perturbations, where visual corruption and linguistic noise jointly induce distribution shifts that degrade task-level execution. Existing robustness approaches typically rely on joint training with perturbed data, treating robustness as a static objective, which leads to conflicting optimization between robustness and task fidelity. In this work, we propose STRONG-VLA, a decoupled fine-tuning framework that explicitly separates robustness acquisition from task-aligned refinement. In Stage I, the model is exposed to a curriculum of multimodal perturbations with increasing difficulty, enabling progressive robustness learning under controlled distribution shifts. In Stage II, the model is re-aligned with clean task distributions to recover execution fidelity while preserving robustness. We further establish a comprehensive benchmark with 28 perturbation types spanning both textual and visual modalities, grounded in realistic sources of sensor noise, occlusion, and instruction corruption. Extensive experiments on the LIBERO benchmark show that STRONG-VLA consistently improves task success rates across multiple VLA architectures. On OpenVLA, our method achieves gains of up to 12.60% under seen perturbations and 7.77% under unseen perturbations. Notably, similar or larger improvements are observed on OpenVLA-OFT (+14.48% / +13.81%) and \pi_0 (+16.49% / +5.58%), demonstrating strong cross-architecture generalization. Real-world experiments on an AIRBOT robotic platform further validate its practical effectiveness. These results highlight the importance of decoupled optimization for multimodal robustness and establish STRONG-VLA as a simple yet principled framework for robust embodied control.

Abstract:
Click-through rate (CTR) prediction increasingly leverages multimodal content together with knowledge derived from large language models (LLMs) to complement sparse behavioral signals. Yet it remains unclear how such external semantics can be incorporated into behavior-dominant CTR models without disturbing collaborative representations. In this paper, our preliminary experiments show that models using textual semantics alone often perform comparably to those using both textual and visual semantics, yet degrade steadily under item-text mismatch. These observations suggest that the key challenge is not stronger multimodal fusion, but selective admission of reliable external semantics. To address this issue, we propose a tri-subspace semantic admission framework with grounding evidence, called TriSAGE. It first constructs two complementary LLM views for each item, and uses their difference as a candidate semantic increment. It then examines whether this increment is grounded by fine-grained vision--language response patterns, treating the image as grounding evidence rather than directly injecting visual features. Finally, it filters unsupported semantic directions and injects the admitted increment through a behavior-preserving residual adapter, enabling controlled semantic enhancement of item representations. Experiments show that TriSAGE outperforms multimodal and LLM-based methods, with stronger robustness to text misalignment and multimodal noise.

Abstract:
Generative face video coding (GFVC) is effective for ultra-low-bitrate talking-head communication, but practical video conferencing often involves hand gestures and face-hand occlusions that are not well handled by existing face-centric pipelines. In such hand-active scenarios, facial and hand motion follow heterogeneous dynamics that complicate compact representation, motion transformation, and reconstruction. To address this, we present a hybrid-prior framework for interactive generative video coding. Our method combines a semantically disentangled structural prior, encoded by compact face-hand landmarks for motion alignment and pose guidance, with a pretrained generative prior, instantiated by a conditional latent diffusion decoder for reconstructing newly emerged or heavily occluded content. Concretely, we introduce a motion-factorized alignment module that decouples facial and hand motion, and a diffusion-based reconstruction stage conditioned on both pose maps and motion-transferred features. We further construct a hand-active benchmark tailored to interactive video conferencing. Experiments show that the proposed method achieves superior reconstruction quality and compression performance over VVC and representative GFVC baselines on both full-frame and hand-region metrics. In particular, it achieves bitrate savings of 60.6% and 76.1% over VVC in terms of LPIPS and DISTS, while also yielding clear improvements in hand-sensitive regions.

Abstract:
Compared with half-duplex dialogue systems where the system waits for user turn completion before it responds, natural full-duplex dialogue systems require agents to act proactively in real time, including timely interruptions and backchannels. This creates a key challenge: improving turn timing without sacrificing response quality. To address limitations in realistic proactive turn-taking, we build a generalized style-aware full-duplex framework with three key components. Firstly, we propose LPS-TC, a Lightweight Proactive Speech Turn Controller for plug-and-play integration. It features a fine-grained action space covering both reactive and proactive turn behaviors, enabling half-duplex models with full-duplex capabilities and enhancing existing full-duplex models with superior timing control. Secondly, we construct WildTurn, a large-scale, real-world English dataset containing approximately 2,981 hours of filtered multi-turn stereo conversations from face-to-face and telephone conversations, annotated with five turn-taking and five backchanneling styles. Trained on WildTurn, LPS-TC exhibits rich spoken dynamics that are not captured by existing static full-duplex benchmarks. Thirdly, we introduce a two-tier evaluation scheme that assesses both chunk-level timing precision and turn-level interaction quality under realistic streaming constraints. Our experiments, integrating LPS-TC with half-duplex models like Qwen2.5-Omni and full-duplex models like Freeze-Omni, showcase its superior performance in timing appropriateness and response quality. Our framework also demonstrates fine-grained style controllability and strong generalizability, enabling more natural and human-like spoken interactions.

Authors: Zhicheng Liang (School of Science and Engineering, The Chinese University of Hong Kong, Shenzhen), Boyan Li (School of Science and Engineering, The Chinese University of Hong Kong, Shenzhen), Zijian Cao (School of Science and Engineering, The Chinese University of Hong Kong, Shenzhen), Dayou Zhang (School of Information and Electronics, Capital Normal University), Wei Sun (School of Science and Engineering, The Chinese University of Hong Kong, Shenzhen), Zizhou Gong (School of Science and Engineering, The Chinese University of Hong Kong, Shenzhen), Fangxin Wang (School of Science and Engineering, The Chinese University of Hong Kong, Shenzhen)

Abstract:
Streaming large-scale 3D Gaussian Splatting (3DGS) scenes to bandwidth- and memory-constrained clients is an open systems problem. Replacement-based level-of-detail (LOD) pipelines bound client memory but re-transmit overlapping content across levels, wasting 57--62% of bandwidth on redundancy. Vanilla additive streaming transmits each primitive once yet produces incoherent early views with no path to memory-efficient rendering. We present \method, the first \emphSculpted Additive progressive 3DGS streaming system. \method is entirely training-free: a one-time offline stage partitions a pretrained scene into spatial tiles, constructs append-only coarse-to-fine levels via hybrid saliency--spatial sampling, and applies support-aware inflation that enlarges sparse early Gaussians to produce visually coherent partial views. At runtime, a value-per-byte scheduler prioritizes tiles by viewpoint saliency and completion, while inflated coarse levels serve as a persistent LOD cache that enables bounded-memory rendering without re-transmission. Across four benchmark scenes spanning 2.5M to 81M Gaussians, \method reduces bandwidth redundancy to 11.1%, requires 3.2× fewer bytes than the strongest replacement-based baseline to reach 27\,dB PSNR, and sustains real-time rendering on consumer mobile devices---including an 810M-splat stress test at 60\,FPS on mobile phones and 22\,FPS under a 128\,MB VRAM budget.

Abstract:
\beginabstract Surface defect segmentation plays a crucial role in industrial quality inspection, but existing supervised learning paradigms are often limited by the scarcity of defect samples. Few-shot segmentation enables rapid adaptation to novel categories with only a few annotated samples; however, traditional methods usually suffer from insufficient semantic guidance and loss of fine-grained details when dealing with complex industrial backgrounds and defect patterns. To address these issues, the Text-Guided Prior Decoupling and Alignment framework (TDANet) is proposed. A Text-Guided Hierarchical Prior Refinement module is designed to leverage semantic cues from CLIP together with multi-scale visual priors to generate high-quality prior masks, providing reliable global guidance for segmentation. Then, the SemanticVisual Cross Alignment mechanism is introduced to enable bidirectional interaction between global semantic features and visual features, while explicitly decoupling foreground and background under the guidance of the prior mask. Furthermore, to enhance the representation of fine-grained textures and boundary structures before alignment, a SpatialFrequency Synergy enhancement module is adopted. Extensive experiments on the FSSD-12 and Surface Defect-4^i benchmarks demonstrate that the proposed method achieves state-of-the-art performance under both 1-shot and 5-shot settings. \endabstract

Abstract:
Accurate medical image segmentation demands fast and precise delineation of anatomical boundaries across different imaging modalities. However, prevailing Transformer-based models rely on similarity-driven global attention that may aggregate inconsistent structures and blur anatomical boundaries, while quadratic complexity slows training and inference on high-resolution images. To address these limitations, we propose FSP-Former, a Fourier Spectral Projected Transformer for structure-aware and lightweight attention modeling. Concretely, FSP-Former derives spectral features from token affinity structures to capture global anatomical organization. The dominant components of these features define a compact projection subspace, which explicitly encodes the principal structural patterns in medical images. By projecting queries, keys, and values into this structure-aware spectral space, attention aggregation is inherently constrained to respect anatomical boundaries, ensuring structure-consistent interactions. This mechanism effectively compresses global context into a low-rank form, enabling lightweight attention computation. To obtain such spectral representations efficiently, we exploit the Fourier diagonalization of circulant operators, yielding closed-form spectral responses without explicit eigen-decomposition. Overall, this spectral projection framework preserves anatomical coherence and boundary delineation while significantly reducing computational and memory costs. We evaluate FSP-Former across different imaging modalities including MRI in BraTS 2017, CT in Synapse, and dermoscopic images in ISIC 2018. FSP-Former improves Dice by 1.8% and reduces HD by 15.7% over state-of-the-art Transformer baselines, while accelerating inference by 15.6%. These results demonstrate the effectiveness and efficiency of Fourier spectral projection.

Abstract:
Academic papers are a primary carrier of scientific knowledge, yet most of this knowledge remains locked in PDFs that are optimized for human reading rather than machine use. For Multimodal Large Language Models (MLLMs), the core challenge is not only perception, but representation: scientific pages interleave text with Structured Academic Elements (SAEs) such as tables, formulas, charts, and pseudocode, whose structure, data, and logic are poorly preserved by common surrogates like Markdown. We therefore propose Compilable Academic Document Parsing (CADP), a paradigm that reconstructs a full page as contextual \LaTeX plus executable Python, so that structure-preserving elements and executable chart representations can be reconstructed, recompiled, and directly verified against the source page. To support this setting, we introduce \textscCADP-Bench, an expert-verified benchmark of full academic pages containing tightly coupled text and multiple SAE types, evaluated through a re-injection compilation protocol. We further study current capabilities using SOTA MLLMs and an exploratory multi-agent baseline that incorporates common agentic techniques. Results show that even frontier models still struggle to produce high-fidelity executable reconstructions, highlighting substantial room for improvement in structure-aware scientific document parsing. \textscCADP-Bench is released for future research.

Abstract:
Video Large Language Models (VideoLLMs) have demonstrated impressive capabilities in video understanding, yet the massive number of input video tokens incurs a significant computational burden for deployment. Existing methods mainly prune video tokens at input level while neglecting the inherent information structure embedded in videos and large language models (LLMs). To address this, we propose HieraVid, a hierarchical pruning framework that progressively and dynamically reduces visual redundancy. Based on two observations that videos possess the segment-frame structure and LLMs internally propagate multi-modal information unidirectionally, we decompose pruning into three levels: 1) segment-level, where video tokens are first temporally segmented and spatially merged; 2) frame-level, where similar frames within the same segment are jointly pruned to preserve diversity; 3) layer-level, redundancy gradually shrinks as LLM layer increases w/o compromising performance. We conduct extensive experiments on four widely used video understanding benchmarks to comprehensively evaluate the effectiveness of HieraVid. Remarkably, with only 30% of tokens retained, HieraVid achieves new state-of-the-art performance, while maintaining over 98% and 99% of the performance of LLaVA-Video-7B and LLaVA-OneVision-7B, respectively.

Abstract:
Vision-language models (VLMs) such as CLIP exhibit strong zero-shot generalization but remain highly vulnerable to adversarial perturbations. Recent zero-shot defenses largely attribute this vulnerability to text-guided attention shift and thus focus on spatial attention rectification. We reveal a contrasting failure mode, termed spatial-semantic decoupling: adversarial images can preserve accurate text-guided localization while still causing misclassification. This indicates that the attack primarily induces semantic drift in the embedding dimensions, which is largely invisible to spatial attention objectives. Moreover, we find that naively aligning the full embedding space is counterproductive due to dimensional heterogeneity in VLM representations, where redundant tail dimensions are disproportionately corrupted and dominate the alignment loss. To address these issues, we propose Semantic Purification via dimensional Alignment (SPA). SPA identifies a semantic core by ranking dimensions with inter-sample variance from a frozen teacher, and performs teacher-anchored alignment only on the selected dimensions via dimension-aware pooling, while suppressing the influence of noisy tail dimensions during optimization. Extensive experiments on 16 diverse datasets demonstrate that SPA consistently improves zero-shot adversarial robustness under strong attacks (PGD-100, AutoAttack, and C\&W) while maintaining, and often improving, clean accuracy over prior state-of-the-art methods.

Abstract:
Virtual Try-On (VTON) has achieved progress with the advent of diffusion models. However, existing methods ranging from specialized VTON pipelines to general multimodal editors often suffer from limited local perception and cross-source misalignment. As a result, existing VTON often exhibit semantic inconsistencies at both local and global levels, such as attribute distortions caused by mask-based localization fragility and mechanical copy-paste artifacts due to global semantic misalignment. In this paper, we propose DSR-Tryon, a mask-free framework driven by a Dual-Scale Semantic Reasoning paradigm that mimics the human progressive cognitive process, ranging from fine-grained attributes to holistic synergy. Our framework addresses the semantic limitations through two key modules. First, the Local Semantic Thinker (LST) serves as a cognitive front-end to achieve a deep perception of garment semantics. With this semantic perception, the LST enables an effective disentanglement that accurately distills relevant latents while filtering out noisy attributes. Second, the Global Semantic Alignment (GSA) module acts as a perceptual coordinator. Through representation alignment with pre-trained visual models, GSA ensures that the transferred garment is consistent with the target person and textual prompts, thus alleviating global semantic inconsistencies. Experiments demonstrate that DSR-Tryon exhibits superior in-the-wild zero-shot generalization although trained exclusively on simplistic flat-lay garment images. Our DSR-Tryon also significantly outperforms both specialized VTON methods and general image editing frameworks in attribute preservation and semantic metrics. The proposed DSR-Tryon successfully resolves semantic inconsistencies through a progressive semantic cognitive process, establishing a novel and low-cost paradigm for versatile and coherent VTON.

Abstract:
Immersive 3D telepresence promises a transformative paradigm for remote interaction, enabling participants to perceive a compelling sense of shared physical presence. However, existing systems rely on costly multi-view capture hardware, which limits their widespread adoption. In this paper, we present SingleGS, an end-to-end high-fidelity monocular 3D telepresence system built entirely on consumer-grade devices. Our system comprises a monocular RGB camera, an edge server for 3D reconstruction and rendering, and a glasses-free 3D tablet with view-tracking. To address the challenge of photorealistic reconstruction from ambiguous monocular input within strict latency budgets, we develop a lightweight pipeline that infers 3D Gaussian representations directly from single RGB frames by decomposing the process into geometry initialization and Gaussian attribute estimation, with tailored designs. We implement a prototype of SingleGS, and experimental results demonstrate that SingleGS achieves state-of-the-art visual quality, especially in close-range views, while maintaining at least 28 frames per second with an end-to-end latency of around 90 milliseconds.

Abstract:
Previous voxel-based semantic scene completion (SSC) algorithms are limited by low resolution and high computational requirements, while the partial inputs of point cloud-based SSC methods are typically generated from complementary RGB-D images or acquired through expensive LiDAR sensors. To address these limitations, we proposed ScenIma, which reconstructs a refined semantic scene point cloud from a single image. Specifically, ScenIma first employs the proposed Adaptive Pose Correction Algorithm (APCA), aligning the degraded scenes with the world coordinate system. Then, a Scene Spatial-aware Transformer Encoder (SSTE), designed with our Scene Flash Attention Block (SFAB), encoding the aligned degraded 3D scenes and produces structure-completed coarse semantic point clouds. Finally, our constructed Scene Semantic-Spatial-aware Transformer Decoder (S3TD) simultaneously refines both scene geometry and semantics, generating the final output. Furthermore, we proposed the KITTI-SSC dataset. Experiments on multiple datasets demonstrate that, ours outperforms existing point cloud-based SSC algorithms from a single image, while requiring only 28% of the GPU memory of these algorithms.

Abstract:
Audio-visual (AV) deepfake detectors have become increasingly effective by exploiting temporal coherence between speech and facial motion. In this paper, we show that this strength can also become a shared vulnerability: many AV detectors appear to rely on similar cross-modal synchronization statistics as discriminative cues. Motivated by this observation, we propose a coherence-guided adversarial attack that manipulates alignment priors. Specifically, we construct a local cross-modal alignment distribution from pre-trained AV encoders and characterize it with two complementary terms: expected absolute deviation from the temporal center, which captures off-center synchronization, and alignment entropy, which captures diffuse or uncertain correspondence. We combine these terms into a unified coherence score and optimize it directly to generate bounded joint perturbations on both video and audio. Experiments on FakeAVCeleb show that the proposed method substantially degrades representative AV deepfake detectors while preserving perceptual quality. Beyond attack effectiveness, our results highlight cross-modal coherence as a potential robustness bottleneck in AV deepfake detection.

Abstract:
Remote Sensing Image Change Captioning (RSICC) aims to describe the differences between two images with a short description. Existing methods conventionally concentrate on optical imagery under clear-sky conditions, failing to account for complex multimodal environmental influences (e.g., clouds, haze, low-light conditions, as well as SAR and multispectral modalities). Moreover, these methods merely generate coarse-grained change captions, leading to the omission/misjudgment of critical change information, particularly in complex scenarios. To address these limitations, we propose a Fine-grained RSICC (FRSICC) task tailored for complex multimodal environments. To effectively capture critical change information under such environments, we propose an environment-aware trajectory and object-location guided refinement framework for the FRSICC task, termed FRSICC-TR. Specifically, to quantify image environmental information, we propose Environment-Aware Trajectory Generation, which generates the collaboration trajectory among different agents. We further propose Dynamic Image-Enhanced Caption Generation that dynamically adapts to image-specific environmental perturbations for targeted enhancement to generate a high-quality initial change caption. To better capture fine-grained changes, we design Object-Location-Guided Caption Refinement to iteratively recover missed changed objects and correct imprecise spatial location information. We finally construct a large-scale FRSICC dataset, called FRSI-CC. Extensive experiments demonstrate that FRSICC-TR achieves SOTA performance in fine-grained captioning.

Abstract:
Instruction-guided image editing allows users to specify modifications in natural language. Recent advances show that Diffusion Transformers (DiTs) surpass U-Net-based diffusion models in both scalability and performance. However, real-world scenarios often involve applying multiple instructions simultaneously. Step-by-step editing accumulates errors and degrades image quality, while putting all instructions into a single prompt typically leads to incomplete edits. To address these challenges, we introduce Instruction Influence Disentanglement (IID), a training-free framework that enables simultaneous handling of multiple instructions within a single denoising process for DiT-based models. By analyzing the self-attention mechanisms of DiTs, we observe distinct attention patterns under multi-instruction settings and design instruction-specific masks to disentangle their influence. These masks guide the editing process to achieve localized modifications while preserving consistency in non-edited regions. Extensive experiments demonstrate that IID improves both fidelity and instruction completion, while also reducing computational overhead compared to existing approaches.

Abstract:
Omni-modal Large Language Models (OLLMs) that integrate visual, auditory, and textual processing face severe safety risks. They exhibit fragile defenses against audio-visual joint harmful inputs and demonstrate inconsistent safety performance across different modalities, enabling simple modality-switching jailbreaks. However, existing safety benchmarks fail to comprehensively assess these risks due to the absence of audio-visual joint samples, limited modality coverage, and lack of parallel test cases for cross-modal consistency evaluation. To address these gaps, we introduce Omni-SafetyBench, the first comprehensive parallel benchmark for OLLM safety evaluation, featuring 23,328 test instances across 24 modality variations derived from 972 seed samples. Recognizing that complex inputs pose comprehension challenges and that cross-modal consistency is critical for OLLM safety, we propose tailored metrics: a Safety-score based on Conditional Attack Success Rate (C-ASR) and Conditional Refusal Rate (C-RR), and a Cross-Modal Safety Consistency score (CMSC-score). Evaluating 11 state-of-the-art OLLMs reveals severe vulnerabilities: only 3 models exceed 0.6 in both metrics, with safety degrading sharply for audio-visual inputs. Furthermore, evaluation of existing safety alignment methods on Omni-SafetyBench identifies fundamental challenges in OLLM safety alignment, highlighting urgent needs for enhanced research in this domain.

Abstract:
Split Federated Learning (SFL) facilitates privacy-preserving collaborative training with reduced client-side overhead. However, its split architecture introduces unique attack surfaces, rendering it vulnerable to diverse poisoning attacks. Most existing defenses fail to exploit the split paradigm, limiting their ability to detect and contain malicious behaviors at an early stage. To bridge this gap, we propose Target-Oriented Feature Decoupling (TOFD), a unified framework that jointly enables proactive detection and robust optimization against a wide range of poisoning attacks. TOFD operates in three stages: (1) Target Inference, which identifies potential attack targets by refining class-wise safe zones via class-specific Margin Perturbation (MP); (2) Sample Purification, which adaptively filters poisoned smashed data using thresholds calibrated through cross-class minmax normalization of MP; and (3) Decoupling Optimization, which leverages an adversarial guidance model to capture attack-induced patterns and decouple their influence during optimization, thereby suppressing residual adversarial effects. We provide theoretical guarantees for the convergence of TOFD. Extensive experiments on five datasets demonstrate that TOFD consistently outperforms state-of-the-art defenses under diverse attack scenarios, achieving superior robustness with low computational overhead suitable for practical deployment.

Abstract:
Dual-stream Single Image Reflection Separation (SIRS) methods jointly estimate the transmission layer (T) and reflection layer (R) for superior separation. However, real-world training must use I - T as a proxy for R, which is severely contaminated by non-linear ISP effects and dynamic range clipping, creating a data fidelity gap that degrades both branches. Prior methods either tolerate this contamination or suppress it via low-pass filtering, destroying fine-grained structure of R. We formulate Reflection Proxy Purification as a novel task and propose Ref-Gen, a generative data engine built on a pre-trained Diffusion Transformer that purifies contaminated proxies through a Hybrid Condition Module for cross-granularity structural-semantic conditioning and a Cross-Modal Transmission-Suppression Guidance for dual-space T-layer repulsion. A flexible prompt strategy enables fully automatic or user-guided generation, and a physical fidelity guard enforces pixel-domain consistency. Experiments show that Ref-Gen's purified labels consistently boost four state-of-the-art SIRS models across five benchmarks, with recovered reflections achieving the best pixel-level fidelity and strong semantic clarity under vision-language evaluation.

Abstract:
Spiking Neural Networks (SNNs) employing Time-to-First-Spike (TTFS) coding enable extreme energy efficiency by constraining neurons to At-Most-One-Spike (AMOS). However, the inherent AMOS constraint hinders scaling to deep architectures like ImageNet due to information loss in residual pathways and low discriminative power in high-dimensional spaces. Consequently, existing TTFS SNNs often fail to optimize effectively on large-scale datasets. We propose Deep-TTFS, a framework for scaling high-performance TTFS SNNs through high-fidelity temporal fusion. We introduce the ADD-FirstSpike mechanism to guarantee stable gradients under the AMOS constraint. Furthermore, we employ Polynomial Synapses and Temporal Neural Threshold Decay to model time-variant weight trajectories, effectively mitigating irreversible firing errors. Finally, our Temporal Fusion Decoder resolves concurrent firing conflicts while enabling asynchronous early-stop inference. Experimental results demonstrate that Deep-TTFS is the first framework to successfully scale directly training step-by-step TTFS SNNs to ImageNet. Deep-TTFS reduces synaptic operations and energy consumption by over 5× compared to traditional rate coding SNNs, which effectively bridges the gap between AMOS sparsity and deep-layer performance requirements.

Abstract:
Vision-Language Navigation (VLN) requires embodied agents to interpret natural language instructions and navigate through complex 3D environments. While recent advancements employing 3D Gaussian Splatting (3DGS) provide high-fidelity spatial and semantic memory, these representations are fundamentally passive. They construct scene maps strictly from historical and current observations, leaving agents vulnerable to short-sighted decision-making when faced with occlusions or unobserved regions. To address this limitation, we propose Generative 3D Gaussian World Models, a novel framework that transitions 3DGS from a static memory buffer to an active and predictive world model for VLN. By leveraging language instructions and partial egocentric data, our model dynamically completes both the geometry and semantics of unobserved 3D Gaussians. This enables Active Look-Ahead Rendering, allowing the agent to synthesize and evaluate anticipated views of future candidate waypoints before physical movement. Furthermore, we introduce an uncertainty-driven exploration mechanism that adaptively triggers generative planning only upon high perceptual ambiguity, ensuring on-demand computational efficiency. Our method consistently surpasses state-of-the-art baselines on R2R and REVERIE, showcasing superior performance in long-horizon reasoning and highly occluded scenarios.

Abstract:
Diffusion models~(DMs) have gained significant attention for their powerful image synthesis capabilities. However, the applications of DMs may expose sensitive user inputs, thereby raising substantial privacy concerns. Existing privacy-preserving methods offer rigorous privacy guarantees but incur excessive computational and communication overhead. While several acceleration techniques have been developed, directly applying them to DMs often results in severe image quality degradation. To the best of our knowledge, we are the first to identify this degradation and trace its root cause to the coupling between the acceleration-induced bias and the iterative nature of DM sampling. Based on this insight, we propose PrismDM, the first secure DM sampling framework that jointly considers efficiency and accuracy. We propose an accuracy-aware acceleration method to mitigate the bias via multi-objective optimization, along with a feature alignment algorithm that stabilizes the sampling trajectory, thereby improving both sample efficiency and quality. Compared to state-of-the-art methods, PrismDM improves sample quality by 1.56× ～ 14.66×, reduces inference time by 1.07× ～ 3.15×, and minimizes communication overhead by 1.18× ～ 4.20×.

Abstract:
Lifelong Visible-Infrared Person Re-identification (LVI-ReID) aims to continuously learn cross-modality discriminative knowledge from streaming data while preserving previously acquired knowledge. This task is particularly challenging due to the joint effects of cross-modality discrepancy and catastrophic forgetting. Existing methods primarily rely on data replay to mitigate forgetting, which raises privacy concerns and incurs additional memory overhead, limiting their practicality in real-world deployments. In this paper, we propose a novel Multi-level Harmonious Knowledge Learning (MHKL) framework for replay-free LVI-ReID, which aims to learn discriminative, compatible, and adaptive knowledge. The key idea is to balance the model's stability and plasticity through multi-level knowledge distillation across discriminative boundaries, feature space, and identity relations. Specifically, we introduce a Prior-guided Discriminativity Learning paradigm to enforce cross-space discriminative consistency via triplet-based constraints. A Reconstruction-driven Compatibility Learning strategy is further designed to reconstruct compatibility-aware representations, promoting feature space compatibility between the new and old models. In addition, a Relation-aware Adaptability Learning mechanism is proposed to refine and distill cross-modality relational knowledge, thereby enhancing identity-level adaptability across modalities. Extensive experiments on multiple LVI-ReID benchmarks demonstrate that our method effectively mitigates catastrophic forgetting and maintains strong adaptability to new data.

Authors: Yu Liu (Institute of Information Engineering, Chinese Academy of Sciences), Jiahui Liu (Central Conservatory of Music), Liu Zhilin (University of Electronic Science and Technology of China), Cong Cao (Institute of Information EngineeringChinese Academy of Sciences), Fangfang Yuan (Institute of Information Engineering, Chinese Academy of Sciences), Yuling Yang (Institute of Information Engineering, Chinese Academy of Sciences), Pin Xu (Institute of Information Engineering, Chinese Academy of Sciences), Yanbing Liu (Institute of Information EngineeringChinese Academy of Sciences)

Abstract:
Audio-language models generate confident music descriptions that are unsupported by audio. We present the first music-specific, layer-wise, multi-paradigm empirical study of hallucination in audio-language models and formulate it as a hierarchical perceptual grounding failure across five layers: sound events, temporal properties, tonal attributes, style, and emotion. We introduce MuseDiag, a multi-paradigm diagnostic framework with contradiction-based verification, and evaluate nine models (four open-weight and five proprietary API models). We find that (1) vocal misperception is a universal weakness across all nine models, tonal perception is a major axis of architectural differentiation, and Audio-Flamingo-3 leads the primary claim-weighted aggregates while Qwen2.5-Omni remains competitive under layer-balanced analysis; (2) affirmative bias, generation-mode effects, and layer-specific perceptual limitations are each empirically associated with the observed patterns, with convergent evidence from multiple analyses rather than strict causal attribution; and (3) our two training-free mitigation methods, Audio-Dependency-Aware Decoding for Music (ADD-M) and Taxonomy-Guided Perceptual Anchoring (TPA), can reduce hallucination in probing, but their gains vary by model and often do not carry over to free-form generation, showing that music hallucination mitigation must be evaluated across paradigms.

Abstract:
In semi-supervised medical image segmentation, consistency regularization has been widely adopted in clinical-oriented settings due to its efficiency and plug-and-play nature. However, as background noise increases substantially in 3D medical images, foreground consistency constraints are often dominated by visually salient but anatomically irrelevant regions, which limits the generalizability of the existing anatomy-guided and learnable module-augmented methods. To mitigate this issue, we rethink foreground consistency from the perspective of intrinsic 3D medical image properties and propose ForeMatch, a semi-supervised framework without introducing additional learnable components. Specifically, ForeMatch introduces a region selection strategy that samples partially overlapping sub-volumes with distinguishable background statistics, enforcing consistent foreground predictions at the input level. At the feature level, we further propose an adaptive spectral dropout mechanism that modulates feature perturbations based on spectral entropy distributions, suppressing redundant textures while preserving discriminative foreground cues. These two components are coupled through a unified foreground consistency loss that jointly optimizes region-wise and voxel-wise consistency under a similarity-aware weighting scheme. Extensive experiments on three public 3D medical image segmentation benchmarks show that ForeMatch achieves state-of-the-art boundary accuracy with competitive region overlap, while introducing no additional learnable parameters and maintaining high computational efficiency.

Abstract:
Navigation instruction generation from ego-centric RGB video in continuous environments is an important yet challenging task for human--robot interaction and scalable dataset construction. Prior instruction generators assume discrete viewpoint graphs with panoramic observations, where trajectory structure is explicit; in continuous environments, however, the agent receives only a dense RGB stream, making trajectory cues difficult to recover. We propose VTInstructor, the first VLN instruction generation framework for continuous environments. Our key idea is to convert implicit trajectory geometry into explicit visual trajectory prompts: EDTC condenses long RGB trajectories into navigation-critical keyframes, VTP overlays path, turn, and goal cues onto these anchors, VTMod injects the resulting trajectory signals into the visual encoder, and VT-GRPO further calibrates this spatial injection during training, all without requiring a navigation graph, pre-built map, or scene reconstruction. On the challenging R2R-CE and RxR-CE Val Unseen benchmarks, VTInstructor sets a new state of the art across all standard NLG metrics, surpassing the strongest baseline by +0.357 CIDEr and +0.109 CIDEr, respectively. Beyond automatic metrics, VTInstructor-generated instructions raise a frozen follower's success rate to 63.3%, a +14.7 percentage-point gain over the best competing instruction source, and provide consistent data augmentation gains of +3 SR points on downstream navigation tasks.

Abstract:
As generative image editing advances, image manipulation localization (IML) must handle both traditional manipulations with conspicuous forensic artifacts and diffusion-generated edits that appear locally realistic. Existing methods typically rely on either low-level forensic cues or high-level semantics alone, leading to a fundamental micro--macro gap. To bridge this gap, we propose FASA, a unified framework for localizing both traditional and diffusion-generated manipulations. Specifically, we extract manipulation-sensitive frequency cues through an adaptive dual-band DCT module and learn manipulation-aware semantic priors via patch-level contrastive alignment on frozen CLIP representations. We then inject these priors into a hierarchical frequency pathway through a semantic-frequency side adapter for multi-scale feature interaction, and employ a prototype-guided, frequency-gated mask decoder to integrate semantic consistency with boundary-aware localization for tampered region prediction. Extensive experiments on OpenSDI and multiple traditional manipulation benchmarks demonstrate state-of-the-art localization performance, strong cross-generator and cross-dataset generalization, and robust performance under common image degradations.

Abstract:
Class-Incremental Learning (CIL) aims to sequentially learn new classes while mitigating catastrophic forgetting of previously learned knowledge. Conventional CIL approaches implicitly assume that classes are morphologically static, focusing primarily on preserving previously learned representations as new classes are introduced. In practice, however, instances of the same semantic class may undergo substantial morphological evolution, such as a larva turning into a butterfly. Consequently, a model must both discriminate between classes and adapt to evolving appearances within a single class. To systematically address this challenge, we formalize Stage-Aware CIL (Stage-CIL), a paradigm in which each class is learned progressively through distinct morphological stages. We further introduce Stage-Bench, a 10-domain, two-stage benchmark and protocol for evaluating both inter-class forgetting and stage-level degradation within classes. Finally, we propose STAGE, an evolution-aware reference baseline that disentangles semantic identity from evolution dynamics through a fixed-size memory pool, enabling stage-aware prediction of later morphological forms from earlier representations. Extensive experiments show that conventional CIL reductions and existing continual-learning baselines remain insufficient under Stage-CIL, while STAGE consistently outperforms strong competitors, demonstrating the promise of explicit evolution-aware modeling for this new setting.

Abstract:
Retrieval-augmented generation (RAG) systems increasingly combine multi-turn reasoning with heterogeneous evidence sources such as text passages and knowledge graphs. Existing hybrid RAG pipelines, however, usually rely on fixed graph-text fusion and optimize retrieval behavior mainly through answer correctness and retrieval cost. Such designs overlook a key signal available during retrieval: different evidence channels can have different levels of uncertainty for each sub-query. We propose Seek-R1, an uncertainty-aware multi-turn RAG framework that learns to route, retrieve, and answer with graph-text evidence. Seek-R1 introduces entropy-guided hybrid retrieval, which adaptively weights text and graph evidence according to retrieval confidence; a confidence-calibrated reward, which augments accuracy and cost feedback with an evidence-confidence term; and variance-adaptive GRPO, which compresses policy-ratio updates when group rewards provide weak preference signals. The framework preserves standard hybrid RAG training as a special case: fixed fusion, no confidence reward, and a unit adaptive exponent recover the conventional graph-text GRPO pipeline. Experiments on knowledge-intensive QA benchmarks show that Seek-R1 improves average answer quality, with gains concentrated on multi-hop QA, while maintaining efficient retrieval behavior.

Abstract:
Wearable inertial motion capture (MoCap) provides a portable, occlusion-free, and privacy-preserving alternative to camera-based systems, but its accuracy depends on tightly attached sensorsan intrusive and uncomfortable requirement for daily use. Embedding IMUs into loose-fitting garments is a desirable alternative, yet sensorbody displacement introduces severe, structured, and location-dependent corruption that breaks standard inertial pipelines. We propose GID (Garment Inertial Denoiser), a lightweight, plug-and-play Transformer that \emphfactorizes loose-wear MoCap into three stages: (i) location-specific denoising, (ii) adaptive cross-wear fusion, and (iii) general pose prediction. GID uses a \emphlocation-aware expert architecture, where a shared spatio-temporal backbone models global motion while per-IMU expert heads specialize in local garment dynamics, and a lightweight fusion module ensures cross-part consistency. This inductive bias enables stable training and effective learning from limited paired loosetight IMU data. We also introduce GarMoCap, a combined public and newly collected dataset covering diverse users, motions, and garments. Experiments show that GID enables accurate, real-time denoising from single-user training and generalizes across unseen users, motions, and GarMoCap garment layouts. GID is plug-and-play for representation-compatible tight-IMU pose estimators, requiring no downstream retraining and improving multiple evaluated estimators such as PIP and ASIP.

Abstract:
Multimodal action reasoning is a challenging problem due to redundant backgrounds, occlusions, viewpoint changes, etc. Existing approaches typically rely on either image or skeleton alone and fail to jointly model scene context and 3D human actions. To achieve action reasoning more accurately, we construct an unconventional multimodal large language model (MLLM) that employs compact skeleton manifold rough embeddings to complement defective image inputs sufficiently. Specifically, a novel rough tokenizer on hyperbolic manifold is proposed to learn hierarchical local-global geometric topologies of 3D skeleton sequences to build a comprehensible motion vocabulary for MLLM. Furthermore, to mitigate hallucination risks caused by cross-modal noise, we introduce a skeleton-image complementary fusion method that dynamically aligns and synchronizes multimodal information. Finally, we design meticulously a multimodal anticipation enhancement fine-tuning strategy based on a four-stage training paradigm to enhance spatio-temporal action reasoning capacity of MLLM effectively. Extensive experiments and visualization analysis demonstrate the outperformance of our MLLM across diverse benchmarks.

Abstract:
Due to an over-reliance on language priors and insufficient cross-modal alignment, large visionlanguage models (LVLMs) often suffer from visual hallucinations, i.e., systematic inconsistencies between generated text and visual content. While recent training-free methods mitigate hallucinations by enhancing vision-sensitive attention heads, they rely on heuristic layer selection and fixed enhancement strategies, and fail to explicitly model inference-time visionlanguage alignment. In this work, we propose a Dynamic VisionLanguage Alignment (DVLA) framework for mitigating visual hallucinations in LVLMs. DVLA consists of three principled components: (1) Dynamic Layer Selection, which identifies vision-sensitive layers by quantifying hidden-state discrepancies between multimodal and text-only representations; (2) VisionText Fusion Optimization, which regularizes cross-modal attention distributions to improve robustness and consistency; and (3) Confidence-Aware Dynamic Enhancement, which adaptively modulates enhancement strength based on multimodal divergence and attention confidence. Results on the CHAIR, POPE, and LLaVA-Bench benchmarks show that our DVLA dynamically aligns vision and text representations at inference time, substantially reducing visual hallucinations with minimal overhead and no loss in language quality. On CHAIR, it achieves a 12.7% reduction in caption-level hallucinations (CHAIR_S) compared to the latest state-of-the-art HIRE method.

Abstract:
Cloth-Changing Person Re-Identification (CC-ReID) aims to identify individuals across cameras over extended periods, where clothing changes require clothing-invariant identity representations. Existing multi-branch methods combine complementary cues such as facial and body features with quality-aware weighting, but optimize each branch independently for individual instances. Such an instance-wise learning paradigm lacks explicit constraints to align representations of the same identity across instances, leading to cross-instance discriminative misalignment when branch reliability varies. To address this, we propose a Consistency-Aware Branch Alignment Network (CABAN), which promotes cross-instance discriminative consistency by supervising body feature learning with reliable facial embeddings before fusion. CABAN consists of three sequential modules. The Relative-Margin Quality Estimator (RMQE) produces annotation-free quality scores from EMA-stabilized discriminative margins, capturing feature discriminability rather than perceptual sharpness. The Quality-Guided Branch Alignment (QGBA) module leverages these quality scores to align body representations with the discriminative dimensions of reliable facial features using quality-weighted contrastive learning and logit distillation. Finally, the Quality-Guided Residual Fusion (QGRF) module performs reliability-aware residual fusion by considering both feature quality and semantic consistency, ensuring robust integration even under degraded facial representations. Extensive experiments on multiple CC-ReID datasets demonstrate that our method achieves state-of-the-art performance.

Abstract:
Radio astronomy plays a crucial role in understanding the universe, particularly within the realm of non-thermal astrophysics. Images of celestial objects are derived from the signals (called visibility) measured by radio telescopes. Such imaging results, called dirty images, contain artifacts due to factors such as sparsity and therefore require reconstruction to improve imaging quality. Existing methods typically restrict reconstruction to a unimodal domain, either to the dirty image after imaging or to the sparse visibility prior to imaging. Focusing solely on each unimodal reconstruction results in the loss of complementary in-context information in either the visibility or image domain, leading to an incomplete modeling of mutual dependency and consistency. To address these challenges, we propose CDCRec, a multimodal radio interferometric data reconstruction method that explicitly models cross-domain consistency. We design a hierarchical multi-task and multi-stage framework to enhance the exploration of interplays between domains during reconstruction. Our experimental results demonstrate that CDCRec improves imaging performance through enhanced cross-domain correlation extraction. In particular, our self-supervised complementary modeling strategy is better than current methods at interferometric domain translations that rely heavily on recovering dense information from constrained source-domain data.

Abstract:
Camera-free 3D human pose estimation has attracted growing attention for privacy-sensitive applications such as healthcare, rehabilitation, and smart living. Among radio-frequency sensing modalities, WiFi CSI and mmWave radar provide complementary cues for device-free pose estimation: mmWave offers relatively stronger local geometric evidence, while WiFi provides coarse but informative motion and structural context. However, existing multimodal fusion methods typically rely on global and relatively uniform aggregation, which is insufficient for dual-radio pose regression where modality usefulness is highly heterogeneous and can vary across body joints. In this paper, we present SEE, an interpretable dual-radio framework for camera-free 3D human pose estimation. Instead of treating fusion as a monolithic global operation, SEE formulates dual-radio perception as structured evidence modeling followed by joint-wise reliability routing. Specifically, modality-specific encoders first extract WiFi and mmWave features, and a lightweight cross-modal interaction module enables contextual evidence exchange. The interacted representations are then explicitly decomposed into shared evidence and modality-private evidence, allowing the model to preserve both cross-modal consensus and sensor-specific cues. Based on these structured representations, a joint-aware reliability router adaptively determines, for each body joint, how much the final prediction should rely on WiFi-private, mmWave-private, or shared evidence, yielding a more effective and interpretable fusion process. Extensive experiments on a large-scale multimodal human sensing benchmark show that SEE consistently outperforms single-modal and conventional fusion baselines, validating the benefit of modeling multimodal collaboration at the joint level for reliable and transparent camera-free 3D pose estimation.

Abstract:
3D Gaussian Splatting (3DGS) has recently achieved impressive performance in novel view synthesis and real-time rendering. However, its geometric reconstruction quality in indoor scenes remains limited, especially in large weak-texture regions such as walls, ceilings, and floors, where sparse initialization and purely photometric supervision often lead to floaters, distorted surfaces, and structural ambiguities. To address this issue, we propose a planar prior guided 3D Gaussian Splatting framework for indoor scene reconstruction. Our method first extracts dense geometric priors, including depth and normal maps, from input images to provide robust structural guidance for optimization. Based on these priors, we further construct planar priors through multi-granularity image segmentation and geometric consistency verification, enabling reliable identification of planar regions in complex indoor scenes. We then incorporate region-aware geometric optimization into the 3DGS pipeline, including planar region constraints, non-planar local geometric regularization, and global geometry supervision, to jointly improve structural fidelity and surface consistency. Extensive experiments on multiple indoor datasets demonstrate that the proposed method produces more accurate meshes and more faithful rendering results than recent 3DGS-based baselines, particularly in weak-texture and piecewise-planar areas.

Abstract:
Diffusion alignment methods typically rely on static reward models or fixed datasets to guide optimization. However, as the generative model improves, frozen evaluation criteria fail to adapt to the shifting alignment frontiers, which progress from satisfying basic constraints to pursuing fine-grained criteria. To expand this ceiling, we propose Self-Evolving Preference Optimization (SEPO), a framework that continuously refines its evaluation criteria alongside the models post-training. SEPO leverages a multi-role agent system to maintain a dynamic natural language playbook. Through a closed-loop mechanism of judgment, critique, and arbitration, this playbook evolves to capture increasingly sophisticated visual preferences. To utilize the evolved rules, we introduce a semantic injection strategy. We incorporate rule-based visual descriptions as augmented conditions only during DPO training. This allows the model to explicitly internalize specific visual attributes into its parameters, enabling superior generation with original prompts at inference time. Extensive experiments in both offline and online settings across SDXL and FLUX architectures demonstrate SEPO's superiority. It achieves state-of-the-art performance, improving ImageReward by 15.4% and HPSv2 by 5.2%, while offering interpretable insights into the alignment process.

Abstract:
Generating executable CAD code from dimension-annotated orthographic drawings is a challenging task requiring geometric understanding, procedural reasoning, and precise numerical prediction. Existing vision-language approaches typically formulate this problem as one-shot generation, preventing the model from inspecting intermediate CAD results and correcting early mistakes, often leading to non-executable code or geometrically inconsistent outputs. In this paper, we propose IterCAD, an iterative framework that reformulates orthographic-view-to-CAD generation as a progressive program repair process. Instead of predicting the final CAD code in a single pass, IterCAD repeatedly analyzes the current CAD result, reasons about its discrepancy with the target views, and explicitly decides whether to \textttREVISE the code or \textttSTOP the refinement process. To make iterative repair learnable, we further construct IterCAD-RS, a structured revise-or-stop supervision set containing both repairable intermediate CAD states and already-correct states, and develop a three-stage training strategy for initial generation, revision learning, and multi-turn RL optimization. By closing the loop between visual understanding, geometric verification, and code refinement, IterCAD progressively corrects structural and parametric errors. Experiments on CADExpert show that IterCAD consistently improves code executability and geometric fidelity over strong one-shot baselines.

Abstract:
Traditional video-induced physiological datasets usually rely on whole-trial labels, which introduce temporal label noise in dynamic emotion recognition. We present FIRMED, a peak-centered multimodal dataset based on an immediate-recall annotation paradigm, with synchronized EEG, ECG, GSR, PPG, and facial recordings from 35 participants. FIRMED provides event-centered timestamps, emotion labels, and intensity annotations, and its annotation quality is supported by subjective and physiological validation. Benchmark experiments show that FIRMED consistently outperforms whole-trial labeling, yielding an average gain of 3.8 percentage points across eight EEG-based classifiers, with further improvements under multimodal fusion. FIRMED provides a practical benchmark for temporally localized supervision in multimodal affective computing.

Abstract:
UAV visual grounding in real-world applications requires localizing a target referred to by language while both the target and the UAV move. Existing UAV grounding datasets mainly focus on images, while the few video-based benchmarks are still dominated by appearance and spatial cues. As a result, they do not adequately capture two key challenges: motion-centric grounding and drastic cross-view appearance changes caused by UAV ego-motion. To address this gap, we introduce MoRe-UAV, a large-scale benchmark for motion-aware visual grounding in UAV videos. MoRe-UAV contains 22,225 video-expression pairs and 7,415,622 annotated frames, covering diverse aerial scenes with moving targets and substantial viewpoint changes. We build the dataset through a scalable human-in-the-loop pipeline for efficient annotation with quality control. We establish an initial benchmark on MoRe-UAV with spatio-temporal video grounding methods, multimodal large language models, and hybrid MLLM+tracking pipelines. We further provide a stronger baseline with a Motion-aware Prefix Adapter and a Multi-view Alignment Adapter to enhance motion reasoning and cross-view alignment. Experiments show that existing methods struggle on MoRe-UAV and remain far below human performance, highlighting substantial room for future research on motion-aware and multi-view grounding in UAV videos. Project page: https://more-uav.github.io/

Abstract:
State-of-the-art models for audio-driven digital human generation have achieved photo-realistic results in talking-head synthesis. However, extending these models to singing-head synthesis remains challenging due to a significant Domain Gap: singing requires more exaggerated expressions, vivid jaw openings, and precise rhythmic synchronization. Current models, primarily trained on speech datasets, often struggle with ``rhythmic drift'' and constrained dynamics. To address this, we introduce Hi-Singers, the first large-scale, high-quality, in-the-wild video dataset specifically tailored for singing head synthesis. Hi-Singers undergoes a rigorous automated and manual filtering pipeline, ensuring strict thematic adherence, high-resolution rendering, and stable motion, resulting in 29,608 video segments totaling approximately 170 hours. We further establish a dedicated evaluation benchmark balanced across linguistic and musical styles. Extensive experiments across diverse architectures---including 3D-coefficient and diffusion-based models---demonstrate that Hi-Singers consistently and significantly improves performance across all dimensions. Specifically, it enables models to achieve superior visual realism, enhanced lip-sync consistency, and more precise rhythmic dynamics, effectively bridging the domain gap and setting a new performance standard for the singing synthesis task. The dataset is available at \urlhttps://huggingface.co/datasets/CharlesZhang-USTC/Hi-Singers.

Abstract:
Speculative Jacobi decoding (SJD) is a widely used method for accelerating autoregressive (AR) image generation, but its effectiveness is often limited by token selection ambiguity. Recently, existing SJD methods mainly attempt to address this problem by relaxing verification conditions, yet the root causing such token selection ambiguity remains unclear. To figure out such reason, in this paper, we conduct a detailed analysis, and then find that image semantics are often encoded across multiple consecutive tokens, while current methods verify tokens individually, which breaks semantic continuity and amplifies token ambiguity. To this end, instead of performing the speculative verification on token level, we turn into the token-phrase level, and then present a novel speculative Jacobi decoding with phase verification for accelerating AR image generation. Specifically, we first constructs a token phrase library from large-scale image datasets to capture token phase that represent meaningful semantic. Then, we treat the token phrase library as a prior and design a token-phrase-level verification strategy to perform parallel speculative verification. In particular, our method is plug-and-play and can seamless integration with existing SJD methods. Extensive experiments on various datasets show our method can achieve significant acceleration after applying it into existing methods, while strictly maintaining high visual quality.

Abstract:
Cross-image comparative reasoning remains challenging for vision-language models (VLMs), especially when correct prediction requires fine-grained attribute grounding and globally consistent reasoning. We present CoRe, a unified framework for this problem. CoRe includes: (i) CoRe-20K, a large-scale triplet-based training set automatically constructed from structured visual metadata through a multi-expert collaborative pipeline, covering counting, depth, distance, and spatial relations; (ii) TriSR, a structured reward framework that jointly supervises attribute grounding, judgment alignment, and triplet consistency under GRPO optimization; and (iii) CoRe-Bench, the first benchmark dedicated to fine-grained cross-image comparative reasoning. Experiments show that CoRe substantially outperforms existing VLMs on CoRe-Bench while remaining competitive on standard multimodal benchmarks, achieving a 28.2-point gain in partial accuracy over the strongest baseline.

Abstract:
Vision-Language-Action (VLA) models have recently attracted growing attention in end-to-end (E2E) autonomous driving. While recent works have begun exploring Chain-of-Thought (CoT) reasoning to decompose complex driving tasks, they predominantly rely on purely textual reasoning. Driving, however, is an inherently spatial and physical task. Textual-only CoT struggles to capture dynamic geometric changes, frequently leading to numerical hallucinations and unreliable trajectory planning. To bridge the gap between semantic reasoning and the physical action space, we propose CoT4AD, a novel VLA framework that introduces a Hybrid Visual-Language CoT. Grounded in human cognitive Situation Awareness (SA) theory, CoT4AD strictly enforces a causal reasoning sequence: 3D spatial perception, semantic comprehension via Visual Question Answering (VQA), and physical projection via VLM-conditioned future scene diffusion. This hybrid chain explicitly forces the VLM to respect physical laws and spatial constraints before outputting driving trajectories. Furthermore, to balance explicit interpretability with the stringent latency demands of real-world deployment, we introduce an Implicit CoT model via Latent Alignment. This approach distills the multi-step causal knowledge directly into the VLM's hidden states, enabling highly efficient inference without sacrificing reasoning capability. Extensive experiments on both real-world and simulated benchmarks demonstrate that CoT4AD achieves state-of-the-art performance. Notably, it reaches a leading Driving Score of 81.22 on the Bench2Drive closed-loop benchmark and significantly reduces collision rates in nuScenes open-loop evaluations, proving the necessity and effectiveness of physics-grounded visual reasoning in E2E autonomous driving.

Abstract:
\beginabstract The rapid growth of 3D sensing technologies has led to an explosion of point cloud data, making point cloud geometry compression (PCGC) increasingly critical. Recent studies have explored large language models (LLMs) for data compression by leveraging their powerful sequence modeling capabilities. However, existing LLM-based PCGC methods treat point clouds as flat sequences, overlooking the intrinsic hierarchical structure of tree-based geometric representations. This mismatch limits the ability of LLMs to capture coarse-to-fine spatial dependencies, resulting in suboptimal compression performance. To address this issue, we propose TRLM-PCGC, a tree-structure-aware framework that explicitly incorporates hierarchical spatial priors into LLM-based point cloud compression. Specifically, we design a hierarchical tokenization scheme that converts point clouds into structure-preserving token sequences via Z-order chunking and dodecatree-based serialization, while level tokens are inserted to encode tree-depth information. To better align spatial tokens with the language model vocabulary, we further introduce an occupancy-guided token mapping strategy that associates structurally important tokens with higher-frequency tokens in the pretrained language model. In addition, we propose a Tree-Structure-Aware (TSA) loss that rebalances optimization across different tree levels, mitigating the prediction difficulty imbalance during autoregressive modeling. Extensive experiments demonstrate that TRLM-PCGC consistently outperforms existing methods. Our method achieves average bitrate reductions of 30.1% on MVUB and 29.2% on 8iVFB compared with G-PCCv23, while also surpassing the LLM-PCGC baseline. On the SemanticKITTI dataset, our method achieves a BD-rate reduction of 5.78% compared with the state-of-the-art method SCP-EHEM. \endabstract

Abstract:
Autoregressive (AR) image generation has shown strong potential for scalable high-fidelity synthesis by modeling images as discrete token sequences. However, traditional next token prediction (NTP) continues to suffer from sparse and myopic supervision, insufficiently discriminative representations, and high training cost caused by dense computation over the full token sequence. To address these issues, we propose multi-token autoregressive (MTAR), a unified training framework that improves autoregressive image generation from three aspects: prediction objectives, representation regularization, and training efficiency. Specifically, MTAR introduces multi-token prediction (MTP) to alleviate the sparsity and myopia of traditional NTP by imposing joint supervision on multiple future tokens; employs token-level contrastive regularization (TCR) to explicitly enhance the separability of sampled token representations and thereby improve representation discriminability; and incorporates semantic dropping (SD) as a semantics-aware training acceleration strategy to reduce redundant computation on low-information tokens while preserving informative learning signals. All three components are applied only during training and introduce no additional overhead during autoregressive inference. On ImageNet, MTAR achieves a better balance between generation quality and training efficiency. Compared with LlamaGen, MTAR achieves up to 0.95 lower FID and 39% faster training. Moreover, even with only 1/3 of the training iterations, it still attains performance comparable to or better than the baseline, substantially reducing training time.

Abstract:
Crafting high-quality scientific surveys demands not only profound literature insights but also intuitive, professional diagrams to facilitate comprehension. However, existing automated survey generation systems predominantly rely on shallow retrieval and mere information stacking across massive literature, failing to uncover the intrinsic logic of technological evolution. Furthermore, these systems are incapable of producing expert-level academic diagrams, resulting in significant limitations in both content depth and visual representation within the generated surveys. To address these critical gaps, we propose the Deep Reasoning and Multimodal Agentic Framework for Evolutionary Survey Synthesis (MESS). This framework is designed to mine literature depth through semantic expansion and recursive retrieval, while synergizing with vision models to generate professional academic diagrams. Notably, MESS pioneers a Limitation-Driven Recursive Retrieval mechanism: the system proactively extracts unresolved limitations from cornerstone papers as seeds to directionally retrieve subsequent solution literature. Through this mechanism, MESS automatically constructs evolutionary chains between papers, synthesizing survey content with a rigorous and traceable academic trajectory. Furthermore, to objectively measure the true value of automated surveys, we establish a multi-dimensional closed-loop evaluation system, introducing novel quantitative metrics such as Logical Contribution Density (LID) and Semantic Drift Rate (SDR). Extensive experiments on an independently curated dataset of 100 diverse research topics demonstrate that MESS significantly outperforms state-of-the-art baselines. It not only achieves unprecedented logical density in text but also pioneers the generation of academically rigorous, hallucination-free multimodal diagrams, approaching human-expert performance.

Abstract:
High-resolution monocular depth models based on ViTs achieve strong zero-shot performance, but their quadratic attention cost limits practical deployment. Reducing the input resolution sacrifices fine-grained spatial detail important for dense prediction. Internal token compression can lower cost without changing the input interface by reducing the token count after patch embedding, but simple token pooling consistently degrades depth accuracy because later transformer blocks and the decoder depend on spatially structured intermediate features. We therefore propose a lightweight adaptation framework for compressed-token depth models. The encoder is compressed at an intermediate stage, after which the remaining blocks are adapted with RepLoRA, a reparameterizable low-rank module that combines linear and spatial convolution branches in qkv projections and is deployed in a branch-free fused form at inference. Positional embeddings are jointly tuned, and the decoder preserves its original four-feature interface by combining features from before and after compression. Training uses approximately 200K pseudo-depth images with offline teacher supervision. Experiments on NYUv2, KITTI, and DA-2K show that RepLoRA consistently recovers the degradation introduced by pooling alone and provides a better accuracy-efficiency trade-off under the fixed-input setting. In representative compression settings, the proposed models reduce GFLOPs from 42.7 to 21.8--29.4, reduce attention cost from 25.03M to 5.47M--13.29M, and increase TensorRT throughput on RTX 3080 from 76.3 FPS to 112.0--164.5 FPS without changing the input interface.

Abstract:
Audio-driven instrument performance motion generation plays a pivotal role in applications such as digital content creation and instrument pedagogy systems. However, instrument performance is inherently continuous, long-duration, and rich in semantics. Synthesizing long-term performance motions that precisely align with musical rhythms while adhering to a performer's unique style remains a significant challenge. Furthermore, integrating interpretable semantic constraints into audio-conditioned motion generation poses an additional challenge. To address these limitations, we propose CLARA (Cycle-consistency based Latent-diffusion model for Audio-driven Realistic piano performance Animation), a novel framework for audio-driven piano performance motion generation that employs a transformer-based latent diffusion model combined with a powerful music representation model. First, to synthesize complex, long-term motions that precisely respond to the musical descriptions, our model generates motion primitives conditioned on prior motion history and current audio inputs, thereby facilitating the continuous generation of extended motion sequences. Second, to capture the intricate dynamics of the performer-instrument interaction, we introduce an exemplar-based style encoder that guides the diffusion generation process conditioned on reference performance clips, tailoring the output to the performer's distinct personal style without compromising performance synchronization quality. Finally, to enhance the physical interpretability of performance motions, we train a motion-to-music model that predicts acoustic-related targets directly from motion primitives. Through acoustic consistency supervision, these predictions facilitate alignment between the generated motions and the semantic context of the musical audio, effectively improving the kinematic-acoustic consistency of the interaction. Experimental results demonstrate that our proposed approach outperforms state-of-the-art methods, yielding highly natural and expressive piano performance motions. The source code will be released.

Authors: Shibo Gao (Beijing Jiaotong University), Chongxiao Wang (HUJING Digital Media & Entertainment Group), Chenglong Huang (HUJING Digital Media & Entertainment Group), Jie Ma (HUJING Digital Media & Entertainment Group), Haolin Shi (HUJING Digital Media & Entertainment Group), Fei Ding (HUJING Digital Media & Entertainment Group), Jing Li (HUJING Digital Media & Entertainment Group), Qiang Lyu (University of Chinese Academy of Sciences), Yangyang Liu (Institute of Automation, Chinese Academy of Sciences), Yang Liu (Tongji University), Jun Liu (Lancaster University), Linlin Huang (Beijing Jiaotong University), Peipei Yang (Institute of Automation, Chinese Academy of Sciences)

Abstract:
Real-world video reasoning often involves multimodal, multi-source inputs, whereas existing video reasoning tasks typically assume a simplified video-text setting, limiting identity matching and person-centric reasoning. To bridge this gap, we introduce the Identity-conditioned Queries (ICQ) task, in which models are required to jointly associate and interpret an input video and a reference image of a person, and leverage this conditioning to address identity grounding, behavior understanding, and temporal reasoning, among other challenges. Building on ICQ, we present ISYV (I Seek You in Videos), a systematic solution comprising three components: (1) ISYV-Bench, a challenging evaluation benchmark with 1,377 real-world complex videos and 1,377 questionanswer pairs, organized into six difficulty levels spanning capabilities from identity recognition to causal reasoning; (2) ISYV-75K, a large-scale training set of 75K high-quality samples constructed via automated annotation, multi-stage verification, and manual review; and (3) ISYV-Framework, containing an ICQ-oriented model and training strategy for learning to exploit informative video shots without additional shot-level annotations. Extensive experiments show that both mainstream closed-source and open-source MLLMs struggle on ISYV-Bench, especially in cross-domain identity matching and long-horizon tracking. ISYV-Model outperforms strong baselines and in some aspects approaches closed-source performance. Overall, ISYV provides a unified task definition, scalable datasets/benchmarks, and modeling insights for person-centric video reasoning.

Abstract:
Internet videos are a promising complement to motion-capture data for text-to-motion generation, but most web shots do not contain a fully visible body. Existing pipelines either keep all recovered joints as if they were equally reliable targets or discard incomplete clips altogether, yet both clip-level policies are mismatched to web motion with mixed-quality body parts. We study this setting as part-level credible supervision, where observed parts provide reliable learning targets while unobserved parts should not directly define the loss. We propose RoPAR, a part-aware text-to-motion framework that uses only the credible body parts as direct supervision and infers the remaining parts from text and visible context. Specifically, we divide the body into five kinematic parts, estimate part credibility from web videos, learn a latent motion space from the credible part motions, and use masked generation to reconstruct the full motion sequence. We also build K700-M, a dataset of 198k motion-text pairs mined from Kinetics-700 for studying this setting. On K700-M, RoPAR outperforms both clip-level alternatives under matched Keep-All and Discard protocols, reducing the best baseline FID by 19.1% and improving the best baseline R@1 by 20.3% in the scale-preserving Keep-All setting. These results suggest that, for web-recovered motion, keeping partially useful clips and restricting supervision to image-supported body parts is more effective than clip-level filtering.

Abstract:
2D Gaussian Splatting (2DGS) stores images as explicit primitives and decodes them by rasterization, running at real-time speed with no learned decoder. However, current methods lack a principled, spatially adaptive Rate-Distortion (RD) mechanism. They allocate primitives based on heuristic posterior errors, which leads to suboptimal capacity distribution and slow convergence. Furthermore, their use of global uniform quantization is fundamentally incompatible with adaptive bitrate control. To address these, we propose ScoutGS, a 2DGS image representation and compression framework centered around a prior-guided patch-level RD allocation strategy. It partitions the image with a content-adaptive quadtree and pre-trains a lightweight dual-stream RD surrogate to predict local coding costs. This surrogate assigns initial Gaussian budgets before optimization and supports a global Lagrangian search for patch-wise quantization profiles at a target bitrate. On Kodak and DIV2K, ScoutGS outperforms previous 2DGS methods in both representation and compression while keeping the same 2DGS decoder. With 5K primitives on Kodak, ScoutGS converges 38% faster than state-of-the-art 2DGS baselines and exceeds their final quality. Under target compression ratios on the Kodak dataset, ScoutGS achieves a Bjntegaard Delta (BD) PSNR gain of 0.93\,dB and a BD-rate reduction of 24.45% over baselines, while keeping the realized bitrate within 8% of the target.

Abstract:
Sleep staging is essential for sleep quality evaluation and personalized treatment of sleep disorders. Traditional polysomnography (PSG)-based multimodal methods provide accurate sleep staging but require complex setup and cause discomfort to users, which hinders their daily applications. Specifically, their unimodal adaptation is limited because cross-modal knowledge complements disrupt the optimization trajectory of the primary modality, while the distribution discrepancy between multimodal training and unimodal inference further degrades generalization. To address these problems, this paper proposes CMGC, a curriculum-regulated multimodal gradient complement framework for ubiquitous sleep staging. To mitigate the distribution mismatch between multimodal training and unimodal inference, we design a curriculum learning scheme that progressively attenuates the updates of auxiliary modality and multimodal interactions. Moreover, to enable compatible transfer of multimodal knowledge while preserving the unimodal learning trajectory, we exploit the principal gradient subspace of the primary modality to inject auxiliary gradients onto its complementary directions. Our CMGC achieves state-of-the-art performance on three multimodal sleep datasets. CMGC demonstrates significant performance improvement in unimodal deployment, highlighting its effectiveness for ubiquitous sleep monitoring.

Abstract:
Large language models (LLMs) have demonstrated significant potential in semantic audio-visual navigation (SAVN). However, audio signals are inherently sparse and intermittent, and existing methods often perform poorly under silent conditions because they fail to maintain persistent and spatially grounded guidance when the audio becomes silent. To address this challenge, we propose a novel Memory-Grounded Navigation (MGN) framework, whose key idea is to decouple navigation decisions from instantaneous observations and ground them instead in persistent audio-spatial memory organized for structured decision-making. Specifically, MGN first converts transient audio cues into temporally accumulated spatial evidence, then organizes this evidence together with visual semantics into a decision-oriented graph memory, and finally performs structured reasoning over this memory to select exploration actions. In this way, the agent can continue making robust and goal-oriented decisions even when the audio goal becomes silent. Extensive experiments show that MGN achieves state-of-the-art Success When Silent (SWS), outperforming existing LLM-based methods by 7.1%.

Abstract:
3D Gaussian Splatting (3DGS) enables real-time novel view synthesis but requires storing a large number of primitive parameters, limiting deployment on memory-constrained devices. Existing compression methods largely ignore the non-uniform sensitivity of human perception and therefore preserve redundant anchors and feature channels in perceptually insensitive regions. We propose KLT-GS, a perception-aware 3DGS compression framework that uses the number of KarhunenLove Transform (KLT)-based critical perceptual components (CPCs) as a feed-forward prior to jointly guide anchor pruning and feature truncation. Specifically, we estimate anchor perceptual importance for adaptive pruning and anchor perceptual sensitivity for dynamic feature truncation. Experiments on five benchmark datasets show that KLT-GS reduces model size by 49.6% on average over HAC++ while maintaining competitive objective and subjective quality, yielding an improved compression-quality trade-off.

Abstract:
Generating lifelike facial expressions in embodied robots requires not only accurately mapping semantic facial representations to servo control signals under physical actuation constraints, but also overcoming deviations arising from silicone skin drift and actuator or sensor inaccuracies. Such deviations accumulate during operation, resulting in drift, jitter, and unstable expression generation. Moreover, randomly collected motor data often fail to reflect natural expression semantics and real-world embodied actuation responses. To overcome these challenges, we present ServoFormer, a Transformer-based framework for embodied facial expression control, together with a self-supervised data construction pipeline. By retargeting natural human expressions to a humanoid robot and synchronously capturing its physical responses, we build a temporally aligned embodied facial expression dataset of 39,090 frames. Based on this dataset, ServoFormer maps blendshape sequences to servo commands using an encoder--decoder architecture with Locally Constrained Self-Attention to suppress long-range mechanical noise propagation and Strictly Aligned Cross-Attention to enforce frame-level correspondence. Experiments show that ServoFormer significantly outperforms existing methods in single-frame accuracy, velocity consistency, and expression recognizability, generating more coherent and lifelike robotic facial expressions.

Abstract:
Audio-driven emotional talking face generation aims to synthesize realistic videos with expressive facial dynamics. However, existing methods struggle to balance controllability and visual fidelity. Although implicit representations capture rich semantics, they lack structural guidance, often resulting in averaged emotional expressions. In contrast, explicit geometric methods offer better control over facial expressions but tend to sacrifice high-frequency texture details. To address it, we propose GemTalk, a diffusion-based framework that combines the semantic richness of implicit representations with the structural precision of explicit geometric priors. We introduce a Vision-guided Audio Emotion Projection (V-AEP) module to extract implicit emotional lip and expression features. At the same time, a Diffusion-based Geometric Priors Generator (D-GPG) generates identity-aware blendshape coefficients as explicit structural priors. Crucially, our Geometry-guided Emotion Modulation (GEM) module leverages these geometric priors to recalibrate the magnitude of implicit features, enabling precise, continuous control over emotional expressions, especially emotion intensity, without sacrificing visual quality. Extensive experiments show GemTalk achieves superior performance in photo-realism, and facial emotional dynamics.

Authors: Hongyuan Liu (University of Electronic Science and Technology of China), Qinli Yang (University of Electronic Science and Technology of China), Wen Li (University of Bristol), Zhong Zhang (University of Electronic Science and Technology of China), Jiaming Liu (University of Electronic Science and Technology of China), Wei Han (University of Electronic Science and Technology of China), Zhili Qin (Institute of Applied Electronics, China Academy of Engineering Physics), Jinxia Guo (University of Electronic Science and Technology of China), Junming Shao (University of Electronic Science and Technology of China)

Abstract:
Vision-Language Models (VLMs) such as CLIP learn a shared embedding space for images and text, yet their representations remain geometrically separated, a phenomenon known as the modality gap. This gap limits tasks requiring cross-modal interchangeability, such as captioning and joint clustering. Existing post-processing approaches can partially improve cross-modal compatibility; however, we show through geometric analysis that they primarily reduce the global centroid offset while leaving the underlying distributional mismatch intact. We decompose the modality gap into a Centroid Gap and a Distribution Gap, and demonstrate that the Distribution Gap is the true predictor of cross-modal task quality (R^2 = 0.986), whereas the commonly used Raw Gap is misleading (R^2 = 0.691). Motivated by this observation, we propose TPC-CMA (Three-Phase Curriculum for Cross-Modal Alignment), a fine-tuning framework that explicitly reduces both components. The proposed CMA jointly mitigates centroid offsets and reshapes the distributional structure, while a three-phase curriculum with gradient-aware scheduling progressively introduces alignment during training to enable stable optimization. Experiments demonstrate that our method significantly improves cross-modal alignment. With \alpha_\texttarget=0.05, the modality gap is reduced by 66.6% with only 4.84% accuracy drop. Under stronger alignment (\alpha_\texttarget=0.5), the gap is reduced by 82.3%, clustering ARI improves from 0.318 to 0.516, and captioning CIDEr increases by 57.1% over the original model.

Abstract:
Harmful meme detection aims to identify whether a meme carries harmful content by understanding and leveraging both its visual and textual modalities. The natural language rationales generated by Large Language Models (LLMs) delve into the implicit meaning of memes, which can facilitate the determination of whether the meme is harmful or harmless. The harmful memes often have a specific target scope, such as individual impact, group identity, or social context. Analyzing memes from the perspectives of potential targets can provide sound rationales for harmful content inference. However, rationales generated by existing methods are not oriented towards these specific targets. Besides, the LLM-generated rationales also suffer from the issues of hallucinations, ambiguities, and even errors, that will mislead the harmful meme detection. This paper proposes a Rationale Adaptive Fusion Target-guided Agent Network (RAFTNet) that first generates target scope oriented rationales, then screens and fuses reliable rationale information for harmful meme detection. In RAFTNet, we design a Hierarchical Multi-Agent Target Rationale Generation (HMA-TRG) module to acquire visual and textual cues from memes, which are next used to generate individual impact, group identity and social context rationales respectively. Then an Adaptive Rationale Fusion Detection Network (ARF-DN) is constructed to screen reliable rationales and fuse them with multimodal meme features for harmful detection. Further, we introduce a lightweight RAFTNet-D based on knowledge distillation, offering a plug-and-play solution enabling efficient deployment in practical scenarios. Extensive experiments on four public benchmarks show that our RAFTNet significantly outperforms the state-of-the-art algorithms, and RAFTNet-D maintains competitive performance without substantial parameters and external LLM knowledge.

Abstract:
3D Gaussian Splatting (3DGS) enables high-fidelity immersive rendering, but scene streaming for live 6DoF viewing still depends on anticipating and prioritizing content before user attention shifts. Because the client cannot deliver the whole scene at once, the core systems problem is not simply how much data can be sent, but which content should receive the limited budget before attention moves. Current systems commonly organize requests spatially, whereas user attention often concentrates on semantic objects. This mismatch wastes transmission budget on collateral context and weakens post-saccade recovery under tight runtime constraints. In our live evaluation, one important limitation in this setting is not only prediction accuracy, but also whether the request interface can directly expose attended content. We propose OC-GS (Object-Centric Gaussian Streaming), a two-stage framework for demand-aligned delivery. In the offline stage, semantic--saliency co-distillation converts frozen 3DGS into executable object-addressable transmission units comprising macro background nodes, hierarchical chains for large objects, and atomic micro-packs for fine details. In the online stage, object-level demand prediction and Lyapunov stability-aware control convert short-horizon demand estimates into feasible low-churn requests under coupled resource budgets. Crucially, representation, prediction, and control are coupled through the same executable object-level action space. Against native L3GS/LapisGS baselines and an A3FR variant with only a thin transport adapter, OC-GS primarily improves recovery and stability rather than chasing the full-resident upper bound: in live eye-tracked 6DoF Meta Quest sessions on Mip-NeRF360, ScanNet++, and Replica, it reduces recovery latency from 139--192~ms to 72--78~ms and refetch ratio from 5.4--9.4% to 0.8--2.3%, while improving Foveated-PSNR by 0.87--1.77~dB over A3FR. Our project page is available at \urlhttps://naislab.cn/OC_GS.

Abstract:
Deep learning has shown strong potential in medical image analysis, but most existing methods rely on large-scale annotations and a closed-world assumption that rarely holds in clinical practice. Although Generalized Category Discovery (GCD) has advanced rapidly on natural images, it remains underexplored in medical imaging. To address this issue, we propose MedXplore, a unified framework for reliable and unbiased medical GCD, optimizing from both perceptual and decision levels. Specifically, at the perceptual level, taking a frequency domain perspective, Frequency-SNR Adaptive Attention and Consistency (FAAC) performs learnable full-spectrum filtering and global-local energy contrast activation to not only highlight local abnormal signals relative to the global context, but also provide reliable semantic anchors for patch consistency learning. At the decision level, Adaptive Cosine-Angular Margin (ACAM) adjusts angular margins using semantic difficulty and feature confidence to balance intra-class compactness and inter-class separability. Together, the two modules improve lesion-sensitive representation learning and mitigate old-class bias. Experiments on multiple benchmarks show an average 8.5% gain in All accuracy over the strongest competing methods. On Kvasir, MedXplore reduces false-old errors from 14.50% to 0.80%, demonstrating strong robustness under severe old-new ambiguity.

Abstract:
Visual token compression has emerged as a critical technique for reducing the inference cost of Vision-Language Models. Existing methods select tokens based on text relevance, visual diversity, or a fixed combination of both---treating the relevance-diversity balance as a universal constant independent of the input. We conduct a systematic analysis to challenge this assumption. At the dataset level, grid-searching the optimal relevance-diversity weight per benchmark reveals substantial cross-task variation. More critically, we further maximize per-sample answer likelihood to recover sample-level oracle weights, finding that their distributions not only differ markedly across benchmarks but also exhibit high within-benchmark variance. Based on these observations, we propose SAVER, a framework for Sample-Aware relevance-diversity balanced Visual tokEn Reduction. We first derive a Geometric Intent Prior (GIP) from the spatial concentration of text-relevant tokens to infer a sample-specific relevance-diversity weight, and then apply this weight to jointly select tokens via a Weighted Text Centroid (WTC) for relevance scoring and a Visual Marginal Coverage Gain (VMCG) for diversity. SAVER requires no additional training and achieves sample-aware balancing entirely at inference time. Extensive experiments demonstrate that SAVER consistently outperforms state-of-the-art baselines across diverse benchmarks, retaining 96.5% of full-model performance on LLaVA-1.5-7B with only 11.1% of the original visual tokens, and further generalizing to LLaVA-NeXT and Mini-Gemini.

Abstract:
The rapid evolution of diffusion models has democratized face swapping but also raises concerns about privacy and identity security. Existing proactive defenses, often adapted from image editing attacks, prove ineffective in this context. We attribute this failure to an oversight of the structural resilience and the unique static conditional guidance mechanism inherent in face swapping systems. To address this, we propose VoidFace, a systemic defense method that views face swapping as a coupled identity pathway. By injecting perturbations at critical bottlenecks, VoidFace induces cascading disruption throughout the pipeline. Specifically, we first introduce localization disruption and identity erasure to degrade physical regression and semantic embeddings, thereby impairing the accurate modeling of the source face. We then intervene in the generative domain by decoupling attention mechanisms to sever identity injection, and corrupting intermediate diffusion features to prevent the reconstruction of source identity. To ensure visual imperceptibility, we perform adversarial search in the latent manifold, guided by a perceptual adaptive strategy to balance attack potency with image quality. Extensive experiments show that VoidFace outperforms existing defenses across various diffusion-based swapping models, while producing adversarial faces with superior visual quality.

Abstract:
Multimodal gradient balancing methods modulate encoder gradients with a shared scalar per modality, implicitly assuming that corruption is uniform across the training batch. In practice, corruption is sample-heterogeneous: within a single mini-batch, different samples may have different modalities corrupted. We prove that under this heterogeneous corruption model, any batch-level sample-agnostic linear estimator with a shared modulation parameter incurs an irreducible bias with respect to the clean-data gradient, and that sample-level all-or-nothing gating is the unique unbiased strategy within a natural distribution-free estimator class. Motivated by this result, we propose Sample-Adaptive Gradient Gating (SAGG), which makes a binary retain-or-discard decision per sample via an online feature-norm quality test and incorporates a truncation mechanism for variance control. We prove that SAGG-based SGD converges at the standard O(1/?T) rate to stationary points of the clean loss without a corruption-dependent error floor, and derive a certified robustness radius for the independent-encoder architecture that connects per-modality Lipschitz constants to the classification margin. Experiments on Kinetics-Sounds and UCF-101 under Gaussian noise injection, partial modality missing, and natural contribution imbalance show that SAGG consistently outperforms ten existing methods, with the largest gains in high-corruption regimes where batch-level bias is most severe.

Authors: Jiaxian Xu (Guangdong Provincial Key Laboratory of Ultra High Definition Immersive Media Technology, Shenzhen Graduate School, Peking University), Wei Jiang (Guangdong Provincial Key Laboratory of Ultra High Definition Immersive Media Technology, Shenzhen Graduate School, Peking University), Feng Wang (Guangdong Provincial Key Laboratory of Ultra High Definition Immersive Media Technology, Shenzhen Graduate School, Peking University), Hongwei Hu (Ant Group), Qi Wang (Ant Group), Qunshan Gu (Ant Group), Feng Gao (School of Arts, Peking University), Chao Wang (Pengcheng Laboratory), Ronggang Wang (Guangdong Provincial Key Laboratory of Ultra High Definition Immersive Media Technology, Shenzhen Graduate School, Peking University)

Abstract:
Recent advances in one-step diffusion-based image compression leverage large pre-trained text-to-image latent diffusion models. These methods enable visually realistic reconstruction at ultra-low bitrates with a single diffusion step. Despite these advances, existing methods suffer from suboptimal encoder designs and mismatched latent distributions. As a result, the extracted latents fail to effectively guide one-step diffusion, and the generative capacity of the pre-trained VAE decoder is not fully utilized. To address these issues, we propose ADSEC, which enhances perceptual quality through highly expressive latents and distribution alignment. Specifically, we design a flexible single-stage encoder equipped with Adaptive Feature Extraction Modules to produce highly expressive latents for guiding one-step diffusion. We further introduce Latent Space Adversarial Training to align the latents with the pre-trained VAE latent distribution, improving generalization and fully leveraging generative priors. Additionally, a joint pixel-domain perceptual loss is employed to provide complementary perceptual supervision. Experiments demonstrate that ADSEC achieves state-of-the-art perceptual quality. Compared to StableCodec, ADSEC achieves 20% to 30% bit-rate savings on the CLIC 2020 dataset, measured by FID and DISTS. Furthermore, ADSEC offers faster decoding and lower memory consumption than prior one-step diffusion-based methods.

Abstract:
Stray animals frequently experience poor welfare and face complex risks in urban environments. While Virtual Reality (VR) perspective-taking effectively fosters empathy, existing approaches primarily rely on single-user, passive embodiment, limiting the understanding of dynamic human-animal interactions. We present Pawspective, an asymmetric multisensory VR system that integrates reciprocal humananimal role-play to explore cross-species empathy. Set in a shared urban simulation, two participants concurrently experience the environment through fundamentally distinct audiovisual and interactive modalities: a human worker managing time and financial constraints, and an injured stray dog navigating survival. Their gameplay is driven by continuous mutual influence, where interdependent choices directly impact both players' outcomes. Results from our counterbalanced role-swapping study (N=32) show that participants reported more positive attitudes toward stray dogs after completing the Pawspective experience, while order-dependent differences were observed in state empathy and in-game prosocial behavior. These findings suggest that Pawspectives integrated system design offers a promising approach to empathy-oriented interactive multimedia experiences.

Abstract:
While neural lossless image compression has advanced significantly with learned entropy models, lossless video compression remains largely unexplored in the neural setting. We present NeuralLVC, a neural lossless video codec that combines masked diffusion with an I/P-frame architecture for exploiting temporal redundancy. Our I-frame model compresses individual frames using bijective linear tokenization that guarantees exact pixel reconstruction. The P-frame model compresses temporal differences between consecutive frames, conditioned on the previous decoded frame via a lightweight reference embedding that adds only 1.3% trainable parameters. Group-wise decoding enables controllable speed--compression trade-offs. Our codec is lossless in the input domain: for video, it reconstructs YUV420 planes exactly; for image evaluation, RGB channels are reconstructed exactly. Experiments on 9 Xiph CIF sequences and on 24 HD sequences at 720p and 1080p (Xiph, UVG, and the modern BVI-AOM benchmark) show that NeuralLVC outperforms H.264 and H.265 lossless by a significant margin at every resolution tested. We verify exact reconstruction through end-to-end encode--decode testing with arithmetic coding. These results suggest that masked diffusion with temporal conditioning is a promising direction for neural lossless video compression.

Abstract:
Multimodal Large Language Models (MLLMs) have demonstrated remarkable performance in visionlanguage reasoning. However, the hallucination issue continues to limit their reliability in high-risk scenarios. Existing research predominantly focuses on post-hoc output-layer correction and heuristic interventions, while few studies investigate the underlying hallucination mechanisms from the perspective of internal semantic evolution. Through empirical analysis, this paper reveals that the reasoning process of mainstream MLLMs follows a three-phase semantic evolution patternexplorati-on, anchoring, and convergence. Hallucinations arise when the semantic evolution fails to converge toward the visual evidence. We theoretically prove that ensuring such convergence effectively suppresses hallucination generation and propose a phase-specific intervention method: during the anchoring phase,visual information is strengthened to guide the model toward reasoning directions grounded in true visual evidence. In the convergence stage, critical visual details are enhanced to ensure convergence to the correct visual evidence.The proposed method requires no additional training, can be flexibly adapted to various architectures, and extensive experiments validate its effectiveness in mitigating hallucinations.

Abstract:
Layout-guided text-to-3D generation has significantly advanced the controllable synthesis of multi-object scenes. However, existing methods still struggle to maintain consistent and precise spatial relationships, frequently exhibiting inaccurate relative positions and ambiguous object orientations. To bridge this gap, we introduce SpatialGS, a novel, plug-and-play framework that seamlessly incorporates relative spatial regularization into Gaussian Splatting. Specifically, we propose two core modules: Relative Position Regularization (RPR) and Relative Orientation Regularization (ROR), which explicitly enforce accurate inter-object spatial relationships during optimization. To further stabilize this generative process, we design an Alternating Global-to-Local Sampling Strategy (AGS) that harmonizes global scene context with local object details. Extensive experiments demonstrate that SpatialGS can be easily integrated into existing Score Distillation Sampling (SDS)-based pipelines, significantly outperforming state-of-the-art baselines in both spatial consistency and overall compositional quality to yield physically plausible 3D scenes.

Abstract:
Large VisionLanguage Models (LVLMs) are increasingly expected to perform visual question answering (VQA) over planar media. However, existing planar VQA benchmarks typically emphasize isolated challenges: some emphasize long-document understanding with limited reasoning depth, while others require complex visual reasoning but remain restricted to single-page, noise-free settings. Moreover, through theoretical analysis, we identify the impact of irrelevant visual tokens, which leads to measurable performance degradation but has received little attention with respect to systematic quantification. To address these limitations, we introduce TestHallVQA, a multi-image VQA benchmark that simultaneously embodies document-level scale and the difficulty of human examinations, while providing comprehensive task coverage. Leveraging TestHallVQA's ability to controllably inject multi-level contextual redundancy, we further propose a novel metric, F1-R, which jointly quantifies LVLMs' computational reasoning capability and their evidence retrieval robustness against document-level redundancy. Extensive experiments and analyses on mainstream LVLMs reveal their latent deficiencies across multiple dimensions, offering concrete insights and directions for future research.

Abstract:
Video multimodal large language models have advanced significantly, yet fine-grained motion-temporal understanding remains fragile. The core bottleneck is not only sparse frame sampling, but also the lack of a complete temporal modeling pipeline for explicitly representing frame-to-frame change, enabling appearance-motion interaction, and optimizing temporal direction sensitivity. We propose COMET, a temporally grounded framework that systematically strengthens video MLLMs through explicit temporal representation, appearance-motion fusion, and direction-aware optimization. Architecturally, COMET introduces a temporal motion branch built on Taylor frame differences and injects its motion evidence into the appearance stream via temporal attention bias-enhanced cross-attention. For optimization, COMET combines temporal prior distillation with a forward-reverse TC-GRPO stage that turns temporal order into a direct learning signal and strengthens the model's use of directional motion patterns encoded by the temporal motion branch. The method achieves consistent overall improvements with a pronounced motion-temporal bias: on Qwen3-VL-8B, action-centric tasks (STAR, SSv2) improve by 4.9% on average, temporal reasoning tasks (NExT-QA, CLEVRER, LLaVA-178K) by 2.1% over BL-GRPO, while static perception tasks (PerceptionTest) remain on par. The same gain pattern also transfers to InternVL2.5-8B, indicating that COMET generalizes across model families.

Abstract:
Point clouds are essential for immersive media applications such as autonomous driving and virtual reality, yet their large volume poses challenges for storage and transmission. To address this, 3D point clouds are projected onto 2D video frames using Video-based Point Cloud Compression (V-PCC), allowing them to be compressed with standard video codecs. However, the encoding processparticularly inter codingremains computationally intensive due to complex partition decisions. Thus, this paper focuses on accelerating inter-coding partition for both geometry and attribute sequences. Specifically, we study three unique relationships of point cloud videos: the correlation between geometry and attribute sequences, the near-layer and far-layer dependency within projected frames, and the temporal correlation across adjacent frames. Each of these relationships exhibits a strong correlation. Based on these relationships, we design a dedicated dataset, develop an LSTM network that effectively captures these characteristics, and propose a tailored loss function that explicitly leverages them, enabling efficient far-layer CU partition prediction. Experiments on standard V-PCC sequences show that our method reduces encoding time by an average of 74.11% while maintaining compression performance, achieving BD-Rate reductions of 0.12% (D1) and 0.07% (D2), and outperforming other existing fast CU decision approaches.

Abstract:
Visual search is ubiquitous in daily life, from locating keys in a messy apartment to finding a tool in a cluttered workshop or an application icon on a desktop screen. While visual attention is to a large extent driven by image content, research in psychology indicates that user's cognitive states have a significant influence on attention allocation. A promising way to access information about cognitive states is via Electroencephalography (EEG) recordings. Despite this potential, previous works in computational attention modeling have largely disregarded EEG, and related studies in psychology operate with abstract, simplistic stimuli, or even artificially guided search procedures. We bridge this gap with two contributions. First, we present \dataset, the largest publicly available multimodal dataset of free visual search with realistic stimuli. The dataset contains synchronized EEG and eye-tracking recordings from 36 participants across 140 naturalistic scenes, covering two scenarios: searching for desktop icons and finding tools in cluttered workshop scenes. Second, we introduce \method, the first method for human attention prediction in visual search that integrates EEG signals. Our approach consists of a EEG-guided refinement stage that adapts the output of a state-of-the-art saliency prediction method to the user's current cognitive state. In comprehensive evaluations on \dataset, we show that \method\ consistently outperforms all baselines and ablation conditions.

Abstract:
Recent advances in sign language understanding have achieved impressive success on short, single-sentence videos, yet their performance drops sharply when applied to long, continuous sign language videos. To bridge this gap, we focus on a challenging and realistic setting: Visual-only Sentence-level Sign Language Segmentation (Vis-SSLS), which aims to partition continuous sign language videos into non-overlapping sentence-level segments without any caption assistance, serving as a crucial prerequisite for downstream recognition and translation tasks. However, sentence transitions in sign language are often smooth and visually ambiguous, lacking explicit pauses or posture resets. As a result, static frame representations may fail to capture the subtle temporal changes that indicate sentence boundaries. To address this challenge, we propose SignShift, a difference-aware segmentation framework that explicitly models frame-to-frame feature variation as semantic cues for sentence boundary detection. First, to model the feature variation, we design a Temporal Difference Module, which incorporates full-frame, facial, and hand cues, and employs inter-frame differencing to learn multi-scale temporal variations that capture both fine-grained local kinematics and global semantic transitions. Second, to mitigate over- and under-segmentation issues, we design a Segment Count Prediction module, which predicts the number of sentences to guide boundary selection. Extensive experiments on benchmark datasets demonstrate that SignShift substantially outperforms existing methods, validating its effectiveness.

Abstract:
Referring Audio-Visual Segmentation (Ref-AVS) seeks to localize and segment target objects in video frames based on visual, auditory, and textual referring cues. The task is challenging because the relevance of different modalities varies across referring expressions and scenes, while existing methods typically treat multimodal cues as homogeneous inputs for fusion, prompting, or reasoning, making them vulnerable to irrelevant or misleading modalities. To address this problem, we propose PRIMED, inspired by the biased competition theory in cognitive neuroscience, which explicitly models both visual perception and language-driven prior modulation, and enables more accurate Ref-AVS by adaptive modality suppression. Specifically, a Modality Prior Decoder first estimates whether the referring expression relies primarily on audio, vision, or their joint interaction, generating a modality prior to adaptively guide high-level attention. A Token Compressor further extracts compact global visual tokens from high-level features and shares them across Cross-modal Biased Competition Fusion modules to provide hierarchical global context. Additionally, we introduce a Spatial-Aware Semantic Alignment loss to further enhance foreground-background discrimination through contrastive learning. Extensive experiments on the Ref-AVS benchmark demonstrate that PRIMED achieves state-of-the-art overall performance.

Authors: Zhenrong Cheng (School of Intelligence Science and Technology, State Key Laboratory of General Artificial Intelligence, Peking University), Hengyi Cai (Chinese Academy of Sciences), Yuchen Li (Shanghai Jiao Tong University), Hao Sun (School of Intelligence Science and Technology, State Key Laboratory of General Artificial Intelligence, Peking University), Shuaiqiang Wang (Baidu Inc.), Chunyu Yang (Ucap Cloud Information Technology Co.,Ltd), Dawei Yin (Baidu Inc.), Yan Zhang (School of Intelligence Science and Technology, State Key Laboratory of General Artificial Intelligence, Peking University)

Abstract:
High-performance image--text retrieval (ITR) relies on large-scale multimodal data with fine-grained semantic alignment, yet real-world data exhibits pronounced \emphcross-modal information asymmetry: images are often rich while texts are sparse or ambiguous, leading to biased representations and unstable retrieval. To address this, we propose AIM (Asymmetry-aware Information Modeling), an information-theoretic framework that explicitly models and exploits modality asymmetry via a unified information signal. Our metric jointly captures \emphinformation deficiency (through directional likelihood gaps from InfoNCE) and \emphinformation ambiguity (via entropy-based uncertainty), yielding a sample-wise estimate of information gain. Based on this signal, AIM follows a coherent \emphdiagnose--augment--optimize paradigm: it identifies imbalance regions in an information space, performs targeted MLLM-driven augmentation to construct the INFO-ITR dataset, and applies an Information-Theoretic Weighted Loss (ITWL) to prioritize informative samples during training. Extensive experiments on multiple benchmarks show that AIM consistently outperforms strong baselines and improves robustness under noisy, sparse, and imbalanced supervision. These results highlight the importance of explicitly modeling information asymmetry for stable and balanced cross-modal alignment.

Abstract:
Scene-text aware cross-modal retrieval requires models to jointly capture visual text content and object semantics. However, existing methods often exhibit an inherent imbalance: they either over-rely on fine-grained textual cues and become sensitive to noise, or overlook critical textual information, resulting in suboptimal alignment. To address this issue, we propose HRU, a parameter-efficient framework that Harmonizes Reading and Understanding via a two-stage progressive learning strategy and orthogonality constraints. In the first stage, visual prompts disentangle object-level and text-level representations, enabling adaptive scene text encoding while preserving visual semantics. In the second stage, a fusion mechanism progressively aligns and refines these features, achieving coherent heterogeneous visual information integration. With a parameter-efficient design, HRU can be seamlessly adapted to multimodal backbones of varying scales with minimal fine-tuning overhead. Extensive experiments demonstrate that HRU significantly improves alignment and retrieval robustness, achieving RSUM scores of 474.5, 380.3, and 546.3 on CTC-1k, CTC-5k, and TextCaps, respectively, with fewer than 30K additional trainable parameters. These results establish HRU as an effective and unified solution for text-rich visual understanding and retrieval.

Abstract:
Despite remarkable progress in vision-language pretrained models, existing approaches such as CLIP still struggle to capture fine-grained semantic differences under high global similarity, limiting cross-modal retrieval performance. Existing methods typically construct hard negatives through semantic manipulation, yet these negatives are often overly separable from positives and thus insufficient for learning subtle discriminative cues. To address this issue, we propose CLIP-SPSE, a novel framework that enhances cross-modal representation learning via dual optimization: fine-grained semantic element-perturbed negative generation and semantic-enhanced contrastive encoding. We argue that effective hard negatives should satisfy both negativeness and hardness. Under these assumptions, we construct challenging negatives using triplets of fine-grained semantic elements. Specifically, an LLM generates perturbed negative texts, which are validated by a multi-LLM ensemble for semantic plausibility and discriminative hardness, while an image editing model synthesizes visual negatives under weighted similarity constraints to preserve negativeness with minimal perturbation to global semantics. Extensive experiments on benchmarks demonstrate that CLIP-SPSE significantly improves cross-modal retrieval.

Abstract:
Accurate boundary-aware segmentation of mycelium images is essential for fungal growth monitoring, morphological analysis, and agricultural pathological detection. Mycelium complex boundaries render edge regions uncertain, challenging traditional segmentation models. To quantify boundary uncertainty and exploit it for segmentation refinement, studies have adopted Evidential Neural Networks (ENNs) for distribution-level modeling. However, existing ENN methods misestimate uncertainty, as their optimization objectives over-sharpen predictive distributions, yielding overconfident estimations that oversmooth edge details and weaken multi-scale representations. Therefore, we propose Evidential Adaptive Constraint and Guidance (EACG), a novel framework that recalibrates overconfident predictions and enables adaptive feature aggregation. Specifically, EACG models uncertainty via structural constraints and integrates adaptive constraint relaxation to calibrate predictions. Furthermore, it employs calibrated uncertainty to adaptively aggregate multi-scale features. Experiments on MyceliumSeg and three external datasets demonstrate that EACG achieves state-of-the-art boundary segmentation performance.

Abstract:
Multimodal Retrieval-Augmented Generation (RAG) enhances Large Vision-Language Models (LVLMs) with external knowledge, yet prevailing frameworks are bottlenecked by static top-k retrieval. This rigid paradigm creates a fundamental trade-off: an expansive retrieval window introduces excessive visual distractors, while a restricted one precipitates recall collapse. To address this limitation, we present CoRe-RAG, a multimodal retrieval framework that performs adaptive evidence calibration instead of relying on a fixed ranking pipeline. Specifically, CoRe-RAG employs an uncertainty-aware calibration mechanism to adaptively filter high-recall candidates, constructing candidate sets that retain key evidence while discarding visually plausible distractors. Within the calibrated candidate set, we further introduce a utility-aware alignment strategy that prioritizes downstream generation utility over semantic similarity. Experiments on challenging multimodal benchmarks demonstrate that CoRe-RAG consistently improves retrieval quality and downstream performance, with gains of up to 15% on complex evidence localization tasks.

Abstract:
Multimodal Federated Learning (MFL) faces client-level modality missingness and Non-IID data distributions, which can induce substantial representation drift and hinder effective cross-client collaboration. Existing approaches commonly rely on generative imputation, external auxiliary data, or isolated unimodal training to bridge modality gaps, often introducing additional computational, communication, or privacy concerns. We propose FedTaste, a parameter-efficient framework for topology-aware structural transfer in MFL with missing modalities. Rather than relying solely on first-order alignment, FedTaste exploits transferable group-level semantic relations within a shared semantic-group space. Full-modality clients use frozen vision-language encoders with trainable Modality-Adaptive Structural Prompts to distill joint relational structure, which the server consolidates into a confidence-aware global blueprint. Missing-modality clients optimize only the prompts attached to their available branches and align their partial representations with the blueprint through topology and spectral regularization. FedTaste therefore avoids explicit modality imputation and raw-data exchange during federated optimization while preserving shared semantic structure across clients. Experiments on standard cross-modal retrieval benchmarks show that FedTaste consistently outperforms representative baselines under challenging Non-IID and modality-heterogeneous settings while maintaining favorable communication efficiency.

Abstract:
Existing robotic imitation learning methods rely on per-step point-wise supervision that regresses instantaneous actions without explicitly modeling the temporal dynamics of action sequences. This zeroth-order view yields oscillatory trajectories and hurts task success. To address this limitation, we propose HoTFlow, a higher-order temporal trajectory generation framework for robotic imitation learning. HoTFlow models action generation as a temporally continuous process by jointly aligning actions and their temporal variations with second-order constraints, and combining task-space dynamical consistency with decoupled supervision for continuous and discrete variables. Moreover, we reformulate the implicit optimization objective in MP1 as explicit velocity regression, augmented by a temporally localized bounded guidance mechanism. Consequently, HoTFlow yields more reliable direction estimation at test time and improves the stability of both training and inference. Extensive experiments on simulated and real-world robotic manipulation tasks demonstrate that HoTFlow achieves higher task success rates and smoother, more stable action trajectories, while preserving single-step inference efficiency.

Abstract:
Scene text image super-resolution (STISR) seeks to restore high-resolution text images with both visual clarity and textual fidelity. Existing methods often introduce OCR models to provide semantic priors for super-resolution. However, we observe a counter-intuitive failure mode: even when the low-resolution input is still readable, OCR-guided methods may generate incorrect characters after super-resolution. Beyond character errors, such guidance can also entangle text semantics with visual generation, leading to undesired hallucinations in character style and background content. This reveals that semantic priors are not universally beneficial in STISR, when structural evidence is sufficient, over-reliance on semantics can instead induce hallucination. Motivated by this observation, we propose HR-Diffusion, a hallucination-resistant diffusion framework for STISR that dynamically coordinates structure and semantics according to input readability. Our key idea is that semantics determines what to write, while structure constrains how it should be formed. Accordingly, we design a dual-branch diffusion architecture with an image diffusion branch (IDB) guided by OCR-derived semantic priors and a segmentation diffusion branch (SDB) specialized for character structure modeling. We further introduce a Triple-Evidence Gating Module (TEGM) to estimate input readability using OCR confidence, structural latents, and image-generation latents as three types of evidence, and to adaptively balance semantic and structural guidance during denoising. Extensive experiments demonstrate that our adaptive coordination of structural and semantic guidance effectively mitigates semantics-induced hallucinations while preserving character topology, visual style, and textual fidelity.

Abstract:
White balance (WB) aims to eliminate illumination-induced color casts and recover a neutral rendering, yet in consumer imaging, strictly neutral correction does not always yield the most visually preferred result, especially in scenes with mixed illumination, dominant color casts, or low-light conditions. In such cases, unreliable illuminant estimation may lead to results that are neither sufficiently faithful nor visually pleasing, suggesting that perceptual aesthetics can serve as an auxiliary guidance signal in these challenging cases when strict neutrality is difficult to achieve. We address this issue by formulating an aesthetic-aware white balance refinement (A^2WB) as a consumer-oriented extension of Automatic WB (AWB). Rather than regarding the task as general image enhancement, we frame it as a WB-centered iterative refinement process that keeps chromatic correction as the primary goal while allowing appearance refinement when it is advantageous. To achieve this, we propose A^2WB, a policy-based iterative framework operating over a disentangled action space with two components: core WB-related actions for chromatic refinement and optional appearance actions for auxiliary perceptual improvement. The framework is trained with a joint objective that balances aesthetic improvement, reference consistency, and artifact regularization. In this manner, A^2WB bridges the gap between neutral AWB and consumer-preferred rendering in a more systematic manner than either one-shot correction or general enhancement. Experimental and subjective results show that A^2WB improves perceptual preference while maintaining visually acceptable WB-related color consistency.

Abstract:
A paramount challenge in controllable diffusion models is the inherent conflict between guidance signals, such as structure and style, which compete within the latent feature space and lead to visual artifacts. Existing methods attempt to mitigate these conflicts through architectural modifications or heuristic fusion, but often ignore their root cause: the lack of a shared contextual basis among guidance signals. Without a unified structural prior, content and style remain entangled in the feature space, leading to unstable attention dynamics. This paper introduces Rationale, a novel hybrid framework that resolves this conflict by rationally combining a lightweight, trainable style encoder with two parameter-free, inference-time attention guidance modules. Our approach first leverages a dedicated projection network, Resampler, trained on our newly curated, style-balanced Subject-Materials-Style-Set dataset, to encode a reference image into a rich style representation. Then, during generation, our two modules ensure its harmonious application: 1) Relative Self-Attention Guidance actively corrects for structural distortions by steering self-attention queries towards a stable, pre-computed structural prior. 2) Style-Aligned Sparse Attention achieves precise, localized stylization by decoupling content and style at the query-level in the cross-attention mechanism. This hybrid design combines the rich expressive power of a trained encoder with the fine-grained control of dedicated, training-free guidance, avoiding the high costs of test-time optimization. Rationale demonstrates superior performance in structure-preserving stylization and photorealistic material transfer, validated by extensive quantitative and qualitative evaluations, achieving a state-of-the-art LPIPS score of 0.483 and highly preferred stylization quality by human evaluators.

Abstract:
Automatically generating cinematically expressive camera trajectories through 3D scenes from natural language descriptions is a challenging task of high practical value, with applications ranging from real-estate advertising to virtual tour creation. Existing methods either lack true 3D spatial awareness by relying on 2D image priors, or treat trajectory generation as a geometric path planning problem divorced from cinematographic semantics. We present CinemaTraj, a framework that reframes camera trajectory planning as a language-grounded spatial reasoning problem. Given a set of RGB-D images and a user prompt, CinemaTraj equips an LLM agent with a structured 3D scene graph: the agent decomposes the prompt into a sequence of atomic cinematographic movements (dolly, orbit, crane, pan, tilt, zoom, arc). Each movement is instantiated via a novel parametric trajectory representation that is both cinematographically expressive and optimizable for collision avoidance. The scene graph acts as a structured spatial prior, grounding the agents reasoning in accurate geometric and semantic knowledge of the environment. CinemaTraj further generates synchronized voiceover and subtitles aligned with camera motion, producing narrated cinematic video outputs. We evaluate CinemaTraj on real-world ScanNet++ environments, and show that it produces prompt-faithful, collision-free trajectories with high cinematographic quality, outperforming existing approaches on prompt alignment, trajectory quality, and safety metrics.

Abstract:
Multimodal recommendation leverages item multimodal features alongside collaborative signals to capture user preferences. While item-item graphs have become a key building block in advanced models, existing methods typically construct them with noisy similarity edges and limit their role to a single function of item-item representation propagation, leaving substantial potential untapped. In this paper, we propose IIMRec, a framework that constructs a single high-quality item-item graph during preprocessing and systematically reuses it across three stages of the recommendation pipeline: representation enhancement, interaction graph enhancement, and optimization enhancement. The graph is built by fusing semantic and co-occurrence signals, then refined via Neighborhood Consistency Edge Reweighting (NCER), which applies the triadic closure principle to amplify structurally reliable edges and suppress spurious ones. Once constructed, the graph is leveraged in three complementary ways: (1) Item-item propagation with a Residual II Gate (RIG) that adaptively controls per-item absorption of semantic neighborhood signals for representation enhancement; (2) A content-guided UI graph expansion that introduces virtual user-item edges through high-confidence semantic neighbors for interaction graph enhancement; (3) II-Neighbor BPR Augmentation (INA) that treats top neighbors of positive items as discounted soft positives for optimization enhancement. All components are computed or cached during preprocessing or add negligible per-batch cost, making IIMRec both effective and efficient. We provide theoretical analysis showing that NCER reduces the spectral noise-to-signal ratio, RIG converges to a non-degenerate gating regime, and INA yields a tighter generalization bound. Extensive experiments on four datasets demonstrate that IIMRec consistently outperforms state-of-the-art baselines while running faster and consuming less GPU memory, with particularly strong gains under cold-start and sparse-interaction conditions.

Abstract:
Effectively addressing the joint effects of non-IID data and long-tailed distribution remains a key challenge in federated learning. While fine-tuning vision-language models (VLMs) like CLIP has shown promise in tackling non-IID scenarios, it often degrades tail-class performance when both heterogeneity and long-tailed distributions are present, even leading to overall performance falling below the zero-shot baseline. To overcome this, we propose Class-Aware Prompt Tuning (CAPT), a novel framework leveraging pre-trained VLMs to effectively learn from non-IID and long-tailed data. Specifically, CAPT introduces a dual-prompt mechanism that synergizes class-aware and general prompts, enabling the model to preserve class-specific information while capturing global trends. Furthermore, to guide the prompts in learning targeted knowledge, we propose a heterogeneity-aware clustering strategy with two key components: similarity-based clusters aggregate class-aware prompts to reinforce tail-specific knowledge, while heterogeneity-based clusters aggregate general prompts to balance diverse representations and promote robust generalization. Extensive experiments demonstrate that CAPT significantly outperforms state-of-the-art methods, e.g., by improving overall accuracy by 6.1% on ImageNet-LT and 8.8% on Places-LT over baselines, while also substantially enhancing tail-class performance across other benchmarks.

Abstract:
Unified Multimodal Models (UMMs) unify multimodal understanding and image generation, but often suffer from an understandinggeneration gap, where correct semantic interpretation does not translate into faithful visual outputs. This gap stems from the lack of reflection and correction over intermediate generation processes, as existing methods supervise the final outputs. Since these intermediate states are implicit and not directly observable, we are motivated to exploit UMMs own multimodal understanding as semantic feedback to supervise and correct the generation process. In this paper, we propose a training-free framework to improve the generation ability of UMMs through CoT-based reflecting and correcting the intermediate visual representations. The proposed framework consists of two complementary modules. The CoT-based Reflection module performs global semantic diagnosis and refines textual conditions to correct high-level inconsistencies, while the Semantic-Driven Region Correction module evaluates token-level reliability and selectively resamples unreliable tokens to prevent local error accumulation. By injecting semantic feedback into intermediate stages, generation trajectories become semantically observable and controllable. Experiments demonstrate substantial improvements in compositional synthesis, boosting the base model from 68% to 79% on GenEval and achieving a 14% absolute gain on T2I-CompBench.

Abstract:
Controllable and navigable panoramic content generation is essential for immersive virtual reality applications. However, existing approaches often lack precise camera controllability or fail to capture the semantic context of navigation. To address this, we introduce PanoTask, a novel task-driven framework for generating panoramic videos that follow semantically meaningful trajectories conditioned on scene context and user intent. As no existing dataset provides aligned video-text-trajectory annotations for this formulation, we construct PanoTask-10k, a large-scale dataset containing approximately 10k video-text-trajectory triplets. Building upon this dataset, we propose a two-stage task-to-video pipeline. First, a multi-modal autoregressive model predicts coherent camera trajectories conditioned on task descriptions and environmental context. Then, a diffusion-based video generator produces high fidelity panoramic videos. To ensure motion consistency between the generated video and the predicted trajectory, we introduce a trajectory-conditioned structured noise initialization. Furthermore, we incorporate ERP-aware geometric constraints in latent space to ensure panoramic consistency. Experiments show that our framework outperforms existing baselines in visual fidelity and geometric consistency, advancing the generation of semantically navigable panoramic videos.

Abstract:
Translating medical images into accurate natural language descriptions is a critical challenge in cross-modal understanding. However, existing RAG-based methods typically retrieve references according to visual similarity, where visually similar images may correspond to different clinical findings and report semantics. Visually matched but semantically mismatched samples can introduce misleading multimodal evidence, thereby reducing the accuracy of generated reports. To address this challenge, we propose a hypothesis-contrastive differentiable retrieval augmented generation (HINT-RAG) framework, a perceive-query-disambiguate cross-modal learning paradigm for ambiguity-aware medical report generation. Specifically, the perceive module formulates uncertainty-aware hypotheses to capture plausible clinical semantics from the input image. The query module leverages these hypotheses to perform contrastive differentiable retrieval, aiming to identify discriminative evidence from clinically confusable candidates rather than merely visually similar samples. The disambiguate module adaptively fuses visual features and retrieved evidence to resolve cross-modal ambiguity before report decoding. Extensive experiments on public datasets show that HINT-RAG achieves leading performance in both natural language generation and clinical efficacy metrics.

Abstract:
Desktop interaction streams provide a continuous, privacy-sensitive record of interleaved user tasks. Transforming these streams into task-organized personal logs on-device faces two main challenges: exhaustive Vision-Language Model~(VLM) processing strains local resources, and global stream processing causes cross-task context pollution. We present FOCAL (Filtered On-device Continuous Activity Logging), a privacy-first multi-agent system utilizing a unified \emphfilter--plan--log architecture. It cascades a lightweight Filter Agent for noise suppression, a text-only Brain Agent for task attribution, a Record Agent for selective visual reasoning, and a task-isolated Memory Agent for context-coherent summarization. Experiments on DesktopBench (comprising 2,572 screenshots across 420 complex sessions) show FOCAL reduces total token consumption by 60.4% and VLM call count by 72.3% versus a baseline, while boosting Key Information Recall~(KIR) from 0.38 to 0.61. Crucially, under A\toB\toA task interruptions, FOCAL maintains Task~Acc~0.81 and KIR~0.80, whereas the baseline collapses to Task~Acc~0.03. FOCAL pioneers the efficient, on-device summarization of instruction-free desktop streams into multi-perspective personal logs.

Abstract:
We present SignDPO, a novel multi-level Direct Preference Optimisation (DPO) framework designed to enhance the alignment of skeleton-based Sign Language Translation. While current skeleton-based models have made significant progress using Maximum Likelihood Estimation, they are primarily constrained by an imitation-based paradigm that lacks discriminative sensitivity to the fine-grained spatio-temporal nuances of sign language, often leading to semantic drift. To address this, SignDPO shifts the optimisation goal from simple sequence mimicry to structured preference alignment across spatial, temporal, and linguistic dimensions. Our framework involves three key designs. First, we introduce a hierarchical perturbation strategy to construct spatial and temporal non-preferred samples at both global and local granularities automatically. Second, we propose a self-guiding mechanism that leverages decoder cross-attention scores to identify and perturb semantically salient skeletal regions, forcing the model to distinguish genuine sign signals from structural distortions. Third, we establish an automated language-level preference generator by fine-tuning a dedicated perturbation model, capturing complex output-level failure modes without manual annotation. Extensive experiments on three widely adopted benchmarks, CSL-Daily, How2Sign, and OpenASL, demonstrate that SignDPO consistently outperforms state-of-the-art gloss-free methods and even rivals established gloss-based ones. Our results suggest that multi-level preference alignment is a powerful paradigm for bridging the gap between high-entropy skeletal trajectories and discrete linguistic semantics.

Abstract:
Generalized Referring Expression Grounding aims to detect and segment an indefinite number of objects according to a given expression. Current methods can be categorized into global-level and object-level discrimination according to how they discriminate against no-target scenarios. The former methods segment target objects and discriminate no-target scenarios separately, leading to inconsistency between the mask and discrimination results. The latter methods generate per-object boxes or masks with corresponding object discrimination scores, leading to a lack of a global perspective to explore mismatches. To address this, we propose the Think Twice approach, which features a unique Two-level Mismatch Discriminator that performs discrimination at both object-level and global-level. Object-level ensures that the model identifies high-scoring objects within the target samples, mitigating the global-level's inconsistency limitation. Global-level incorporates the box prompts, and leverages cross attention at different feature levels, enhancing discrimination accuracy from a global perspective and overcoming the object-level's drawback. By leveraging their complementary strengths, each level addresses the core limitation of the other, resulting in more robust performance. The proposed discriminator can be readily integrated with existing detection models to empower traditional architectures with the capability to handle the generalized setting. Additionally, we propose a Ground-Truth Mixed-Prompt Generation algorithm to enhance the prompt stability in the training stage. Experiments on the gRefCOCO and Ref-ZOM datasets demonstrate the superior performance of our Think Twice approach in detecting mismatches, achieving a new state-of-the-art.

Abstract:
Generating high-fidelity, city-scale 3D environments plays a crucial role in applications such as virtual reality, gaming, and urban digital twins. However, the scarcity of real-world city-scale 3D training data forces existing methods to either stitch independently generated assets or learn from synthetic urban corpora, often sacrificing realism and global consistency. In contrast, 2D image data and their corresponding generative models are abundant and encode rich visual priors. To bridge this gap, we propose CityPaint, a novel framework that leverages 2D diffusion models to generate globally consistent city-scale 3D textures without relying on city-scale 3D training data. Specifically, CityPaint integrates geometric priors with the generative power of 2D diffusion models, introducing a robust top-K camera baking strategy for iterative texture projection. To address the artifacts of 2D-to-3D projection, we further propose a decoupled refinement mechanism that separately optimizes terrain and building textures. Extensive experiments show that CityPaint produces globally coherent textures with fine local details and achieves favorable results against strong baselines in both visual quality and cross-view consistency.

Abstract:
Score distillation sampling has become a cornerstone paradigm for text-to-3D generation, yet it is limited by mode-seeking behavior that leads to over-smoothing. Variational Score Distillation (VSD) offers a principled alternative by casting synthesis as probabilistic inference, but the computational cost increases substantially with the number of independently optimized particles. To address this trade-off, we introduce MorphDreamer, a novel framework for variational 3D scene generation that achieves both high-fidelity diversity and remarkable computational efficiency. Rather than operating on independent particles, MorphDreamer decomposes the originally intractable global optimization into a sequence of well-conditioned local subproblems by a progressive particle splitting strategy, effectively navigating the non-convex landscape and avoiding local minima. In addition, we uncover a critical misalignment between the entropy-maximizing nature of VSD and the inherently unstructured, discrete representation of 3D Gaussians, which introduces geometric ambiguity. To resolve this, MorphDreamer rebalances the variational lower bound with a Weighted Entropy-Regularized VSD (WER-VSD) objective, striking a principled balance between geometric determinacy and textural richness. Extensive experiments show that MorphDreamer reconciles diversity and fidelity with high efficiency, offering a pivotal advancement toward practical and expressive text-driven 3D content creation.

Abstract:
Procedural materials, typically represented as computational graphs, are widely adopted in 3D content creation because they enable editable, reusable, and highly controllable appearance design. To automate this expertise-intensive and time-consuming process, we study the problem of reference-guided stylized procedural material authoring for multi-object 3D scenes. This task poses three main challenges: (1) existing methods mostly assume single-object material optimization, which is insufficient for multi-object scenes; (2) stylized procedural materials are strongly coupled with object semantics and geometry; (3) procedural material authoring requires joint optimization over both structure and parameters, resulting in a highly complex optimization space. To address these challenges, we propose SceneMat, a reference-guided framework for procedural material authoring. Given a text or image reference, SceneMat first performs scene analysis to extract semantic, geometric, and spatial cues; it then refines scene-level references into object-level cues and performs material initialization; finally, it conducts material optimization through restricted VLM-driven iterative editing based on a structured intermediate representation. Experimental results show that the proposed method outperforms existing baselines in both image-guided and text-guided settings, achieving superior object-level and scene-level visual quality, as well as higher optimization correctness and efficiency.

Abstract:
Linguistic steganography enables covert communication through natural language. Existing methods heavily rely on token-level operations and struggle to maintain reliability under word- and sentence-level textual perturbations. Moreover, variable-length coding-based schemes are highly susceptible to bit-slippage under minor disturbances, as perturbations cause desynchronization between embedded and extracted bit sequences. To address these issues, we propose a robust coverless steganographic framework that operates in the sentence embedding space rather than the token space. Specifically, secret messages are encoded as hierarchical clustering paths in the sentence embedding space, which enhances decoding stability against word- and sentence-level textual perturbations. To tackle the bit-slippage problem, we introduce a Global Resynchronization Mechanism (GRM) that reframes variable-length bitstreams as discrete symbols anchored to semantic subspaces, decoupling local embedding failures from global message recovery. Experimental results demonstrate that under word- and sentence-level perturbations, our approach achieves substantial improvements in robustness, while maintaining effective embedding capacity and exhibiting strong resistance to statistical analysis.

Abstract:
Diffusion models have become the mainstream paradigm for modern visual generation and have substantially advanced multimedia content synthesis, especially in text-to-image and text-to-video tasks. To further align such generative models with human preferences, reinforcement learning (RL) has recently shown strong potential as a post-training strategy. Nevertheless, existing policy gradient-based methods often explore inefficiently, making them vulnerable to local optima that may degrade semantic faithfulness and visual realism. To address these challenges, we present Reflection-Aware GRPO (RA-GRPO), a new RL-based preference alignment framework for diffusion generative models. The core idea is to improve `forward' generation by incorporating `backward' reflection during optimization. We first introduce Diffusion Reflection, which rectifies intermediate sampling trajectories by inverting the diffusion process with a weak estimator, guiding latent states toward higher-probability regions of the true data manifold. Furthermore, we introduce Counterfactual Path Synthesis to implicitly distill these rectified trajectories into the policy, enabling the model to internalize the benefits of search-based exploration without incurring inference-time overhead. Extensive experiments on T2I and T2V models demonstrate that RA-GRPO significantly outperforms existing methods, particularly in mitigating reward hacking and improving generalization. The method remains architecture-agnostic and integrates seamlessly with standard pipelines, suggesting a promising direction for stable preference alignment.

Abstract:
Emotion analysis is a fundamental task in computer vision, but its practical deployment remains constrained by the privacy risks inherent to conventional RGB cameras. Bio-inspired event cameras present a promising hardware-level solution because they capture asynchronous brightness changes, thereby reducing exposure of facial identity details while leveraging high dynamic range for robust perception under challenging illumination conditions. Despite these advantages, existing event-based methods struggle in complex real-world settings due to limited dataset scales, simple acquisition conditions, and reliance on single-modality visual cues. To address these, we establish a challenging tri-modal benchmark with event, audio, and text modalities and propose the Information-Guided Gated Fusion (IGF) framework, which first pre-trains an event encoder on the FAU subset of Emo-DVS to capture fine-grained facial dynamics, then employs adaptive modality gating to suppress modality-specific noise, and finally leverages mutual information maximization to align robust cross-modal representations. To alleviate data scarcity, we introduce Emo-DVS, the first large-scale event-based emotion analysis dataset, which couples dynamic illumination with the Facial Action Unit (FAU) subset and emotion subset. Extensive experiments demonstrate that IGF achieves state-of-the-art performance.

Abstract:
Existing VLMs have achieved strong performance in video understanding, yet they struggle with long-video spatiotemporal reasoning when target objects become invisible, often mistaking "invisible" for "unknown". We define this challenge as hidden-state spatiotemporal reasoning: inferring object states during prolonged invisible intervals from context interactions. To address this, we propose StateTrace, a novel object-centric framework that endows VideoLLMs with an explicit mechanism for hidden state reasoning in long videos. StateTrace builds a reusable spatiotemporal state memory that organizes object trajectories, inter-object relations, and state-transition events into a structured reasoning substrate. At inference time, it retrieves question-relevant state-evolution trajectories and converts them into compact reasoning cues, enabling the model to explicitly reason about why an object disappears, how its state evolves while invisible, and whether that state should persist at query time. We further build HSR-Bench, a diagnostic benchmark for hidden-state reasoning, containing 1,427 video-QA samples from 1,384 unique videos. Extensive experiments across multiple VideoLLMs show that StateTrace consistently improves performance on both public benchmarks and HSR-Bench(e.g., improving VideoLLaMA3 from 39.6 to 64.2 on HSR-Bench).

Abstract:
Recent advances in instruction-driven image editing have made significant progress, yet challenges remain in achieving semantic alignment and precise localized editing. These limitations arise from misalignment between the multimodal large language model (MLLM) instruction understanding and reinforcement learning (RL) tuned diffusion models, as well as inefficient reward allocation in region-specific tasks. To bridge those gaps, we propose SREdit, a self-rewarding RL-based framework with a dual-shared reward mechanism and region-aware reward reweighting. The dual-shared reward mechanism unifies instruction understanding and reward evaluation within a single MLLM, and employs instruction decomposition to align semantic objectives with reward criteria. To further enhance localized editing precision, we introduce a region-aware reward reweighting strategy that spatially reweights the RL loss based on cross-attention maps, thereby focusing optimization on semantically relevant editing regions. Evaluations show SREdit's superior performance, improving overall scores by 3.8% on GEdit-Bench, 2.5% on ImgEdit-Bench, and 3.8% on KRIS-Bench.

Abstract:
Multimedia event extraction (M2E2) aims to predict triggers, ground arguments across text and images, and assemble them into schema-consistent event records. Recent LLM-based approaches have shown strong potential for M2E2, but their intermediate event hypotheses often remain implicit, and event-argument linking is still tightly coupled with role binding. This leaves little opportunity to inspect or revise intermediate event hypotheses and makes predictions brittle to early errors. To bridge this gap, we present ECHO, a multi-agent framework that reframes M2E2 as iterative refinement over an explicit Multimedia Event Hypergraph (MEHG). Instead of relying on implicit linear generation, ECHO performs auditable atomic updates over a shared hypergraph, making intermediate event structures explicit and revisable. Furthermore, we introduce a Link-then-Bind strategy that decouples event-argument linking from role binding, reducing premature semantic commitment during structured prediction. Extensive experiments on the M2E2 benchmark show that ECHO outperforms state-of-the-art approaches, achieving gains of 7.3 and 15.5 F1 points on event mention and argument role, respectively.

Abstract:
The rapid progress of image generation models calls for AI-generated image (AIGI) detectors that are not only accurate but also explainable and reliable. While MLLM-based detectors can provide natural language explanations, existing methods often generate speculative rationales: they rely on vague or hallucinated artifacts, miss subtle localized flaws from the latest generators, and fail to provide evidence that can be visually verified. We present Defake-o3, an explainable AIGI detector that moves from speculative rationales to verifiable evidence. It combines interactive visual search with verifier-guided evidence alignment: the model iteratively zooms into suspicious regions to inspect fine-grained details, while an Evidence Verifier, trained from human verification annotations, provides reinforcement learning rewards that favor grounded evidence and penalize baseless claims. To support this objective, we construct GroundFake, a dataset designed for grounded explainable detection, with localized bounding-box evidence, human verification based on visual grounding and artifact specificity, corrected reasoning trajectories, and valid/invalid evidence supervision. We further introduce FakeFrontier, an out-of-distribution benchmark built from real images and outputs of 10 recent generators, together with an MLLM-based protocol for evaluating evidence quality and persuasiveness. Experiments on GroundFake, FakeFrontier, and additional out-of-distribution benchmarks show that Defake-o3 improves both detection accuracy and explanation quality, producing more localized, verifiable, and persuasive evidence.

Abstract:
Proactive watermarking offers a promising approach for deepfake tamper detection and localization in short-form videos. However, existing methods often decouple audio and visual evidence and assume that watermark signals remain reliable under real-world degradations, making tamper localization vulnerable to multimodal misalignment and compression distortions. Moreover, existing semi-fragile visual watermarking methods often degrade significantly under codec compression because their embedding bands overlap with compression-sensitive frequency regions. To address these limitations, we propose Layered Audio-Visual Anti-tampering Watermarking (LAVA), a calibration-aware audiovisual watermark fusion framework for deepfake tamper detection and localization. LAVA leverages cross-modal watermark fusion and calibration-aware alignment to preserve consistent and reliable tamper evidence under compression and audiovisual asynchrony, enabling robust tamper localization. Extensive experiments demonstrate that LAVA achieves near-perfect detection performance (AP = 0.999), remains robust to compression and multimodal misalignment, and significantly improves tamper localization reliability over existing audiovisual fusion baselines.

Abstract:
\beginabstract Proactive defense methods protect portrait images from unauthorized editing or talking face generation (TFG) by introducing pixel-level protective perturbations, and have already attracted increasing attention for privacy protection. % In real-world use, images inevitably undergo sequences of benign operations during display across devices and subsequent dissemination, such as resizing and color compression, which directly alter pixel values. % Existing studies and robustness defenses have mainly examined individual transformations in isolation, while the effects of sequentially composed transformations remain underexplored. % To address this gap, we systematically evaluate whether representative proactive defenses against unauthorized image editing using GANs and diffusion models, as well as unauthorized talking face generation, remain effective under sequential image transformations. % The evaluated methods span both general purpose and portrait specific defenses and are assessed through qualitative and quantitative evaluation. % Experimental results indicate that defense methods based on pixel-level perturbations struggle to withstand sequences of common image transformations, posing a risk of defense failure in real-world applications. % To further demonstrate that this vulnerability can be exploited at low cost, we introduced Transformation-Induced Purification via Region-wise Super-Resolution (TIP-RSR), a simple training-free purification framework that combines sequential image transformations with off-the-shelf restoration models. % TIP-RSR can efficiently purify protective perturbations while preserving image fidelity and requiring substantially less computation than existing diffusion-based purification methods. % These findings expose a practical vulnerability in current proactive portrait defenses and highlight the need to explicitly account for the effects of sequential real-world transformations when designing future protection mechanisms. \endabstract

Abstract:
Depression is a widespread mental health issue that increasingly affects adolescents globally. Accurate and timely assessment of depression severity is crucial for effective intervention strategies. Although automated depression detection technologies have been widely studied, existing works primarily rely on models derived from single-event datasets. These models fail to account for the variability in individual responses across different contexts and overlook the diverse manifestations of depression among individuals. Furthermore, existing approaches mainly focus on predicting depression scores using scales such as the Beck Depression Inventory (BDI) or the Patient Health Questionnaire-9 (PHQ-9), while often overlooking subjective wellbeing as a complementary dimension of mental health. Additionally, physiological-signal-based studies have predominantly focused on electroencephalogram (EEG) analysis, often disregarding electromyogram (EMG) signals and their interactions with EEG data. To address these limitations, we introduce M3DEP, a multievent, multimodal, and multitask framework for wellbeing-aware mental health assessment based on physiological signals. We develop a dataset comprising 110 participants, capturing four ecological tasks---self-introduction, two text-reading tasks, and walking---together with depression severity, subjective wellbeing, and individual-difference annotations. M3DEP incorporates Big Five trait classification as individual-difference supervision and jointly models frontal EEG and upper-limb EMG as complementary sources of physiological information at the behavioral level. Under a strict subject-independent split, M3DEP achieves an accuracy of 88.89% in PHQ-9-based depression severity classification and 92.54% in GWB-based wellbeing classification. These results indicate the effectiveness of jointly modeling multiple ecological events, complementary physiological modalities, and individual-difference supervision for the concurrent assessment of depression severity and subjective wellbeing. The dataset and accompanying implementation will be made publicly available.

Abstract:
Recent advances in object-centric representation learning have shown that slot attention-based methods can decompose visual scenes into object-centric slot representations without supervision. However, existing approaches typically process foreground and background regions within a shared slot competition space, often resulting in background interference and suboptimal instance discovery performance on real-world data. To address this challenge, we propose Foreground-Aware Slot Attention (FASA), a structured object-centric learning framework. It structures slot competition in a coarse-to-fine manner, starting from foreground-background decomposition and proceeding to fine-grained binding of foreground objects. Specifically, FASA first performs scene-level foreground-background separation through dual-slot competition with clustering-based initialization, which provides a structural prior for slot specialization. We then introduce a masked slot attention mechanism, where background modeling is isolated to a dedicated slot while the remaining slots compete to represent individual foreground objects. To further alleviate fragmented object binding, we introduce pseudo-maskbased structural consistency to regularize slot assignment toward object-complete grouping. Extensive experiments on both synthetic and real-world datasets demonstrate that FASA improves instance-level object discovery and yields robust object-coherent representations for downstream tasks. The code will be made publicly available.

Abstract:
Immersive video streaming typically follows a viewport-aware pipeline that predicts the users future viewport and allocates bitrate accordingly. In 6-DoF viewing, however, this pipeline becomes far more fragile: coupled translational and rotational motion makes viewport prediction much harder, and even small prediction errors can misguide adaptive bitrate (ABR) decisions, causing bandwidth waste and visible quality degradation. As a result, 6-DoF interaction uncertainty becomes a key bottleneck to stable quality of experience (QoE). In this paper, we present Reparo, an enhancement framework for object-centric, viewport-aware adaptive volumetric video streaming. Rather than proposing another standalone predictor or ABR algorithm, Reparo is designed as a model-agnostic and algorithm-agnostic plug-in that strengthens the two most vulnerable stages of existing viewport-aware pipelines. Its viewport prediction enhancement (VPE) module suppresses misleading rotational cues when the attended object largely fills the viewport and adaptively selects the most informative history window to improve viewport hit rate, while its bitrate decision enhancement (BDE) module estimates prediction reliability and exposes this uncertainty to ABR algorithms, turning bitrate adaptation from prediction-blind to uncertainty-aware. Across diverse network conditions and user interactions, Reparo consistently improves existing state-of-the-art solutions, helping them deliver at least 61.53% higher average QoE.

Abstract:
Handwritten mathematical expression recognition (HMER) focuses on automatically generating LaTeX sequences from provided images. Currently, attention-based encoder-decoder models are extensively adopted for this task. However, due to the ever-changing nature of handwriting styles, existing decoding methods, which merely learn character features from formulaic images, struggle to accurately classify similar symbols, even misclassifying the same symbol into different categories. Attention drift during character recognition further exacerbates visual confusion between characters. To tackle these problems, we propose DeMER, a unified framework integrating feature denoising and attention calibration for effective mathematical symbol representation learning. In particular, we first design an Implicit Variational Prototype Module (IVPM), which effectively decouples of prototype and style by modeling character features as Gaussian probability distributions via variational inference. Variational sampling and prototype alignment mechanisms force noisy features to map onto a standardized prototype space. Second, we introduce the Soft Redundancy Aware Coverage (Soft-RAC) attention refinement module for more precise attention localization. In addition, IVPM can be seamlessly integrated into existing RNN-based models for HMER and consistently deliver performance improvements. Experimental results on widely-used benchmarks demonstrate that our method significantly outperforms existing state-of-the-art approaches in both recognition accuracy and structural robustness.

Abstract:
Class incremental medical image segmentation (CIMIS) aims to continuously learn new anatomical classes while preserving previously acquired knowledge. However, existing methods mitigate background shift by aggregating the background channel's probabilities with those of other unlabeled channels during knowledge distillation or ground-truth supervision, which may lead to semantic confusion and impair the segmentation performance of the merged classes. Moreover, these approaches overlook the conflict that arises when pixels labeled as new classes are misclassified as old classes by the previous model, potentially degrading performance on new categories. To alleviate the above issues, we propose Region-Wise Supervision Deconfusion (RWSD) and Prototype-Guided Feature Rectification (PGFR) for CIMIS in this paper. Specifically, RWSD selectively amplifies the responses of target channels while suppressing interference from competing channels, enhancing segmentation performance. PGFR, on the other hand, rectifies feature representations of conflicting pixels using prototype guidance, thereby strengthening the learning of new classes. Extensive experiments on two multi-organ medical image segmentation benchmarks demonstrate that our method consistently outperforms existing state-of-the-art class incremental semantic segmentation approaches.

Abstract:
Category-level object pose estimation from RGB-D images remains challenging due to large intra-class shape variations and incomplete observations in cluttered scenes. Existing keypoint-based methods improve robustness by learning compact and representative object structures from multimodal RGB-D features. However, their structural modeling is still limited by rotation-sensitive geometric representations and a single fixed keypoint ordering. To address this issue, we propose REF-Pose, a novel keypoint-based framework that constructs rotation-consistent multimodal point-wise representations to generate structurally informative keypoints for pose estimation. To reduce the structural bias caused by a single fixed keypoint ordering, REF-Pose further introduces spatially factorized structural modeling, where keypoint features are serialized under multiple complementary spatial orderings, and a randomly selected ordering is modeled at each layer by a hybrid depthwise convolution and bidirectional Mamba architecture. Extensive experiments on CAMERA25, REAL275, and HouseCat6D demonstrate that REF-Pose achieves state-of-the-art performance with notable gains under occlusion and rotation.

Abstract:
Satellite video tracking remains highly challenging due to tiny targets, weak visual evidence, cluttered large-scale backgrounds, and frequent interference from similar objects. Recent sequence-generation trackers offer a concise autoregressive framework by formulating target localization as sequential state prediction. However, their performance still depends heavily on reliable online target representation during inference, making them vulnerable to drift under noisy observations. To address this issue, we propose SMTrack, a satellite video tracker with a subspace memory based on low-rank manifold modeling for more reliable online adaptation. Rather than directly storing raw historical template features, SMTrack models target appearance history as a compact manifold-structured subspace, which preserves stable appearance characteristics while suppressing transient corruption and noise. This memory is maintained in closed loop. An anchor-guided read uses the fixed first-frame template as a reliable query anchor to retrieve historical cues and enhance the dynamic template, while a contamination-aware write screens each observation before admission, suppressing unreliable evidence at the entrance rather than filtering it downstream. Extensive experiments and comparisons on SV248S, SatSOT, and VISO show that our tracker significantly outperforms previous state-of-the-art trackers.

Abstract:
Adaptive video streaming has seen remarkable advancements in recent years, with adaptive bitrate (ABR) algorithms at its core. Nowadays, the learning-based ABR algorithms have become the mainstream choice, owing to their ability to adapt to diverse streaming environments through training. However, our measurement of state-of-the-art ABR algorithms reveals a critical issue: catastrophic forgetting the key environmental features encountered early or infrequently during training are often forgotten or overlooked, leading to significantly degraded performance at runtime. To the best of our knowledge, none of the existing ABR algorithms is capable of overcoming the catastrophic forgetting due to their conventional training paradigms. To tackle this issue, we propose ANOLE, a novel learning-based ABR framework. At its core is a new training method, Multi-Specialist Guidance, where multiple specialist ABR models collaboratively guide the training of a generalist ABR model that is then deployed at runtime. This ensures all the key features are simultaneously learned by the generalist without forgetting. Evaluations show that ANOLE delivers a 10.5% ~ 20.7% improvement in QoE over the state-of-the-art ABR algorithms. Moreover, it exhibits strong robustness and scalability across heterogeneous network conditions, evolving user preferences, and streaming application diversity, while remaining fully compatible with the existing streaming platforms. We have open-sourced ANOLE on GitHub.

Abstract:
Medical image fusion aims to integrate complementary information from diverse imaging modalities to support clinical diagnosis. Existing methods typically apply uniform fusion rules globally, lacking a deep understanding of diagnostic intents and pathological structures. To address these limitations, we propose MIND, a Multimodal Intent-Driven Network via Diffusion Transformers (DiTs) for medical image fusion. Specifically, we utilize BioMedGPT to generate intent-driven fusion texts from source images, guiding the fusion process with pathology-aware diagnostic intents. To combat the loss of 2D spatial continuity caused by 1D sequence flattening in DiTs, we design a Multi-scale Latent Adapter. This module explicitly extracts source image features before serialization, injecting them into the network via strict dimensional alignment to effectively supplement image features. To resolve the semantic shift caused by decoupling image outputs from diagnostic intents, we design a medical semantic consistency loss. This loss ensures deep semantic locking between fused images and fusion texts while maintaining the stability of the underlying physical manifold reconstruction. Comprehensive experiments on the Harvard, BraTS, and GFP datasets reveal that MIND delivers superior fusion quality, significantly improves downstream brain tumor segmentation accuracy, and enables flexible interactive fusion, holding significant promise for intent-driven intelligent clinical decision support systems.

Abstract:
Unsupervised multimodal intent discovery aims to uncover latent intents from unlabeled multimodal dialogues, but remains challenging due to the lack of explicit semantic supervision. Existing methods often provide limited interpretability, as their refinement mainly relies on geometric similarity rather than high-level semantic guidance. To address these limitations, we propose MCSP, a fully unsupervised method that introduces semantic refinement based on concepts into multimodal intent discovery. To obtain reliable semantic evidence for intent discovery, we identify high-quality representative samples for each cluster and use them to support MLLM-guided contrastive reasoning against neighboring clusters, which produces interpretable high-level semantic concepts. Building on these concepts, we perform semantic propagation over a semantically weighted graph to align conceptual information with local structural consistency and generate reliable pseudo-labels for representation refinement. Extensive experiments on three challenging multimodal intent datasets show that MCSP consistently outperforms state-of-the-art methods while producing interpretable clusters grounded in semantic concepts.

Abstract:
Multimodal Emotion Recognition (MER) in conversations requires a delicate balance between capturing long-range temporal dynamics and immediate paralinguistic cues. While Large Language Models (LLMs) have set new benchmarks through generative instruction tuning, they are fundamentally bottlenecked by the "Context-Attention Trade-off," where extensive dialogue histories induce quadratic token redundancy and attention dilution, leading to inference instability and prohibitive latency. In this paper, we propose TC-LLM, a novel framework that reformulates multimodal ERC as a high-efficiency discriminative task to bypass the generative bottleneck. We introduce Hierarchical Textual History Compression (HTHC), which distills the entire conversational timeline into a global semantic skeleton---serving as foundational background knowledge---while explicitly preserving uncompressed local context to sharpen the model's regional awareness of current interactions. To bridge the modality gap, we design an Asymmetric Hybrid Trigger (AHT) mechanism that anchors textual causality via local context and captures acoustic inducements through a residual attention module. By extracting internal hidden states instead of conducting autoregressive decoding, TC-LLM achieves deterministic emotion classification and a stable 7.4× inference acceleration. Extensive experiments on IEMOCAP and MELD datasets demonstrate that TC-LLM achieves state-of-the-art performance, proving that "less is more" in multimodal context modeling.

Abstract:
Open-ended video game glitch detection aims to identify glitches in gameplay videos, describe them in natural language, and localize when they occur. Unlike conventional game glitch understanding tasks which have largely been framed as image-level recognition or closed-form question answering, this task requires reasoning about game-specific dynamics such as mechanics, physics, rendering, animation, and expected state transitions directly over continuous gameplay videos and distinguishing true glitches from unusual but valid in-game events. To support this task, we introduce VideoGlitchBench, the first benchmark for open-ended video game glitch detection with temporal localization. VideoGlitchBench contains 5,238 gameplay videos from 120 games, each annotated with detailed glitch descriptions and precise temporal spans, enabling unified evaluation of semantic understanding and temporal grounding. We further propose GliDe, an agentic framework with three key components: a game-aware contextual memory for informed reasoning, a debate-based reflector for multi-perspective glitch detection and verification, and an event-level grounding module that recovers complete glitch intervals from fragmented temporal evidence. We also design a task-specific evaluation protocol that jointly measures semantic fidelity and temporal accuracy. Experiments show that this task remains highly challenging for current multimodal models, while GliDe achieves substantially stronger performance than corresponding vanilla model baselines. Code and dataset are available at GitHub and Hugging Face.

Abstract:
Audio-Visual Speaker Detection (AVSD) hinges on modeling both individual temporal continuity and inter-personal social context. Existing coupled architectures struggle to reconcile these tasks in shared representation spaces due to conflicting inductive biases: temporal modeling favors low-frequency smoothness, while inter-personal interaction requires high-frequency discriminability. We propose D^2Stream, a decoupled dual-stream framework that explicitly isolates these functionalities into parallel, task-specific branches. Specifically, the Intra-speaker Temporal Continuity (ITC) stream captures longitudinal stability, whereas the Inter-personal Social Relation (ISR) stream models transversal social cues. Quantitative gradient analysis reveals an evolutionary divergence in update directions, stabilizing at 86.1^\circ, which confirms the inherent task conflict and the effectiveness of our structural decoupling. D^2Stream breaks the long-standing performance plateau, achieving a state-of-the-art 95.6% mAP on AVA-ActiveSpeaker and superior generalization on Columbia ASD, all within a lightweight and efficient design.

Abstract:
Open World Object Detection (OWOD) built on multimodal foundation models often suffers from semantic ambiguity caused by unidirectional text-to-vision matching, while rigid outlier penalties may over-suppress unknown objects near known-class decision boundaries. We propose CODE (Cross-Modal CalibratiOn and Dynamic SupprEssion), a unified inference-time framework with three complementary components. Cross-Modal Joint Confidence Calibration injects global visual prototypes to calibrate text-driven known-class predictions. Uncertainty-Guided Universal Objectness Enhancement measures classification hesitation from local visual responses to strengthen potential unknown objects. Dynamic Outlier Suppression via Confidence Margin replaces rigid suppression with a margin-aware adjustment that preserves ambiguous out-of-distribution instances. Experiments on the Real-World Detection benchmark demonstrate that, with the OWL-ViT L/14 backbone, CODE achieves 21.7 U-mAP and 40.8 K-mAP in Task 1, surpassing the previous state of the art by 2.6 and 2.3 points, respectively.

Abstract:
Deploying high-resolution tiny-object perception on edge platforms requires not only accurate localization, but also selecting a small set of informative patches under compute, transport, and latency constraints. We study budgeted tiny-object selection, where a frontend ranks patch centers from a lightweight proxy and a downstream detector processes only the selected regions. DenseScout is a 1.01M-parameter deployment-oriented dense-response selector that removes detector-style box regression and directly optimizes ranked patch-center prioritization. Its contribution lies in the task-specific selector formulation, the alignment among output representation, supervision, and decoding, and its joint design with transport-aware execution and QoS-oriented evaluation. Under unified protocols on VisDrone and DOTA, DenseScout provides stronger low-budget recall than the evaluated detector-derived selectors; controlled fixed-K inspection experiments further demonstrate advantages over selection-style proxy baselines. Cross-platform profiling on Jetson Orin NX and RK3588 shows that deployable utility depends jointly on selector quality, memory movement, and heterogeneous runtime realization. These results support treating edge tiny-object perception as a selection-and-deployment co-design problem rather than evaluating model accuracy and runtime in isolation.

Abstract:
Large Language Models (LLMs) have demonstrated strong capabilities in natural language understanding and code generation, yet remain limited in their ability to manipulate multimodal data. In this work, we investigate an orthogonal direction that does not rely on model scaling or additional training: \emphdesigning data formats that are inherently aligned with LLM processing. Through controlled case studies spanning image, audio, and 3D modalities, we show that the choice of data format alone can lead to performance differences of up to 8× on downstream tasks, without any modification to the underlying model. Our experiments reveal a consistent task--format interaction pattern: natural language representations are more effective for information extraction tasks (achieving query accuracy up to 0.98), whereas programmatic representations (e.g., SVG, ABC notation, and OpenSCAD) are better suited for structural manipulation tasks (achieving edit correctness up to 1.00). We further analyze LLM reasoning behaviors across formats and find that programmatic formats facilitate \emphedit locality, enabling precise modifications through localized changes, while natural language representations implicitly encode higher-level abstractions that reduce the burden of interpreting low-level structure. Additionally, we show that LLMs can autonomously design novel data formats tailored to specific tasks. Overall, our results position \emphformat engineering as a complementary paradigm for improving multimodal LLM capabilities.

Abstract:
Traditional horticultural therapy (HT) delivers proven psychological benefits yet remains inaccessible to hospitalized patients, individuals with mobility impairments, and urban residents due to physical barriers and spatial constraints. This paper presents BreathForest, a proof-of-concept breath-responsive VR system demonstrating the technical feasibility of translating HT's therapeutic mechanisms into passive virtual environments. We introduce \emphlight interaction, a design paradigm where unconscious breathing drives environmental responses without controllers or learned techniques. Through CO_2-based sensing (<100ms latency, >90% accuracy), the system translates respiratory patterns into synchronized multimodal feedback across three progressive forest scenarios. Evaluation with 12 participants reveals strong system responsiveness (M=6.35/7), enhanced embodiment (M=6.12/7), and positive therapeutic impact (M=5.75/7). Our contributions include: (1) establishing light interaction as an accessible therapeutic XR paradigm, (2) providing a replicable translation framework mapping physical HT mechanisms to breath-responsive interactions, and (3) validating CO_2 sensing technical specifications. By preserving HT's experiential elements (agency, responsibility, achievement) while eliminating participation barriers, this work establishes foundations for future clinical validation with excluded populations.

Abstract:
As generative multimedia evolves from static image synthesis to complex, interleaved visual narratives, a foundational bottleneck has emerged: the judgment crisis. While human perception natu- rally synthesizes the temporal and logical flow of a story, automated evaluation systems remain largely blind to sequential continuity, often failing to distinguish between a coherent narrative and a semantically shuffled or contradictory sequence. This work iden- tifies a critical structural gap in current multimodal evaluation paradigms, arguing that the reliance on Large Vision-Language Models (LVLMs) as judges is fundamentally limited by architectural biases. Our analysis reveals a profound performance dichotomy: while models may appear competent in isolated pointwise scoring, they suffer a catastrophic collapse when required to perform pair- wise discrimination of temporal order. We demonstrate that this is not merely a data-scarcity issue but a structural one. Through a series of diagnostic probes, we uncover systematic positional asym- metries, specifically primacy and recency effects, where a models judgment of a story is significantly influenced by the placement of a frame, often more than by its semantic consistency. These bi- ases, potentially rooted in causal masking and rotary embeddings, suggest that current transformer-based judges are inherently ill- equipped for long-form visual reasoning. By exposing these blind spots, we challenge the multimedia community to move beyond snapshot-centric metrics and instead pioneer Temporally-Aware Evaluation paradigms that treat visual sequences as unified logical structures rather than unordered collections of frames.

Abstract:
Scene-dependent video anomaly detection (SDVAD) is a recently emerging and valuable topic. Existing methods typically rely on scene-conditioned generative learning, which implicitly embeds scene context without explicitly considering the intrinsic video scene labels and the connection between foreground events and their corresponding scenes. To this end, we propose a novel discriminative-contrastive learning framework, which can explicitly exploit scene labels and bridge foreground events and scenes for effective SDVAD. First, we deploy a discriminative learning process by directly taking scene class labels from different surveillance cameras as learning targets. It trains deep neural networks (DNNs) to classify events from different scenes, and anomalies are determined based on scene-aware classification probabilities in inference. Second, to fully establish the connection between events and scenes, we develop a contrastive learning strategy, which pulls events closer to their corresponding scenes while pushing them away from other scenes. During inference, the distance between events and scenes is utilized to perform anomaly scoring for further enhancing performance. Finally, we design a joint discriminative-contrastive DNN, where the discriminative module embeds scene modulation signals into multiple layers with learnable weights to properly incorporate scene information, and the contrastive module utilizes a memory bank to augment contrastive samples, thus strengthening the connection between events and scenes. Extensive experiments on commonly-used SDVAD datasets demonstrate that our method achieves state-of-the-art performance with up to 8.5% AUC improvement over existing methods.

Abstract:
Portrait animation methods have achieved substantial visual quality and lip synchronization, but fine-grained manipulation of the eye region still faces a trade-off between input granularity and motion accuracy. Existing methods using emotion labels or coarse text prompts are insufficient for describing subtle ocular dynamics, whereas approaches based on Action Units or driving videos provide higher fidelity at the cost of a heavier input burden. These limitations are still restrictive for beyond-emotion states (e.g., thinking) and drowsiness. In light of the above, we propose CogPortrait, a two-stage framework that generates portrait animations from high-level labels. In the first stage, three Multimodal Large Language Model (MLLM) agents compile high-level labels into facial keypoints: chain-of-thought (CoT) planning and composition agents handle temporal event planning and prototype composition, while a critic enforces semantic-physiological constraints. In the second stage, a DiT-based video generation backbone synthesizes the final animation conditioned on the keypoints, reference portrait, audio, and text prompt, enhanced by a dynamic classifier-free guidance strategy with eye-region-aware reweighting and KTO-based refinement for boundary cases. We further introduce the EMH benchmark covering diverse emotions and beyond-emotion categories with two AU-level metrics for evaluating fine-grained eye-region and head-motion control. Extensive experiments on HDTF and the EMH benchmark demonstrate that CogPortrait achieves more precise eye-region control than existing methods while maintaining superior visual quality and identity consistency.

Abstract:
Weakly Supervised Video Anomaly Detection (WS-VAD) increasingly relies on audio-visual integration to resolve visual ambiguities. However, real-world surveillance suffers from inherent temporal asynchrony and noise-induced modality corruption. Existing methods often fail to explicitly measure modality reliability, leading to a blind fusion of compromised data that inevitably degrades detection performance. To address these intertwined challenges, we propose the Deep Evidential Audio-Visual Learning (DEAL) framework. First, a State Space Model-based Feature Extraction and Alignment (FEA) module employs a novel cross-modal parameter exchange strategy to dynamically resolve temporal misalignment. Second, a Bi-Branch Deep-Evidential Fuser (BDF) utilizes Evidential Neural Networks to quantify epistemic uncertainty, enabling adaptive feature re-weighting. Finally, a Joint-Modal Mutual Learning (JML) strategy is integrated to optimize feature manifolds and calibrate model overconfidence. Extensive experiments on the challenging XD-Violence dataset demonstrate that DEAL effectively suppresses misleading noise and achieves robust semantic alignment without explicit timestamp supervision. By systematically eradicating blind fusion through explicit uncertainty modeling, DEAL significantly outperforms state-of-the-art methods, providing a highly robust and unified paradigm for real-world multimodal anomaly detection.

Abstract:
Audio-driven avatar generation requires realistic lip-sync, expressive motion, and real-time streaming. Recent work achieves the latter via self-forcing with Distribution Matching Distillation (DMD), but this paradigm suffers from a critical failure that has not been systematically characterized: dynamic collapse, where the student model converges to a near-static optimum with high perceptual quality but severely suppressed temporal dynamics. We trace this to two causes: the reverse KL objective in DMD, which biases toward low-motion modes, and unanchored self-conditioning, which creates a feedback loop that amplifies collapse. This is especially harmful for avatars, where even subtle motion loss breaks lip-sync and expression. To address this, we propose DynaForcing, a training framework with three complementary strategies applied at different levels. Specifically, Hybrid Forcing anchors rollouts to ground-truth dynamics at the data level to break the feedback loop. Dynamics-Aware Reward Regularization introduces explicit motion rewards via the RL interpretation of DMD to counteract the reverse KL bias at the loss level. Reference Perturbation perturbs reference images to decouple identity from static details, forcing the model to rely on audio for motion at the conditioning level. We further introduce computation graph pruning and gradient replay, reducing the GPU footprint of self-forcing by over an order of magnitude. Experiments show that DynaForcing recovers dynamics to teacher-comparable levels (Dyn-Deg: 0.31?0.73, Sync-C: 7.03?7.68) while improving visual quality, resolving the quality-dynamics trade-off throughout training without early stopping.

Abstract:
\beginabstract Precise Event Spotting (PES) requires distinguishing visually similar yet semantically distinct adjacent frames, making it fundamentally different from image classification and coarse action recognition. Although self-distillation methods such as DINO have shown strong representation learning ability in images, we find that directly applying them to PES is ineffective: without supervised guidance, subtle but crucial motion cues are often suppressed as noise, leading to representations that are insensitive to precise event boundaries. To address this, we propose Temporal Feature Distillation, a semi-supervised objective that aligns temporally informative backbone features, rather than projection-head outputs, to preserve motion-sensitive and boundary-aware cues for frame-level localization. A supervised warm-up with a ramp-up schedule further stabilizes training by ensuring that meaningful event cues are learned before unlabeled distillation begins. We also introduce Transformer Gate Shift, a multi-scale gated shifting module that injects motion-aware temporal information into Vision Transformers. Experiments on four fine-grained sports benchmarks show consistent improvements over fully supervised and semi-supervised baselines. Under 10% supervision on FSPerf, our method improves mAP by 4.54 points over the strongest competing approach, and with only 80% labeled data, it matches or surpasses the fully supervised 100% baseline on two of the four datasets. \endabstract

Abstract:
Music-grounded mashup video creation is a challenging form of video non-linear editing, where a system must compose a coherent timeline from large collections of source videos while aligning with music rhythm, user intent, story completeness, and long-range structural constraints. Existing approaches typically rely on fixed pipelines or simplified retrieval-and-concatenation paradigms, limiting their ability to adapt to diverse prompts and heterogeneous source materials. In this paper, we present GLANCE , a global-local coordination multi-agent framework for music-grounded nonlinear video editing. GLANCE adopts a bi-loop architecture for better editing practice: an outer loop performs long-horizon planning and task-graph construction, and an inner loop adopts the ''Observe-Think-Act-Verify'' flow for segment-wise editing tasks and their refinements. To address the cross-segment and global conflicts emerging after subtimeline composition, we introduce a dedicated global-local coordination mechanism with both preventive and corrective components, which includes a novelly designed context controller, conflict region decomposition module and a bottom-up dynamic negotiation mechanism. To support rigorous evaluation, we construct MVEBench , a new benchmark that factorizes editing difficulty along task type, prompt specificity, and music length, and propose an agent-as-a-judge evaluation framework for scalable multi-dimensional assessment. Experimental results show that GLANCE consistently outperforms prior research baselines and open-source product baselines under the same backbone models. With GPT-4o-mini as the backbone, GLANCE improves over the strongest baseline by 33.2% and 15.6% on two task settings, respectively, with particularly strong gains on more challenging long-horizon subsets. Human evaluation further confirms the quality of the generated videos and validates the effectiveness of the proposed evaluation framework.

Abstract:
In this paper, we present Swap-JEPA, a novel joint embedding predictive architecture (JEPA) for zero-shot object swapping in live commerce videos. In contrast to conventional approaches that rely on labor-intensive, meticulously paired video data or suffer from representation collapse when fine-tuning diffusion models, Swap-JEPA leverages a pair of expressive latent space aligners to modulate frozen pre-trained diffusion models. Specifically, given multiple reference images of a target object and a source video: (i) We introduce a holistic latent correspondence aligner (HLCA), which models semantic-structural correspondences among the target object, the source object, and the surrounding context. Crucially, HLCA predicts latent representations consistent with unseen viewpoints, thereby enhancing the temporal and cross-view consistency of the target object throughout the generated sequence. (ii) A discrepancy-guided intra-frame aligner (DIA) is designed to bridge the feature distribution gap between the source and target objects within individual frames, facilitating seamless visual integration. (iii) Furthermore, adhering to the JEPA philosophy of decoupling representation learning from generative adaptation, we devise a progressive synthetic-to-real optimization paradigm. HLCA and DIA are first trained on identity-disjoint 3D-rendered pairs via latent matching objectives to prevent representation collapse into identity copying shortcuts. Subsequently, these aligners are frozen, and only contextual adaptation parameters are optimized on real commerce footage within the same latent space. Consequently, Swap-JEPA exhibits strong generalization to diverse, unseen objects, yielding high-fidelity, temporally coherent videos that significantly outperform state-of-the-art methods.

Abstract:
Understanding and reasoning over 3D scenes from point clouds is essential for a broad range of AI applications. Recent 3D large language models have shown promising results, yet they still face two fundamental limitations: (i) insufficient capability for grounded multi-step reasoning in complex 3D scenes, and (ii) a text-centric inference paradigm that lacks effective interaction between textual reasoning and visual evidence. As a result, these models often fail to verify intermediate inferences, disambiguate similar object instances, and avoid hallucinations. To address these limitations, we propose MMCoT^3, the first 3D--Text Interleaved Chain-of-Thought (CoT) framework for 3D scene understanding. The core idea is to explicitly interleave textual reasoning with task-relevant 3D evidence, enabling the model to consult scene observations throughout the reasoning process. In addition, since a textual mention in 3D scenes may correspond to multiple candidate objects, naive object--mention alignment can inject incorrect evidence and mislead subsequent reasoning. To mitigate this multi-instance ambiguity, we design an ambiguity-aware evidence aggregation mechanism that uses query-guided cross-attention to fuse candidate object tokens and inject both evidence and ambiguity cues into the reasoning process. To better equip the model with multi-step reasoning capability, we further develop PointCoT, a large-scale 3D-specific CoT instruction tuning dataset with approximately 176.2K instances spanning captioning, question answering, dialogue, and planning tasks. Extensive experiments on six benchmarks show that our method consistently improves performance and achieves state-of-the-art results, validating the effectiveness of our proposed ambiguity-aware 3D--text interleaved reasoning for 3D scene understanding.

Abstract:
Long-term viewport prediction is a key technique for improving the users' quality of experience of 360 video. However, existing methods suffer from two major limitations. First, they overlook the mismatch between saliency-based visual priors and user attention mechanisms, thus deriving saliency maps that align poorly with real user viewport movements. Meanwhile, they lack effective modeling of the correlation between head and eye behaviors, and fail to produce accurate viewport predictions. To address these issues, this paper proposes BAT-VP, a novel behavior-modulated long-term viewport prediction system for 360 videos. First, we propose a visual graph encoding module to extract prototypes of users' historical visual preferences. Then, we introduce a behavior-modulated visual attention module that dynamically generates user-level saliency maps. Meanwhile, we design a head & eye soft alignment module to model the collaborative behavior between head and eye-movements. Finally, these multimodal outputs are passed through a novel Temporal & Visual Fusion module that enables adaptive collaborative prediction across time horizons. Through comprehensive evaluation on two datasets, BAT-VP outperforms five representative baselines, improving overall and long-term IoU by at least 4.9% and 7.6%, respectively, while maintaining low computational overhead.

Abstract:
Large vision-language models (LVLMs) such as Qwen-VL series have demonstrated strong capabilities in visual understanding and reasoning, particularly for static images. However, extending them to video reasoning remains computationally expensive, often incurring substantial latency and token usage under conventional Chain-of-Thought (CoT) prompting. In this paper, we present a unified framework for efficient video reasoning that combines Chain of Draft (CoD) prompting with adaptive sampling, event compression, and supervised fine-tuning. CoD encourages the model to produce concise and informative intermediate reasoning steps rather than verbose rationales, making it particularly suitable for temporally complex video understanding. We adapt this paradigm to video reasoning and show that our approach achieves comparable or superior accuracy to CoT while reducing token consumption by up to 78% and inference latency by up to 65%. We evaluate our method on multiple video reasoning benchmarks, including MVBench, EgoSchema, PerceptionTest, and VideoMME, and demonstrate consistent gains across diverse tasks. Overall, our results show that concise draft-style reasoning offers an effective path toward deploying video reasoning systems in resource-constrained and real-time settings.

Abstract:
Prompt tuning has emerged as a parameter-efficient strategy to adapt pre-trained vision-language models (VLMs) to biomedical downstream tasks without fine-tuning the backbone. Unlike natural imaging, biomedical images present a unique conflict where diagnostic semantics are frequently obscured by the ubiquitous style variations across scanner, site or modality, etc. However, existing prompt tuning methods mainly learn dense yet continuous prompts in a unified embedding space lacking explicit subspace structures, and thus fail to decouple diagnostic semantics and domain-dependent styles, resulting in poor generalization to unseen domains. To this end, we propose Decoupled Low-rank Prompt Subspace learning (DeLPS), which is designed to decouple diagnostic semantics and domain-dependent style for adapting VLMs to the biomedical domain. Specifically, we parameterize prompts via a set of structured low-rank basis that is decomposed into orthogonal semantic and style subspaces. The semantic subspace is designated to capture domain-invariant diagnostic cues, while the style subspace confines domain-dependent styles. To further enforce this structural decoupling, we devise an asymmetric optimization strategy termed Semantic-Anchored Prompt Distillation, which anchors the semantic subspace to invariant text representations while isolating domain-dependent shifts within the style subspace, thereby ensuring robust generalization to unseen domains. Extensive experiments on 11 biomedical datasets across 7 imaging modalities demonstrate stronger generalization of our proposed DeLPS.

Abstract:
Large Vision-Language Models (LVLMs) demonstrate remarkable multimodal capabilities by encoding text, images, and videos into unified tokens for joint reasoning. However, the excessive number of visual tokens results in overly long input sequences to the large language model (LLM), increasing its inference complexity and latency. To address this issue, visual token pruning methods have been proposed. However, existing methods evaluate tokens in isolation, ignoring their semantic correlations and leading to severe information redundancy in the retained token subset. In this paper, we reformulate token pruning as the task of constructing an optimal information subset. We propose EntropyPruner, a framework that iteratively selects tokens to maximize the marginal information gain. We introduce a novel non-parametric estimation method derived directly from the visual feature space: a tokens self-information is quantified by the entropy of its feature similarity distribution to measure its significance, while its mutual information with the already selected set is estimated through global representational overlap to quantify its redundancy. By unifying self-information and mutual information within the optimization process, EntropyPruner maintains robust performance even at high compression ratios and achieves new state-of-the-art performance across a wide range of models and benchmarks.

Abstract:
Continual Generalized Category Discovery (CGCD) aims to recognize old categories while progressively discovering new ones from an unlabeled dataset. A key challenge lies in the unreliability of pseudo-labels for unlabeled samples, where these errors are repeatedly reinforced during training, leading to accumulated confirmation bias and degraded performance over sessions. Consequently, we propose a Reliability-Aware Pseudo-label (RAP) framework that models the trustworthiness of pseudo-labeled new candidates and adaptively regulates their contribution during optimization. Pseudo-label reliability is estimated via multiple risk signals, including prediction ambiguity, neighborhood inconsistency, and old-new conflict, enabling the suppression of unreliable supervision while preserving informative signals. Meanwhile, we observe that emphasizing new-category learning under such reweighting can inadvertently disturb old-category decision boundaries. RAP therefore incorporates a boundary preservation mechanism to stabilize old-category evidence under old-new competition. Extensive experiments demonstrate consistent improvements on different CGCD benchmarks, validating the effectiveness of our proposed framework.

Abstract:
Longitudinal chest X-ray report generation requires joint understanding of current and prior studies to accurately describe disease evolution. Existing methods often underuse prior images and reports, rely on implicit temporal fusion, and overlook cross-time anatomical misalignment caused by posture and acquisition differences, leading to unreliable progression descriptions. We propose ProAlign, a framework for anatomically aligned and clinically grounded progression modeling. Its Anatomical Feature Alignment module formulates cross-time correspondence as a regularized Optimal Transport problem, enabling anatomically consistent matching without rigid registration and providing a reliable structural basis for temporal comparison. Built on this, a Text-Grounded Progression Estimation module uses the prior report as a semantic prior to suppress irrelevant visual changes and derive a robust Disease Progression Field (DPF) for progression-aware report generation and clinically faithful decoding. By combining anatomically aligned visual comparison with prior-report-guided semantic filtering, ProAlign better distinguishes true pathological evolution from non-pathological variations and reduces clinically unreliable temporal reasoning. On MIMIC-CXR and M-ABN, ProAlign outperforms prior state-of-the-art methods with CE F1 improvements of 3.5% and 3.2%, respectively.

Abstract:
Multimodal Large Language Models (MLLMs) have significantly advanced embodied AI, yet synchronized bimanual coordination introduces substantial challenges in multi-stream multimodal integration. We introduce ST-BiBench, a comprehensive multi-tier framework for evaluating spatio-temporal multimodal coordination. Our approach centers on Strategic Coordination Planning, assessing high-level cross-modal reasoning over multiple action and perception streams. To investigate spatial arm-assignment errorswhere semantically coherent plans fail to align with spatially grounded visual inputswe incorporate Foundational Spatial Grounding to evaluate workspace awareness and arm-selection logic. Furthermore, Fine-Grained Action Control serves as a stress test of whether MLLMs can map multimodal context and structured robot/object states to 16-dimensional continuous dual-arm actions. Evaluating 30+ MLLMs, we uncover a persistent gap between high-level strategic reasoning and fine-grained physical execution. Results reveal that while stronger MLLMs perform well on logic-driven planning, they still exhibit spatial grounding errors and model-dependent multi-stream interference. ST-BiBench provides a platform for identifying critical bottlenecks in multi-stream multimodal fusion and cross-modal alignment for complex embodied tasks.

Abstract:
In recent years, Diffusion Models (DMs) have achieved remarkable success in style transfer and AI-driven artistic creation. However, they have also exacerbated the risk of unauthorized style imitation of original artworks. Existing protection methods, such as Mist and Glaze, mitigate infringement risks by embedding adversarial perturbations into images to disrupt the fine-tuning process, especially for prevalent techniques like DreamBooth and LoRA. Nevertheless, these denfenses exhibit clear limitations when facing training-free diffusion-based style transfer methods, such as StyleID. To address this challenge, we propose Anti-Training-Free Style Transfer in Diffusion Models(AFSD), an active defense framework tailored for training-free style transfer. Specifically, AFSD introduces adaptive perturbations in high-frequency regions and imposes constraints at both the latent representation and self-attention feature levels, thereby effectively disrupting the propagation of style information while preserving perceptual quality. Extensive experiments demonstrate that AFSD significantly degrades the performance of training-free diffusion-based style transfer methods. Our work provides an effective proactive defense strategy for protecting the copyright of artistic works.

Abstract:
Video Temporal Grounding (VTG) has become a critical challenge due to the rapid growth of video content on digital platforms. Traditional methods face difficulties in adapting to token contextualization and implementing stable masking strategies. To address these challenges, we propose a novel Dynamic Masked Mixture-of-Experts with Sparse-Constrained Reinforcement Learning (DSMoE) framework for video temporal grounding. Specifically, the Context-aware Semantic Attention Fusion (CASF) module dynamically adjusts dummy tokens in cross-attention to suppress irrelevant textvideo interactions. The Multi-view Expert Semantic Extraction (MESE) module employs a mixture-of-experts architecture to capture complementary semantic and structural information. The Gate Control Denoising Decoder (GCDD) module further filters noisy activations through gated hidden-layer sparsification. Moreover, the Adaptive Token Fusion Policy Optimizer (FPO) learns the dummy-token sampling policy using reinforcement learning with an L1 constraint, improving both policy sparsity and training stability. Extensive experiments demonstrate that DSMoE consistently outperforms SOTA methods in video temporal grounding and highlight detection, validating its effectiveness and robustness.

Authors: Bin Chen (College of Computer and Data Science, Fuzhou University), Jia-Li Yin (College of Computer and Data Science, Fuzhou University), Zhaozhe Hu (College of Computer and Data Science, Fuzhou University), Genna Ba (College of Computer and Data Science, Fuzhou University), Wei Lin (College of Mathematics and Computer Science, Fujian University of Technology), Xiaolei Liu (National Interdisciplinary Research Center of Engineering Physics, China Academy of Engineering Physics), Lichao Su (College of Computer and Data Science, Fuzhou University), Yuanlong Yu (College of Computer and Data Science, Fuzhou University), Shouling Ji (College of Computer Science and Technology, Zhejiang University)

Abstract:
Vision-language models (VLMs) have been shown to be vulnerable to adversarial examples (AEs), i.e., images that remain visually similar to benign ones but can induce malicious outputs. The risk is exacerbated in realistic scenarios, where the same image may be queried with diverse prompts and an effective AE should remain effective across such prompt variations. Therefore, improving the cross-prompt transferability of AEs is crucial for assessing the practical security of VLMs. Existing cross-prompt attacks mainly improve transferability by searching for harder prompts in a local neighborhood of the prompt embedding space. Despite their effectiveness, this formulation still models prompt variation as a continuous latent perturbation and therefore does not fully capture explicit prompt changes encountered at inference time. In practice, prompts are more often varied through token-level rewriting or reasoning-process modifications induced by self-correction instructions. Motivated by this gap and recent progress on self-correction in VLMs, we propose the Anchored Dual-Track Self-Correction Attack (ADSA) for improving cross-prompt adversarial transferability. ADSA strengthens optimization-time prompt hardening from two complementary directions: Token-level Adversarial Prompt Search (TAPS), which performs budgeted token-level substitutions to construct harder, intent-related prompt variants for cross-prompt optimization, and Self-Correction-Guided Cross-Prompt Optimization (SCG-CPO), which introduces self-correction as a process-level hard prompt condition during adversarial optimization. Moreover, ADSA adopts an anchored design that keeps prompt-side optimization focused on the rewritten query rather than the auxiliary self-correction template. Extensive experiments on multiple VLMs and tasks show that ADSA consistently outperforms prior methods across diverse target outputs and remains effective under self-correction and in-context learning settings.

Abstract:
Point cloud completion is a fundamental yet still challenging problem in 3D vision. Existing methods primarily integrate 3D and 2D modalities by observing structural cues within the data to achieve more comprehensive, robust, and accurate perception. However, the use of fixed 2D modalities makes it difficult to balance computational efficiency and information sufficiency effectively. Unlike previous approaches, our proposed framework adaptively determines and fuses informative 2D modalities based on the intrinsic geometric characteristics of 3D data, achieving higher consistency and accuracy in structural perception and fine-grained geometric detail completion. Specifically, we employ a difficulty-aware mechanism to adaptively select multi-view self-projected depth maps according to the proxy completion difficulty, thereby significantly enhancing feature representation. Furthermore, we introduce a view-driven refinement strategy that collaboratively exploits shape priors and self-similarity, enabling more effective and precise fine-grained completion of missing point clouds. Extensive experiments are conducted on the ReTooth3D, ToothComp3D, and PCN datasets. The results demonstrate that our method consistently achieves leading performance in terms of both efficiency and geometric completeness across diverse datasets, validating its strong generalization capability and robustness.

Abstract:
Action recognition requires not only identifying which frames are individually informative but also how multiple frames synergize into a sub-action, and how these sub-actions jointly support a coherent action prediction. Existing methods address specific aspects of the issue by either optimizing short-term motion dynamics or long-range sequence modeling, but lack a unified framework that integrates them into a complete reasoning chain. We address this gap by revisiting action recognition through the Volterra-series expansion, which jointly models the first-order contributions of individual temporal evidence and higher-order interactions that capture co-occurrence, dynamics, and causal cues. Building on this view, we present Volterra-inspired Reasoning Framework (VolRF), a unified action recognition pipeline with four stages: (1) frame-level information quantification, (2) local frame-pair interaction, (3) segment-level sub-action aggregation, and (4) long-range ordered sub-action composition. This view yields a modular framework in which each stage can be independently enabled, replaced, and evaluated. We evaluate contributions of each stage via extensive ablation studies. Experiments on three public benchmarks and one challenging derived subset show that VolRF consistently improves over strong baselines. Moreover, we release the framework with interfaces and switches of modules in each stage, which supports different uni-modal/multi-modal implementations and evaluates the effectiveness using unified metrics to foster modular advances in action recognition.

Abstract:
Large-scale vision-language models (VLMs) exhibit remarkable zero-shot capabilities, yet the internal mechanisms driving their anomaly detection (AD) performance remain poorly understood. Current methods predominantly treat VLMs as black-box feature extractors, assuming that anomaly-specific knowledge must be acquired through external adapters or memory banks. In this paper, we challenge this assumption by arguing that anomaly knowledge is intrinsically embedded within pre-trained models but remains latent and under-activated. We hypothesize that this knowledge is concentrated within a sparse subset of anomaly-sensitive neurons. To validate this, we propose latent anomaly knowledge excavation (LAKE), a training-free framework that identifies and elicits these critical neuronal signals using only a minimal set of normal samples. By isolating these sensitive neurons, LAKE constructs a highly compact normality representation that integrates visual structural deviations with cross-modal semantic activations. Extensive experiments on industrial AD benchmarks demonstrate that LAKE achieves state-of-the-art performance while providing intrinsic, neuron-level interpretability. Ultimately, our work advocates for a paradigm shift: redefining anomaly detection as the targeted activation of latent pre-trained knowledge rather than the acquisition of a downstream task. Our codes and models will be publicly released.

Abstract:
Point cloud completion is inherently ill-posed due to severe sparsity and ambiguity in partial observations. Existing multi-view methods alleviate this by incorporating 2D semantics, but often rely on learned attention and fixed fusion, which lack geometric consistency and adaptability. We propose ProjFormer, a cross modal framework that enforces geometry-consistent 2D--3D interaction through explicit projection and adaptive feature routing. A Projective Guided View Attention module aligns 3D points with multi-view features via deterministic projection, enabling efficient and geometrically consistent aggregation. Building on this, a geometry-aware routing network performs point-wise adaptive fusion of structural and observation-driven features for progressive refinement. Experiments show that, under a lightweight design, ProjFormer delivers competitive performance with improved structural completeness.

Abstract:
Point cloud completion aims to recover complete 3D shapes from partial observations. Recent methods introduce single-view images as auxiliary guidance, yet visual features are bound to the camera viewpoint and carry little information about occluded regions. Text generated by a Vision-Language Model can complement images with view-invariant structural priors. However, treating such text as just another feature source and fusing it directly with image features discards its compositional structure and even degrades quality. We argue that the true value of text lies in decomposing an object into semantic parts, and propose CRCNet, a multimodal framework built on a compose-and-route paradigm. Each part phrase selects relevant visual features via cross-attention to form a concept prototype, and each coarse point retrieves prototypes by its affinity to the part phrases for targeted refinement. We further introduce a Geometry-Gated Semantic Dispatch (GGSD) module: while routing decides which prototypes each point attends to, a proximity-aware gate controls how much guidance it receives, based on its geometric displacement from the observed surface. Since each point is explicitly routed through part-level phrases, our framework is highly interpretable. Experiments on ShapeNet-ViPC show that CRCNet achieves state-of-the-art performance and generalizes strongly to unseen categories and to real-world LiDAR scans.

Abstract:
Electroencephalography (EEG), as a non-invasive physiological signal, plays an increasingly important role in human-centered multimedia systems, supporting applications ranging from emotion recognition and cognitive load assessment to motor-imagery-based control. However, due to the substantial differences in temporal scales, neural patterns, and label semantics across EEG tasks, it remains challenging to unify heterogeneous downstream tasks within a single model, while reconciling shared and task-specific EEG features. To this end, we propose UniNeuro, a unified EEG foundation framework that seamlessly integrates heterogeneous-task pretraining and downstream adaptation into a single discriminative learning pipeline. Within the backbone, we introduce a task-conditional Mixture-of-Experts (MoE) with gradient reversal, enforcing shared experts to capture cross-task features while dynamically routed experts focus on task-specific features, naturally decoupling the two types of representations. At the downstream adaptation stage, we propose Hierarchical Prototype-Guided Contrastive Learning (HPGCL) to explicitly structure the latent space at both task and label levels, enhancing task-level clustering and within-task class separability. Pretrained on over 10,000 hours of EEG data and evaluated on six diverse tasks, UniNeuro shows strong unified multi-task performance and narrows the gap with single-task fine-tuning, offering a scalable brain-modality encoder for next-generation interactive multimedia systems.

Abstract:
Non-ideal measurement computed tomography (NICT), which lowers radiation at the cost of image quality, is expanding the clinical use of CT. Although unified models have shown promise in NICT enhancement, most methods require paired data, which is an impractical demand due to inevitable organ motion. Unsupervised approaches attempt to overcome this limitation, but their assumption of homogeneous noise neglects the variability of scanning protocols, leading to poor generalization and potential model collapse. We further observe that distinct scanning protocols, which correspond to different physical imaging processes, produce discrete sub-manifolds in the feature space, contradicting these assumptions and limiting their effectiveness. To address this, we propose an Uncertainty-Guided Manifold Smoothing (UMS) framework to bridge the gaps between sub-manifolds. A classifier in UMS identifies sub-manifolds and predicts uncertainty scores, which guide the generation of diverse samples across the entire manifold. By leveraging the classifiers capability, UMS effectively fills the gaps between discrete sub-manifolds and promotes a continuous and dense feature space. Due to the complexity of the global manifold, it's hard to directly model it. Therefore, we propose to dynamically incorporate the global- and sub-manifold-specific features. Specifically, we design a global- and sub-manifold-driven architecture guided by the classifier, which enables dynamic adaptation to subdomain variations. This dynamic mechanism improves the networks capacity to capture both shared and domain-specific features, thereby improving reconstruction performance. Extensive experiments on public datasets are conducted to validate the effectiveness of our method across different generation paradigms.

Abstract:
Multimodal-attributed graphs (MAGs) are a fundamental data structure for multimodal graph learning (MGL), enabling both graph-centric and modality-centric tasks. However, our empirical analysis reveals inherent topology quality limitations in real-world MAGs, including noisy interactions, missing connections, and task-agnostic relational structures. A single graph derived from generic relationships is therefore unlikely to be universally optimal for diverse downstream tasks. To address this challenge, we propose Task-aware Modality and Topology co-Evolution (TMTE), a novel MGL framework that jointly and iteratively optimizes graph topology and multimodal representations toward the target task. TMTE is motivated by the bidirectional coupling between modality and topology: multimodal attributes induce relational structures, while graph topology shapes modality representations. Concretely, TMTE casts topology evolution as multi-perspective metric learning over modality embeddings with an anchor-based approximation, and formulates modality evolution as smoothness-regularized fusion with cross-modal alignment, yielding a closed-loop task-aware co-evolution process. Extensive experiments on 9 MAG datasets and 1 non-graph multimodal dataset across 6 graph-centric and modality-centric tasks show that TMTE consistently achieves state-of-the-art performance.

Abstract:
Learning dynamic object representations from monocular videos is a crucial challenge in computer vision and graphics, demanding accurate reconstruction and realistic rendering. The recent 3D Gaussian Splatting method offers a powerful representation for dynamic rendering. However, the discrete and unstructured nature of Gaussians makes accurate surface reconstruction difficult. Some methods use squashed 3D Gaussians to improve geometric accuracy, sacrificing rendering fidelity, while others adopt dual-model architectures with high computational cost. To address these limitations, we propose LDSG, which learns Decoupled Structured Gaussians to jointly improve rendering and reconstruction. Specifically, we introduce a dual canonical Gaussian representation to decouple the conflict between appearance and geometry, where the covariance of the geometric Gaussians is derived by inheriting structural priors from the rendering Gaussians. We then propose an Anchor-GS guided hierarchical deformation to mitigate interference between rendering and reconstruction in single-level deformation fields. Subsequently, a spatiotemporally-aware anchor densification scheme adaptively enhances model detail representation. Moreover, we design a virtual-view depth consistency regularization to improve geometric accuracy, which enforces cross-view geometric coherence using synthesized virtual viewpoints. These components form a unified and efficient framework that avoids maintaining separate rendering and reconstruction models while preserving their complementary strengths. Extensive experiments demonstrate that our method achieves state-of-the-art performance in dynamic rendering and reconstruction while maintaining high efficiency.

Abstract:
Radar-camera fusion combines visual semantics with reliable radial depth and Doppler velocity for 3D object detection in autonomous driving. However, sparse and noisy radar measurements, cross-frame misalignment of dynamic objects, and cross-modal semantic discrepancies and reliability variations hinder robust fusion. We propose SPRFusion, a query-based radar-camera fusion framework guided by semantic and physical priors, comprising three complementary components. The Cross-Modal Temporal BEV Alignment (CMTA) module uses radar Doppler cues to estimate a dense displacement field and warp historical camera BEV features, improving temporal consistency for dynamic objects. The Difficulty-Aware Circular Distribution (DACD) strategy initializes object queries from a distance-dependent difficulty field, allocating more queries to regions affected by occlusion, image degradation, and radar sparsity. The Dual-Robust Cross-Modal Decoder calibrates radar-guided query representations using a static LLM-derived physical-semantic codebook and adaptively suppresses features from unreliable modalities through uncertainty-aware gated fusion. Extensive experiments on nuScenes and View-of-Delft (VoD) demonstrate the effectiveness and robustness of SPRFusion. With a ResNet-50 backbone, SPRFusion achieves 62.4% NDS and 55.4% mAP on the nuScenes validation set. With a V2-99 backbone, it achieves 68.2% NDS and 62.9% mAP on the official test set without future frames, test-time augmentation, or model ensembling, demonstrating its effectiveness as a robust multimodal 3D detector.

Abstract:
Multimodal representation has been widely used by autonomous vehicles to integrate the perception results from camera and LiDAR modalities. Despite the rapid development of end-to-end multimodal frameworks for autonomous driving tasks, they latently assume precise cross-modal spatiotemporal registration, which is constrained by the inherent spatiotemporal bias of real-world multimodal data. Although prior works have studied to bridge the camera and LiDAR cross-modal heterogeneity gap, they rarely address the spatiotemporal mismatch explicitly in realistic driving conditions. To this end, we propose RangeCLIP, a CLIP-like cameraLiDAR multimodal pretraining framework for tolerance of spatiotemporal bias without strict calibration. First, we convert LiDAR signals into range images to fit the input shape of 2D vision encoders. To ensure robust multimodal representation against spatiotemporal bias, we introduce a hierarchical contrastive alignment architecture that aligns camera color semantics and LiDAR depth cues at both global and patch levels initially. Then, the global and patch level features are refined via a multi-scale transformer, which enables the top-level alignment after comprehensive semantic interaction. Furthermore, we leverage cross-modal refinement to emphasize foreground-relevant patches while reducing batch-induced variance. Extensive experiments based on the NuScenes dataset show that the proposed RangeCLIP achieves a 22.5% improvement on the Top@1 metric for cross-modal retrieval, and can bring significant downstream robustness gains.

Abstract:
Vision-Language Models (VLMs) excel in zero-shot classification, image captioning, and visual question answering, yet their robustness to wrinkle-like non-rigid deformations remains unclear. We propose a parametric structural attack inspired by three-dimensional fabric wrinkles. The method constructs multi-scale wrinkle fields and couples displacement-field warping with surface-consistent appearance modulation to produce visually plausible perturbations. A hierarchical fitness function balances attack effectiveness and naturalness in a low-dimensional space optimized by a genetic algorithm. Perturbations optimized on a zero-shot classification proxy are transferred to generative tasks without downstream re-optimization. Across multiple state-of-the-art VLMs, our attack causes substantial performance degradation and consistently outperforms representative baselines in captioning and VQA.

Abstract:
The massive adoption of deep facial verification models creates serious privacy risks, motivating adversarial-example-based protection for facial multimedia against unauthorized recognition and tracking. However, existing protection methods rarely exploit explicit target identity priors during optimization, and diffusion protection using unconditional or null-text conditioning offers limited target-aware control, making it challenging to achieve a desirable balance between protection efficacy and visual realism. To address these issues, we propose Adv-Tokens, a diffusion-based method for transferable facial privacy protection. Adv-Tokens models the source-to-target identity shift in FaRL space and maps this residual prior to continuous identity tokens. It then injects these tokens into a frozen diffusion model through a controllable soft-conditioning interface for targeted control during reverse diffusion and refines each image in latent space from a DDPM inversion initialization. This design introduces explicit target identity control into diffusion-based protection while preserving natural appearance and facial structure. Experiments on public datasets and commercial APIs demonstrate superior black-box protection and a better trade-off between protection efficacy and visual fidelity than prior methods.

Abstract:
Face forgery detection (FFD) is crucial for mitigating the threats posed by diverse DeepFake techniques. Cutting-edge detection methods have begun to explore Contrastive Language-Image Pretraining (CLIP) for FFD due to its strong generalizability. However, adapting CLIP for FFD inevitably encounters a discrimination-generalization dilemma: generalizable knowledge from CLIP should be preserved, while task-specific discrimination knowledge (i.e., forgery-related artifacts) requires fine-tuning. How to achieve a trade-off between these two types of knowledge remains a challenge. To tackle this challenge, we propose Bidomain Artifact-Aware CLIP, termed BiA-CLIP, which boosts CLIPs forgery discriminability while preserving its generalizability. We find that CLIPs shallow layers learn fine?grained artifacts that benefit forgery discriminability, whereas its deep layers encode abstract features that support generalizability. Subsequently, a Bidomain Artifact Generator (BAG) is devised to introduce discriminative fine-grained artifacts into CLIP's shallow layers without sacrificing its generalization knowledge. Considering shallow and deep layers contribute oppositely to discrimination and generalization, we present Fine-Grained Artifact Adapter (FGA) to decouple and separately optimize these two objectives. Empirical results show that BiA-CLIP surpasses current state-of-the-art detectors by an average image-level AUC of 6.30% in all cross-dataset scenarios, utilizing only 3.2M trainable parameters.

Abstract:
Deepfake detection faces a critical cross-domain generalization challenge, as existing methods often overfit to dataset-specific artifact shortcuts and lack principled theoretical guidance for generalizable detector design. In this work, we conduct a Rademacher complexity-based generalization bound analysis, revealing that a detector's generalization error upper bound is strongly correlated with the feature-space distance between real and fake training samples. Guided by this insight, we propose GenDet-CLIP, a CLIP-based deepfake detection framework that (i) employs Class-Invariant Feature Augmentation to tighten the generalization bound via collaborative feature alignment, and (ii) introduces a Cosine Balance Consistency Constraint to mitigate the asymmetric convergence bias during detector training. Extensive experiments on mainstream deepfake benchmarks show that GenDet-CLIP achieves state-of-the-art cross-dataset and cross-method performance, improving average frame-level AUC by 1.1%, video-level cross-dataset AUC by 1.7%, and cross-method AUC by 0.6%, validating the effectiveness of our theory-driven design.

Abstract:
Instance segmentation in dense scenes is a challenging problem due to severe object crowding, ambiguous boundaries, and frequent overlaps, which are commonly encountered in applications such as histopathological analysis, remote sensing, and industrial inspection. Existing approaches, including prompt-free and prompt-guided methods, often suffer from excessive false positives and inaccurate boundaries due to conflicting modeling signals between pixel-level geometry and instance-level semantics: prompt-free methods emphasize local geometric consistency but lack instance discrimination, while prompt-guided approaches capture global instance priors yet overlook fine-grained boundary details. Consequently, these paradigms fail to achieve a balanced integration of geometric precision and instance-level reasoning. To address this issue, we propose GIC-SAM, a unified framework that explicitly couples geometric modeling with instance-level reasoning. Specifically, we propose a Semantic-aware Multi-granularity Prompter (SMP) to jointly encode geometric structures and semantic cues, generating boundary-aware representations and instance-aware prompts for accurate instance perception. Building upon this, a Multi-Level Mask Fusion Strategy (MLFS) is introduced to perform instance-level validation and cross-level refinement, enabling more accurate instance predictions by suppressing false positives and improving boundary quality. Additionally, we propose a Robust Prompt Training (RPT) strategy to mitigate the discrepancy between training and inference prompts by explicitly modeling prompt uncertainty, thereby enhancing robustness. Extensive experiments on four dense-scene benchmarks demonstrate that our method consistently outperforms state-of-the-art approaches, achieving superior boundary quality, fewer false positives, and improved robustness in complex scenarios.

Abstract:
Recent advances in generative models have exposed a fundamental weakness of image fake detectors: they often fail to generalize to unseen synthesis methods. We attribute this failure to an optimization pathology in which detectors overfit to source-coupled artifacts and converge to sharp solutions, resulting in unstable decision boundaries under generator shifts. To address this issue, we propose a geometric optimization framework that treats diverse textual prompts as structured semantic perturbations in the joint image--text space. To ensure that these perturbations provide reliable supervision, we introduce an entropy-based anchor selection mechanism that discards ambiguous prompt conditions and retains semantically compatible anchors for each input. Based on these anchors, we further impose two complementary regularizations: (i) perturbation consistency regularization, which enforces stable predictions across nearby semantic conditions and encourages local smoothness of the decision function; and (ii) manifold expansion regularization, which preserves feature diversity in the residual space and alleviates representation collapse. Together, these components steer optimization toward flatter and more transferable solutions. Extensive experiments on multiple benchmarks show that our method substantially improves cross-generator generalization, outperforming previous state-of-the-art methods by 5.65% in average accuracy.

Abstract:
Recently, the integration of generative priors has revolutionized Monocular Depth Estimation (MDE), achieving state-of-the-art (SOTA) performance with remarkably high data efficiency. However, although methods based on Text-to-Image (T2I) models excel in general scenes, their efficacy degrades sharply when confronted with open-world environmental perturbations, such as illumination variations and adverse weather. Since T2I priors are optimized for semantic synthesis rather than preserving pixel-level structure, they struggle to decouple geometry from appearance under severe visual shifts, easily triggering structural hallucinations. Therefore, our core insight is to model open-world depth estimation as consistent geometric inference across multi-style images, fully leveraging the powerful structure-preserving editing priors of Image-to-Image (I2I) models. Based on this, we propose OpenDepth, a robust open-world MDE framework that introduces ``Intra-Domain Awareness''the capability to map diverse stylistic inputs sharing identical geometry into a unified latent space. To achieve this, we first align shallow features within the Diffusion Transformer (DiT) to anchor degraded observations to unified geometric representations. Additionally, we introduce a joint RGB latent reconstruction task, requiring the model to recover clean RGB information from degraded inputs. This multi-stage strategy effectively decouples environmental style interference, compelling the model to focus strictly on underlying geometric recovery. Extensive evaluations show that OpenDepth sets new SOTA records for robust zero-shot depth estimation across standard and extreme weather benchmarks, demonstrating exceptional structural integrity and adaptability.

Abstract:
Multi-view learning and multi-label classification are both vibrant research areas, and their composite task, multi-view multi-label classification, has drawn growing attention. In this paper, we propose an Active Quality Assessment and Hierarchical Aggregation (AQA-HA) framework, which abandons the explicit shared-private feature decoupling and instead enables the model to adaptively extract multi-granularity features at multiple scales. Concretely, we propose a pyramid-style hierarchical aggregation module to progressively fuse coarse-to-fine feature, mitigating information dilution from coarse averaging and avoiding insufficient view decoupling. Besides, an active quality assessment module is proposed to produce instance-level quality scores by combining representation uncertainty, and is adopted to dynamic fusion during both training and inference. To enhance multi-view consistency, we introduce a view quality-guided contrastive learning strategy to conduct effective multi-view aggregation. This design not only improves the flexibility of representation learning, but also allows the model to emphasize more informative views while suppressing noisy or unreliable ones in a task-aware manner. Moreover, the proposed framework can naturally handle both missing views and missing labels within a unified variational learning paradigm. Extensive experiments on multiple benchmarks show that AQA-HA consistently outperforms strong baselines in all metrics, while maintaining stable advantages on fully complete data, demonstrating good robustness and generalization.

Abstract:
Open-set camouflaged object recognition (OSCOR) aims to recognize known camouflaged categories while rejecting unseen ones. Unlike generic open-set recognition (OSR), unseen targets in camouflaged scenes are rarely obvious outliers, since camouflage suppresses discriminative evidence and keeps semantically novel objects visually consistent with their surroundings. This makes existing open-set methods, especially synthesis-based ones, unreliable in OSCOR, as their pseudo-unknowns are typically context-agnostic and poorly matched to real camouflage ambiguity. We address this problem with VLAS, a vision-language anchored synthesis framework for OSCOR. VLAS first builds open-world semantic anchors from a frozen vision-language model, providing plausible unseen directions beyond the closed label set. It then synthesizes pseudo-unseen features conditioned on image-specific camouflage context, making unseen evidence semantically novel yet visually compatible with the current scene. Finally, it learns a stronger rejection boundary through semantic alignment, known-unseen separation, diversity regularization, and uncertainty-aware optimization. Rather than rejecting unknowns as easy outliers, VLAS learns to reject targets that remain plausible under the same scene while being semantically inconsistent with known classes. Extensive experiments on dedicated OSCOR benchmarks and standard OSR datasets show that VLAS establishes a new state of the art (SOTA) in camouflaged open-set recognition while retaining strong generalization in generic settings.

Abstract:
Recent multimodal large language models (MLLMs) have shown promising performance on video quality assessment (VQA) tasks. However, adapting them to new scenarios remains expensive due to large-scale retraining and costly mean opinion score (MOS) annotations. In this paper, we argue that a pretrained MLLM already provides a useful perceptual prior for VQA, and that the main challenge is to efficiently calibrate this prior to the target MOS space. Based on this insight, we propose DPC-VQA, a decoupling perception and calibration framework for video quality assessment. Specifically, DPC-VQA uses a frozen MLLM to provide a base quality estimate and perceptual prior, and employs a lightweight calibration branch to predict a residual correction for target-scenario adaptation. This design avoids costly end-to-end retraining while maintaining reliable performance with lower training and data costs. Extensive experiments on both user-generated content (UGC) and AI-generated content (AIGC) benchmarks show that DPC-VQA achieves competitive performance against representative baselines, while using less than 2% of the trainable parameters of conventional MLLM-based VQA methods and remaining effective with only 20% of MOS labels.

Abstract:
In multi-view anomaly detection, more cross-view information can actually hurt. When multiple inspection views are naively fused in a reconstruction-based pipeline, normal cues from intact views propagate to the decoder, which faithfully reconstructs anomalous regions, collapsing the reconstruction gap the detector depends on. We call this failure mode \emphcross-view information leakage and show that effective multi-view fusion must explicitly restrict the information reaching the decoder. Building on this insight, we present GLAD (Global-Local Attention Driven framework), the first framework combining vision foundation model features with local and global cross-view fusion for multi-view anomaly detection. The Multi-view Merging Attention (MMA) module performs local cross-view fusion at linear complexity with learnable view importance weighting and token-wise gating, letting each view selectively incorporate fine-grained evidence from other views at \mathcalO(N) cost. The Object-Guided Attention (OGA) module captures global context by aggregating class tokens from all views into a single object-level representation and broadcasting it back to patch tokens via temperature-scaled sigmoid gating, replacing the original patch representations rather than adding a residual to preserve the reconstruction gap. Experiments on Real-IAD and MANTA-Tiny show that GLAD outperforms state-of-the-art methods across sample-, image-, and pixel-level metrics, confirming that principled information restriction is key to multi-view anomaly reasoning.

Abstract:
Video Anomaly Retrieval (VAR) aims to retrieve anomalous events from videos using natural language queries, facilitating public safety applications. While extensively studied Video-Text Retrieval (VTR) primarily focuses on querying normal activities, it overlooks the critical need to identify anomalous behaviors in real-world scenarios. Although Video Anomaly Detection (VAD) is dedicated to anomaly understanding, it is typically formulated as a binary classification task based on predefined labels. In contrast, VAR enables active retrieval of specific anomalous events. Therefore, VAR is more practical but also more challenging. To address the challenges arising when directly applying generic VTR models to the VAR task, we propose a novel framework called AVAR, which enables the transition from passive classification to active retrieval. First, we design a Text-Guided Action Targeter (TAT) to extract action-focused representations by leveraging parsed action semantics, thereby suppressing redundant background. Second, we introduce a Dual-Path Anomaly Reasoning Network (DARN), which jointly models contextual information and temporal evolution. A temporal sequence reversal penalty is incorporated to explicitly constrain temporal directionality, enhancing the models sensitivity to action ordering. Finally, we design a Query-Adaptive Soft Routing (QASR) module, which adaptively balances appearance-based matching and logic-based reasoning according to the query semantics. As one of the first systematic explorations on the large-scale SVTA benchmark, our method AVAR achieves state-of-the-art performance. Moreover, it demonstrates strong zero-shot generalization and few-shot adaptability on real-world datasets OOPS! and UCFCrime-AR.

Abstract:
Detecting anomalous accounts in blockchain networks requires jointly analyzing multiple modalities of account behavior, including transaction values (i.e., numerical modality), transaction frequencies (i.e., count modality), and interaction degrees (i.e., structural modality), across thousands of blocks. However, existing approaches inevitably face a fundamental tradeoff between high accuracy and low cost. Model-based approaches process all modalities from raw transactions and achieve high accuracy but are impeded by high cost, while sketch-based approaches compress each modality independently and are orders-of-magnitude more efficient but suffer from estimation errors that degrade detection accuracy. A key limitation of sketch-based approaches is that they cannot exploit correlations across modalities to correct per-modality errors. In this paper, we propose SketchScan, an accurate and low-cost framework for multimodal account anomaly detection in blockchains. SketchScan introduces Multimodal Sketch (MM-Sketch), a compact per-block data structure whose heavy hitter module captures the numerical and count modalities while its HyperLogLog module captures the structural modality. A lightweight neural encoder then learns to reconstruct randomly masked modality entries from the remaining multimodal and temporal context during benign periods. Anomalous accounts violate the learned cross-modal patterns and thus incur significantly higher reconstruction errors, enabling accurate detection. Extensive experiments on real Ethereum data with 94,402 blocks and approximately 9.4M transactions demonstrate that SketchScan improves F1-scores by 6%-35% over state of the arts across all detection problems. It detects all 376 attackers in a documented DDoS attack, and reduces memory by 37% and processing time by 88%.

Abstract:
Gait recognition is a non-intrusive biometric technique for security applications. Existing studies mainly rely on silhouette, skeleton map and parsing representations. Although silhouette captures global body shape, it is structurally sparse and lacks internal details. Skeleton maps encode different body parts but still miss fine shape cues. Parsing provides richer part-level structures, yet its effectiveness depends heavily on the quality of upstream parsing, such as label granularity and boundary precision. These distinct characteristics suggest the need for a more unified view on gait visual representations. To this end, we revisit existing representations from the perspectives of \emphstructural edge density and \emphsemantic richness. Silhouette-based methods rely on sparse boundary structures with single semantic foreground. Skeleton map-based methods are also structurally sparse, but encode different body parts with richer semantic information. In contrast, parsing-based methods use denser structural cues with finer semantic labels. This view reveals an underexplored direction: learning dense part-level structural information without explicit semantic labels. To fill this gap, we introduce Sketch as a new visual modality for gait recognition. Sketch is a representation-level label-free modality that extracts high-frequency structural cues from foreground-masked RGB without encoding explicit semantic part labels. Our experiments evaluate the effectiveness and limitations of Sketch under diverse conditions. The results further demonstrate that Sketch and Parsing effectively complement each other: Sketch compensates for structural boundaries missing in Parsing, while Parsing mitigates texture noise and other irrelevant patterns over-extracted by edge-based Sketch. To better exploit the complementary characteristics of these two modalities, we propose SketchGait, a multimodal framework with two independent streams for modality-specific learning and a lightweight early-fusion branch. Furthermore, by incorporating the Silhouette modality, we extend it to SketchGait++, which further improves the overall performance. Extensive experiments on SUSTech1K and CCPG demonstrate the effectiveness of the proposed modality and framework.

Abstract:
Text-video retrieval has been significantly advanced by pre-trained vision-language models, yet most methods overlook the inherent audio-visual synergy by treating videos as silent frames. Existing audio-enhanced approaches typically rely on monolithic fusion architectures that employ simple gating or cross-attention to aggregate modalities. However, such designs suffer from cross-modal semantic interference, particularly in matched distractor scenarios where physically synchronized but task-irrelevant acoustic signals (e.g., ambient noise) dominate the fusion process, overwhelming the sparse, query-relevant cues. To address this, we propose MoE-Fuser, a sparse Mixture-of-Experts (MoE) framework that reformulates multimodal integration as a query-conditioned dynamic routing process. Specifically, we design: (1) a Gate Router Inter-Modal MoE that utilizes a text-driven semantic filter to selectively dispatch audio-visual tokens, effectively suppressing task-irrelevant modality noise; and (2) a Dual Router Cross-Modal MoE that disentangles top-down semantic intent from bottom-up multimodal correlations through parallel routing branches. Extensive experiments on 4 datasets demonstrate that MoE-Fuser achieves state-of-the-art performance.

Abstract:
Existing multimodal graph retrieval methods, like MMGraphRAG, rely heavily on fine-grained binary entity relations extracted by Large Language Models (LLMs). However, such micro-level graph construction often disrupts the holistic context of long documents, resulting in semantic fragmentation, the "visual island" problem, and cross-page "path disruption". To address these limitations, we propose CHORD (Clustering Hyperedging to Organize Relevant Documents), a dual-layer hypergraph framework for multimodal long-document understanding. CHORD first constructs Text and Visual Physical Hyperedges from text chunks and images, and then applies unsupervised Leiden community detection to organize them into cross-modal Semantic Hyperedges. During retrieval, CHORD matches relevant hyperedges, unpacks and deduplicates the retrieved content, and feeds the high-purity context to an LLM for answer generation. Experiments on DocBench, MMLongBench-Doc, and LongDocURL demonstrate that CHORD delivers strong and robust performance across diverse multimodal long-document benchmarks. It consistently outperforms graph-based baselines and shows clear advantages over Native LVLMs in structured multimodal reasoning scenarios, particularly those involving tables and charts. These results suggest that macro-level semantic organization is a promising and scalable paradigm for multimodal document understanding.

Abstract:
Fine-structure segmentation remains highly challenging. The target structures are thin and sparse, and they are often confused with visually similar background distractors. Generative data augmentation offers a promising way to alleviate data scarcity. However, existing approaches mainly focus on synthesizing the target structures themselves and struggle to jointly model realistic backgrounds with hard distractors. In addition, synthetic samples inevitably vary in quality, and low-quality generations can introduce harmful noise that degrades segmentation performance. To address these issues, we propose QSGen, a quality-scored synthetic data generation framework for fine-structure segmentation. It contains a distractor-aware generator that synthesizes complex backgrounds with diverse hard distractors. To achieve this, we use learnable pseudo-words as rich conditioning and introduce a concept-region alignment strategy for more controllable generation. We further design a structure injection module to incorporate diverse fine structures under explicit mask guidance. This module is built on a dual-UNet architecture with residual gating. To reduce the impact of noisy synthetic data, we also introduce a dual-criterion quality assessor that filters samples according to appearance realism and structural consistency. We evaluate QSGen on two public datasets and one self-collected dataset spanning diverse fine-structured segmentation scenarios. Experimental results show that QSGen outperforms state-of-the-art data augmentation methods. More importantly, it consistently improves the performance of several mainstream segmentation models under a unified augmentation setting.

Abstract:
Unmanned aerial vehicle (UAV) vision-language navigation requires agents to perform autonomous navigation in complex 3D environments by following natural language instructions. Existing methods face critical challenges when processing long sequences of historical observations, including visual redundancy and insufficient temporal context modeling. To address these issues, we propose a UAV vision-language navigation framework that integrates instruction-history interactive filtering with Mamba-based temporal modeling. First, we design an Instruction-Guided Keyframe Selection (IGKS) mechanism that dynamically extracts semantically relevant keyframes based on interaction between the instruction and historical observations, effectively compressing input sequences while preserving spatiotemporal context. Subsequently, we introduce the state space model Mamba to efficiently model temporal dependencies among the selected keyframes, capturing implicit motion trends and environmental evolution patterns during navigation, thereby enhancing the agent's understanding of historical trajectories. Finally, the resulting temporal features are used to predict the next action with a large language model. Experimental results on public UAV VLN dataset demonstrate its effectiveness in terms of navigation success rate and path efficiency, particularly exhibiting stronger robustness and semantic alignment in complex scenarios.

Abstract:
Sketch-based 3D human reconstruction faces significant challenges due to quality degradation in user-drawn sketches. Existing methods directly regress 3D meshes from raw sketches, achieving high-quality results on synthetic sketches but experiencing severe pose distortions on degraded hand-drawn sketches with irregular strokes and inconsistent proportions. We propose ControlPose, a two-stage framework that decouples sketch quality refinement from geometric reasoning. In the first stage, SAC-Diffusion (Sketch Adaptive Control Diffusion) employs a Quality-aware Hierarchical Adaptation (QHA) Block that predicts a learnable quality decay factor to modulate control strength across diffusion model layers, enabling adaptive refinement of stroke-level irregularities while preserving pose-level semantics. In the second stage, a Human Pose Mesh (HPM) Reconstruction Head with Text-Routed mixture of experts (TR-MoE) Decoder performs dual-sketch cross-attention fusion and text-driven expert routing, selectively leveraging pose-informative text descriptions while maintaining robustness to low-quality inputs. Experiments on the Sketch3D dataset demonstrate that ControlPose significantly outperforms existing methods, successfully bridging the performance gap between synthetic and hand-drawn sketches.

Abstract:
Recent advances in customized image generation exhibit strong capabilities in preserving global identity, yet they persistently struggle to synthesize consistent fine-grained details such as text and logos. To bridge this gap, we present FineForge, a unified dual-driven reinforcement learning framework designed to explicitly enhance high-frequency structural fidelity within a single denoising trajectory. First, we formulate an edge-aware terminal reward that provides an output-level, black-box learning signal for local detail accuracy. Second, to overcome the sparse reward dilemma inherent in standard RL, we introduce Advantage-Weighted Local Feature Alignment (AW-LFA). This white-box regularization module dynamically leverages the terminal reward as a gating mechanism to perform dense, patch-level feature distillation, directly aligning the model's internal hidden states with a pristine semantic space. Evaluated on customized generation benchmarks, FineForge demonstrates substantial improvements in text integrity, logo consistency, and structural sharpness, while preserving coherent global layouts.

Abstract:
Flow-based image generators achieve strong semantic synthesis but still produce localized structural distortions, such as malformed hands and limbs. Existing post-training methods typically optimize global scalar rewards, which provide weak spatial credit assignment and may alter well-formed regions while correcting local defects. We present Diagnose-to-Generate (D2G), a grounded alignment framework that separates visual diagnosis from generative optimization. We construct Achilles-600K, a region-annotated dataset for structural anomalies, and train ARGUS through supervised cold-start and Region-Aware Group Relative Policy Optimization to predict anomaly categories and bounding boxes. These diagnoses are converted into confidence-weighted spatial rewards for our Patch-level Flow-GRPO. We further identify that global stochastic exploration can shift semantic regions across group rollouts, making patch-wise relative advantages spatially inconsistent. Our Trajectory-Preserving Stochastic Sampling (TPSS) anchors early denoising with deterministic ODE trajectories and introduces stochasticity only at later steps, thereby limiting intra-group spatial drift while retaining local exploration. Experiments on FLUX.1-dev and SD3.5 show improved structural fidelity with preserved FID and CLIP scores. Consistent gains under HumanRefiner, zero-shot GPT-4.1, and blinded human evaluation further demonstrate that the improvements extend beyond the training reward model.

Abstract:
Audio-capable foundation models enable end-to-end spoken interaction, but they also introduce safety risks beyond transcript content. It remains unclear how much jailbreak capability can arise from matched-text variation in speech delivery rather than from lexical rewriting or broader style transfer. We study this question by holding transcript content fixed and varying six speech-delivery presets whose acoustic attributes may co-vary. We present PJ-Break, a black-box evaluation protocol with presets targeting arousal, authority, and speaking rate, together with AdvAudio-Prosody, a 600-sample benchmark with acoustically verified attributes. On the exact post-QC Qwen2-Audio panel, the Q =1 Panic (38/95), Anger (35/95), and Fast (32/95) presets are all well above Neutral (4/95). The fixed six-query pool covers 44/95 Qwen2-Audio seeds and 15/95 GPT-4o seeds and exceeds a matched-budget StyleBreak reimplementation (27/95) on Qwen2-Audio. A same-voice pool excluding the confounded Commanding condition still reaches 40/95, and a retained-panel ablation shows emotional-delivery audio alone (44/95) is far more effective than emotional text alone (11/95) . Exploratory surrogate diagnostics and pilot mitigation observations are secondary, non-core analyses. Overall, matched-text speech delivery should be treated as a first-class factor in Audio LLM safety evaluation.

Abstract:
Parameter partitioning mitigates catastrophic forgetting in Continual Learning by isolating task-specific parameters. However, in Federated Class-Incremental Learning (FCIL) with heterogeneous client data, this approach requires separate parameter subsets for each client-task combination. As clients scale, this rigid fragmentation severely restricts per-task capacity. To address this, we make two observations. Firstly, we observe that classifier parameters are particularly sensitive to catastrophic forgetting, as Cross-Entropy Loss under partial-class training amplifies logits of present classes while suppressing absent ones, leading to classifier weight collapse. Thus, classifiers should be partitioned, while other parameters can be shared to maximize model capacity. Secondly, in FCIL, different clients' tasks contain overlapping classes, whereas traditional task-level partitioning ignores such correlations. To address this limitation, we partition classifier parameters at the class level rather than the task level, enabling class-specific knowledge sharing across tasks and clients. Based on these insights, we propose FedSAP (Federated Sensitivity-Aware Partitioning). FedSAP partitions classifier parameters by classes, freezes parameters of absent classes during training to prevent overwriting, and maximizes angular separation between classifier boundaries of different tasks, preventing inter-class confusion across tasks. Experiments demonstrate that FedSAP outperforms state-of-the-art baselines.

Abstract:
\beginabstract Low-Rank Adaptation (LoRA) has become a cornerstone of parameter-efficient fine-tuning for large language models (LLMs) and vision-language models (VLMs). While incorporating Sharpness-Aware Minimization (SAM) principles theoretically enhances generalization by prioritizing perturbation-sensitive samples, existing approaches face two critical limitations: prohibitive computational overhead that negates LoRA's efficiency, and a failure to exploit sample heterogeneity by treating all instances uniformly. To address these issues, we propose a \emphSharpness-Score Sampling framework for LoRA (S^3-LoRA). Specifically, to enable efficient per-sample sharpness estimation, we design a lightweight sharpness proxy using only a single perturbed forward pass, guided by Exponential Moving Average (EMA) perturbations. Furthermore, we introduce a drift-corrected scoring mechanism that dynamically distinguishes intrinsic perturbation sensitivity from global optimization drift and persistent sample difficulty, which enables practical online example selection without incurring the full cost of SAM-style estimation. We provide theoretical analysis for the perturbation mechanism and the effectiveness of prioritizing high-sensitivity examples. % Extensive experiments on reasoning, language understanding, code generation, and vision-language benchmarks demonstrate that S^3-LoRA achieves a more favorable efficiency--performance balance for LoRA fine-tuning, yielding stronger or comparable performance with lower training cost. Extensive experiments on reasoning, language understanding, code generation, and vision-language benchmarks demonstrate that S^3-LoRA consistently achieves superior or comparable performance with more than a 1.3× training speedup, establishing new state-of-the-art benchmarks for this field. \endabstract

Abstract:
This paper presents NumAgent, a multi-agent framework for automated physical field simulation. Rather than treating PDE solving as a fixed pipeline, NumAgent formulates it as a closed-loop agent process that unifies solver selection, numerical execution, surrogate construction and experience accumulation within a single framework. To support this formulation, it adopts a hierarchical architecture that adaptively coordinates traditional numerical methods and neural operators across diverse PDE types and complex parameter spaces. It further incorporates experience memory and a reinforcement-learning-inspired value feedback mechanism, enabling continual improvement in retrieval, decision-making, and execution based on historical outcomes. When no suitable neural model is available, NumAgent automatically builds a reference solver, performs parameter sweeps, generates training data, and registers newly trained models for future reuse. Together, these designs move beyond static PDE workflows and provide a general framework for building adaptive and continually improving PDE-solving agents.

Abstract:
Due to the scarcity of annotated data and strict privacy regulations in cross-center deployment, single-domain generalization (SDG) has become imperative for high-precision medical image analysis under resource constraints. Most existing methods rely heavily on static isolated 2D image textures for feature extractions. However, these methods often fail to break the accuracy ceiling as they ignore longitudinal context and inherent structural correlations, causing early perception bottleneck and weak structural robustness. To address these issues, we propose a novel Prior-Aware Spatio-Temporal Evolutionary Distillation Dual-Stream Network (PASTE-D2Net) for accurate medical single-domain generalization, while providing meaningful interpretation on the results. In particular, a new spatio-temporal evolutionary distillation strategy is designed to capture the longitudinal and cross-sectional patterns of diseases, endowing PASTE-D2Net with dynamic evolutionary inference to effectively enhance the prominence of early lesions. Furthermore, a novel structure-texture dual-stream integration mechanism is introduced to compensate for the impacts of missing modalities, promoting the capacity of PASTE-D2Net in anchoring on anatomical structure to enhance its robustness towards complex clinical scenarios. Extensive experimental results on multiple datasets indicate that PASTE-D2Net outperforms the state-of-the-art methods and provides meaningful interpretations, offering a new data-efficient paradigm for clinical deployment.

Abstract:
Multimodal Large Language Models (MLLMs) offer flexibility for Document Information Extraction (DIE) but struggle with complex 2D layouts due to linear decoding constraints. Performance further declines when OCR serialization errors disrupt logical reading order. We present RyO-DIE, a generative framework that calibrates reading order before extraction. A Layout-aware Order Recovery and Detection (LORD) module converts irregular 2D layouts into a 1D structural prior (Line Hint). This hint guides instruction-tuned MLLMs to jointly decode Semantic Entity Recognition (SER) and Relation Extraction (RE) in a single pass. We also introduce UROBench, a benchmark with controlled structural nesting and reading-order labels to evaluate extraction under layout ambiguity. Experiments show RyO-DIE improves Relation Extraction F1 by 41.05% over generative baselines on the UROBench_struct subset. Additionally, RyO-DIE achieves competitive zero-shot performance on real-world datasets like FUNSD and CORD without task-specific fine-tuning.

Abstract:
Most existing LLM-based medical diagnosing approaches focus on specific tasks or modalities (e.g., classification, report generation, or visual question answering across MRI, CT, and X-ray). They lack the ability to diagnose and attribute diverse diseases and generalize poorly to unseen diseases in real-world clinical settings. To address these limitations, this paper proposes LLM-Driven Lifelong Adaptation for Unseen Disease Diagnosing and Attributing (L-UDA), a new task that enables MLLMs to continually learn disease knowledge from sequential medical imaging tasks, generate evidence-based attributing, and make diagnostic conclusions. This paradigm has significant clinical value and may support adaptation to unseen rare diseases. We further identify two critical challenges: 1) Intra-Disease Causal Features Elicitation and 2) Inter-Disease Medical Knowledge Sharing. To address them, we propose EvoDoctor, an evolutionary model that mimics a doctor who continuously acquires new disease knowledge while integrating prior clinical experiences for lifelong growth. EvoDoctor incorporates a Causal-based Disease Features Decoupling (CFD) block and an Expert-based Disease Knowledge Sharing (EKS) block, each targeting one challenge. We also construct an L-UDA dataset to simulate the continual emergence of disease diagnosing and attributing tasks in real-world clinical environments. Extensive experiments demonstrate that EvoDoctor outperforms strong continual learning baselines, highlighting the importance of L-UDA and validating EvoDoctor for disease-specific feature elicitation and cross-disease knowledge sharing.

Abstract:
The rapid advancement of generative AI has led to an explosion of GenAI music on online platforms, raising societal and ethical concerns. Despite its impact on human-centric creative communities, empirical analyses of GenAI music in real-world social environments remain scarce. In this paper, we perform the first large-scale analysis of social GenAI music shared via SoundCloud. Using a dataset of 72,227 GenAI tracks (posted by 18,382 artists from January 1^st, 2022 to June 30^th, 2025), we characterize their audio attributes, upload patterns, and community reception, revealing significant socio-technical frictions and increased toxicity towards GenAI content. We also investigate the phenomena of audio mimicry, finding that many artists are mimicked in their style by AI music. Finally, we design a lightweight multimodal AI music detector and benchmark it against the state-of-the-art. We demonstrate that by integrating social engagement features, tour model achieves performance comparable to heavyweight black-box deep learning models, while also providing explainability. Our work provides insights into the socio-technical dynamics of AI music, advancing the core tenets of responsible multimedia ecosystems.

Abstract:
Controllable Human Image Generation (HIG) aims to synthesize human-centric images that conform to user-specified conditions. However, even with recent diffusion-based approaches such as ControlNet, pose-guided HIG still fails to reliably preserve both structural consistency and fine-grained action accuracy. As a core contribution, we identify and formalize two underlying causes of these failures: structural misalignment and semantic insufficiency. Building on this analysis, we propose Prior-Guided Bidirectional Branch Cooperation (PriCo), a training-free inference framework consisting of two stages. The Structural Alignment stage uses stable guidance from the additional branch to align the structural prior in the initial noise with the target pose, while the Semantic Supplementation stage enhances the controllability of key tokens using pose-consistent fine-grained semantics. Compared with training-based methods, PriCo demonstrates competitive pose consistency and semantic fidelity across multiple datasets.

Abstract:
The Earth system is integral to every aspect of human life, and accurately forecasting the system states is vital in many domains. Modern sensing technologies monitor the Earth system through distributed networks, generating vast volumes of multivariate time series data, such as meteorological factors from weather stations or hydrological data from river networks, exhibiting intricate inter-variable and spatial dependencies. While Transformers are well-suited for processing 1D natural language or 2D vision data, their attention mechanism may struggle to capture underlying geographic structures and high-dimensional correlations inherent in Earth data. In this paper, we introduce Geography-aware Transformer (GeTS), which captures the 3D dependencies underlying the multivariate time series data alternately from the temporal, spatial, and variate views. To tackle the position-insensitivity of the attention mechanism, we apply attention with a novel High-dimensional Position Embedding (HiPE) strategy that meticulously encodes the geographical bias of each Earth observation. HiPE not only effectively integrates the off-the-shelf prior information into attention but also automatically discovers the latent relation in the high-dimensional system. In a set of empirical studies, GeTS achieves consistent state-of-the-art forecasting performance in weather, flood, and air quality, across both global and regional Earth systems.

Abstract:
While Multimodal Large Language Models excel at general visual question answering, their performance degrades significantly on tasks that depend on fine-grained details. This limitation makes it challenging for models to reliably handle tasks such as counting, document understanding, and complex reasoning. A recent strategy to enhance visual details is to perform multi-round inference, using the model's attention to crop a single, rectangular region of interest. This approach is computationally inefficient and is less effective for tasks that require analyzing multiple, arbitrarily shaped regions. To overcome these flaws, we identify a key bottleneck termed attention dilution and formulate the extraction of task-relevant semantic focus as a variational inference problem. Derived from the optimal solution under this formulation, we propose Ruminative Contrastive Attention Calibration (RumiA), a training-free inference method. RumiA computes the Conditional Pointwise Mutual Information of visual tokens via a Layer-Wise Attention Contrast, acting as a semantic unsharp masking filter that helps suppress background priors while maintaining the ability to attend to multiple regions. Capitalizing on the cross-scale stability of this distribution, RumiA leverages an efficient, smaller MLLM to perform a soft key-value calibration on a larger target model during a single inference pass. We evaluated our method on three MLLM families across ten benchmarks, focusing on fine-grained and complex reasoning. Our results show that RumiA achieves competitive improvement over strong baselines and reduces the computational cost.

Abstract:
Multimodal Sentiment Analysis (MSA) requires integrating language, acoustic, and visual signals without sacrificing modality-specific sentiment evidence. Existing methods mainly improve either shared-private decomposition or cross-modal interaction. Although effective, both ultimately depend on how shared and modality-specific evidence is organized before prediction. We observe that, under standard shared-private pipelines, modality heterogeneity often induces a branch-imbalance process: dominant shared patterns accumulate in the shared branch, yielding redundant and modality-biased evidence, while repeated interaction and rigid alignment gradually leak shared information into modality-specific channels and weaken discriminative private representations. As a result, the complementarity between shared and private representations is reduced, limiting robust sentiment reasoning. To address this issue, we propose the Dual-Branch Rebalancing Framework (DBR) on top of a standard multimodal decoupling stage. In the shared branch, a Temporal-Structural Factorization (TSF) module disentangles temporal evolution from structural dependencies and adaptively integrates them to reduce shared redundancy. In the private branch, an Anchor-Guided Private Routing (AGPR) module preserves discriminative modality-specific patterns while allowing controlled cross-modal borrowing. A Bidirectional Rebalancing Fusion (BRF) module then reunifies the two regularized branches in a context-aware manner for final prediction. Extensive experiments on CMU-MOSI, CMU-MOSEI, and MIntRec demonstrate that DBR consistently outperforms the compared baselines. Further analyses show that these improvements come from coordinated mitigation of branch imbalance.

Abstract:
% sec_abstract.tex \beginabstract Cross-modal composition introduces a category of safety risk that is fundamentally multimodal: individually innocuous images and text can become harmful after visual representations enter the language model and interact with textual context, yet existing safety methods treat the connector's projection interface as a black box. Many fine-tuning approaches do not explicitly model representation-level risk, while several safety-unlearning objectives apply sample-uniform pressure across connector activations. This can increase over-refusal and limit robustness to unseen cross-modal threat compositions. We propose \method, a framework that intervenes at the multimodal connector, a controllable visual-to-language projection interface, by estimating and suppressing a low-rank, risk-associated subspace of its visual activations. Risk-Structured Fusion (RSF) extracts explicit visual evidence and fuses it with connector activations through a learned confidence gate, producing image-conditioned risk scores rather than a uniform prior. Connector-Localized Selective Unlearning (CLSU) estimates the visual connector subspace via risk-weighted SVD, suppresses risk-aligned components, and preserves the orthogonal complement. Response-level losses on complete image--instruction--response examples jointly adapt downstream LoRA modules. Iterative Feedback Correction re-estimates the subspace from residual hard samples after each round, progressively tightening the risk boundary. We evaluate the main SafeEraser and VLGuard experiments on LLaVA-1.5-7B and InternVL2.5-8B. SIUO out-of-distribution evaluation is conducted on LLaVA-1.5-7B. Across these settings, \method improves the overall safety--utility trade-off, reduces over-refusal, and shows stronger robustness to out-of-distribution cross-modal threats. These results support the multimodal connector as a targeted intervention point for multimodal safety unlearning. \endabstract

Abstract:
Ultra-long egocentric video understanding requires reasoning over temporally sparse evidence distributed across hours or days, challenging current multimodal models with limited context and the grounding of key video segments. While Chain-of-Tool-Thought (CoTT) agent systems enable iterative retrieval and inspection, they suffer from error propagation due to rigid zoom-in strategies that lack recovery mechanisms. In this work, we address these challenges through SCOUT (Self-Checking Chain-Of-Tool-thought), a recovery-aware agentic framework introducing an adaptive policy that evaluates intermediate tool observations and dynamically trades off exploitation (zoom-in) and exploration (region switching), enabling robust multi-hop reasoning over extremely long horizons. However, training such multi-turn tool-using agents remains challenging, as existing RL methods rely on sparse outcome-level rewards and lack supervision over extended decision trajectories, resulting in suboptimal credit assignment for long-horizon reasoning. To address this, we develop UPS-GRPO, an uncertainty-prioritized policy optimization method that concentrates exploration on high-uncertainty post-tool states while preserving sample efficiency. We further introduce a turn-level advantage decomposition that integrates outcome rewards with tool-grounded temporal alignment rewards for improved credit assignment. Experiments show that SCOUT achieves state-of-the-art results on ultra-long egocentric benchmarks, while remaining competitive on shorter-horizon long-video settings.

Abstract:
Incomplete Multi-view Clustering (IMVC) aims to exploit complementary information from partially observed views to uncover latent semantic structure. Despite recent progress with impute-then-cluster methods, two challenges remain under high missing rates: (i) cross-view semantic misalignment: view-specific distribution shifts impair instance-level cross-view alignment and induce prototype drift, leading to unreliable cross-view correspondences; and (ii) global inconsistency beyond pairwise alignment: with more than two views, independently learned pairwise correspondences can be non-transitive, distorting the cross-view topology. To address these issues, we propose Cycle-Consistent Optimal Transport and Hybrid Imputation (COTHI). COTHI casts cross-view prototype matching as an optimal transport problem with entropy regularization to learn robust soft correspondences, and introduces a cycle-consistency regularizer to enforce globally coherent multi-view alignment. Building on the learned correspondences, we further develop a hybrid imputation strategy that completes missing-view representations by combining prototype-guided semantic transfer with cross-view neighbor fusion. Experiments on five benchmark datasets demonstrate that COTHI consistently outperforms state-of-the-art IMVC methods, especially at high missing rates.

Abstract:
Long video understanding remains a formidable challenge for Multimodal Large Language Models (MLLMs) due to the prohibitive cost of processing dense frame sequences. Prevailing keyframe-selection methods rely on either a single visual-centric metric (e.g., CLIP similarity) or a static fusion of heuristic scores. This "one-size-fits-all" paradigm frequently fails: visual-only metrics are ineffective for plot-driven narrative queries, while indiscriminately adding textual scores introduces severe "modal noise" for purely visual tasks. To break this bottleneck, we propose Q-Gate, a plug-and-play, training-free framework that treats keyframe selection as a dynamic modality routing problem. We decouple retrieval into three lightweight expert streams: Visual Grounding for local details, Global Matching for scene semantics, and Contextual Alignment for subtitle-driven narratives. Crucially, Q-Gate introduces a Query-Modulated Gating Mechanism that uses the in-context reasoning of an LLM to assess query intent and dynamically allocate weights across the experts, activating necessary modalities while "muting" irrelevant ones to maximize the signal-to-noise ratio. Extensive experiments on LongVideoBench and Video-MME across multiple MLLM backbones show that Q-Gate outperforms representative keyframe-selection baselines in most settings, with particularly strong gains on long and medium videos, providing a robust and interpretable solution for scalable video reasoning.

Abstract:
Multifaceted clustering is dedicated to excavating diverse and heterogeneous facets from complex multimedia data, which is a critical task for multimedia content understanding. Existing approaches primarily focus on decomposing a single dataset into multiple subspaces or sequentially generating alternative clusterings. However, these methods are severely constrained by the inherent limited information contained in the data itself, resulting in suboptimal clustering performance and indistinct facet mining results. To address the issue, we propose a novel paradigm shift from Single-to-Multi to Multi-to-Multi (MoM) clustering for multimedia data, which proactively integrates multiple informative views for decoupled multifaceted discovery, breaking the bottleneck of single-source data information scarcity. The proposed MoM framework consists of two core modules: the Dynamic Representation Fusion (DRF) module and the Decoupled Multifaceted Clustering (DMC) module. Specifically, the DRF module implements an adaptive fusion strategy that generates and dynamically fuses diverse representations guided by the anchors derived from a pretrained model, ensuring both fidelity and diversity. The DMC module leverages a structured self-expression model to jointly learn shared and facet-specific affinity matrices, producing well-separated and interpretable clusters. Unlike existing methods that passively decompose limited data into disjoint subspaces, our end-to-end framework actively employs a clustering-oriented representation fusion strategy to learn decoupled features tailored to each clustering facet. Extensive experiments show that MoM consistently outperforms state-of-the-art multifaceted clustering methods, validating the effectiveness of our MoM framework.

Abstract:
Retrieving the source of short edited videos from large-scale long video databases is an essential task in near-duplicate video retrieval (NDVR). Existing methods face a fundamental trade-off between accuracy and scalability: frame-level matching offers high accuracy but scales poorly due to expensive search cost, while video-level descriptors lose fine-grained information. Clip-level methods try to balance this trade-off by dividing videos into a set of clips and encoding each into a descriptor. However, they still struggle with redundant frames in raw video inputs, limiting both representation quality and compression efficiency. To address this, we propose a deduplicate-before-aggregate framework to explicitly decouple redundancy removal from representation learning. We first adopt a scalable preprocessing pipeline to filter low-quality and redundant frames from uniformly sampled inputs. The curated frames enable training a DINOv2 as the frame feature extractor, provide clean inputs for downstream clip-level encoding, and directly reduce retrieval cost. Building on this, we propose Q-Clip, a Perceiver-style model that uses a learnable query to iteratively aggregate frame features into a compact vector via cross-attention. Q-Clip is trained with self-supervised contrastive learning on bootstrap sampled clips from the same temporal interval, yielding clip-level representations robust to augmentations. To support evaluation, we also release the first long-form video retrieval benchmark with 1.3K full-length videos and 1.6K real-world queries. Experiments show that Q-Clip matches frame-level accuracy while achieving order-of-magnitude reductions in storage and search complexity over existing clip-level methods.

Abstract:
Recent feed-forward 3D Gaussian Splatting (3DGS) methods enable efficient novel view synthesis from unposed images in a single forward pass. However, their performance degrades significantly under motion-blurred inputs. Moreover, directly fine-tuning these models on blurry data leads to a counterintuitive phenomenon: reconstruction quality deteriorates as the number of input views increases. We identify the root cause as the lack of cross-view aggregation in the appearance branch---each view is decoded independently, so blurry appearances from multiple views interfere with rather than complement one another. % To this end, we propose BlurSplat, a feed-forward framework that reconstructs sharp 3D Gaussians from blurry multi-view images in seconds. BlurSplat combines \emphOcclusion-Aware Cross-View Appearance Aggregation (OCAA) with teacher--student geometric distillation to fuse complementary appearance cues across views and learn blur-robust geometry without paired 3D annotations. % Experiments on synthetic and real motion blur benchmarks show that BlurSplat performs feed-forward reconstruction in seconds, and with only a few minutes of per-scene refinement, BlurSplat+Opt achieves competitive or superior performance compared to per-scene optimization methods.

Abstract:
4D spatio-temporal reasoning, jointly modeling 3D spatial structure and temporal evolution, is essential for understanding dynamic worlds and enabling embodied interaction. While current Multimodal Large Language Models (MLLMs) show strong capabilities in static scene understanding and coarse-grained 4D tasks, they still have notable limitations in continuous dynamic scene perception, especially in tracking dynamic object evidence for coherent 4D spatio-temporal reasoning. This shortcoming stems mainly from relying on sparse frame-level observations, fragmenting continuous dynamic cues and leaving models unable to disentangle genuine object dynamics from camera-induced apparent motion. Inspired by humans tracking dynamic cues while compensating for viewpoint changes, we propose DynTrace, a training-free framework for 4D spatio-temporal reasoning with two complementary components. Dynamic Trajectory Visualization (DTV) reprojects world-coordinate trajectories onto the image plane, providing geometry-informed visual priors that disentangle genuine object dynamics from camera-induced apparent motion. Meanwhile, the Dynamic Trace Token (DT-Token), organized into a Dynamic Trace Graph (DTG), tracks object-level dynamic cues, trace evolution, and key moments, maintaining continuous dynamic object evidence for coherent 4D reasoning. Together, these two components equip MLLMs with continuously tracked dynamic object evidence, grounded in geometry-informed visual priors and structured spatio-temporal traces. DynTrace consistently improves open-source MLLMs, achieving state-of-the-art results on Dyn-Bench, VLM4D, and DSI-Bench, validating the importance of tracking dynamic object evidence for robust 4D spatio-temporal reasoning.

Abstract:
Unmanned Aerial Vehicle Vision-Language Navigation (UAV-VLN) requires agents to follow language instructions, infer spatial structure from sparse multi-view observations, and execute feasible 3D motion in complex outdoor environments. Despite recent progress with large language models, most existing methods still map vision-language inputs directly to actions, providing limited explicit scene grounding and future-aware spatial reasoning. We propose AirForesight, a current-to-future spatial map imagination framework for UAV-VLN. AirForesight first learns a structured current-map representation from multi-view observations. This representation is jointly supervised by current-map reconstruction and future-trajectory prediction, encouraging it to encode both present scene structure and future motion intent. Under structured causal attention, the current spatial knowledge is propagated to future-map reasoning, and the resulting current and future representations are aggregated to predict the next 3D waypoint. To make spatial imagination more relevant to navigation, we introduce a cross-space planning consistency loss that encourages directional agreement between the predicted map-space trajectory and the expert action direction derived from the ground-truth waypoint displacement. Experiments on OpenUAV and AerialVLN-S, together with extensive ablations, demonstrate strong performance and support the effectiveness and stability of the proposed framework.

Abstract:
Sparse structural learning is crucial for deploying Spiking Neural Networks (SNNs) on resource-constrained neuromorphic hardware. However, most existing sparse structural learning methods for SNNs, largely inherited from Artificial Neural Networks (ANNs), focus on static magnitude to reduce parameters while overlooking spiking activity. As a result, the remaining neurons may fire excessively, undermining energy savings and fundamentally misaligned with the intrinsic spatiotemporal dynamics of SNNs. Inspired by GABA conductance-based synapses, we propose a bio-inspired inhibitory mechanism for SNNs that leverages accumulated firing traces to selectively prune hyperactive redundancy. We further propose a Homeostatic Finetuning strategy that recalibrates BatchNorm biases and reweights gradients based on historical neuronal activity, to compensate for the distribution shift introduced by removing inhibitory connections. Our method naturally evolves a distinct inverted pyramid topology, which preserves information-dense convolutional layers while aggressively compressing parameter-heavy fully connected layers, leading to a better storageefficiency trade-off than conventional magnitude-based pruning. Experimental results on CIFAR-10, DVS-CIFAR10, and ImageNet datasets demonstrate that the proposed topology achieves competitive accuracy and energy trade-offs among existing methods.

Abstract:
Diffusion policies effectively model multimodal action distributions for robotic manipulation, but their iterative denoising requires tens to hundreds of network function evaluations (NFEs) for each control prediction, limiting their applicability to high-frequency closed-loop control and online reinforcement learning (RL). We present a two-stage framework for native one-step generative policies that transfers iterative refinement from inference to training. First, Drift-Based Policy (DBP) uses a fixed-point drifting objective to internalize corrective dynamics into the model parameters, producing multimodal action chunks with a single network evaluation by design. Second, Drift-Based Policy Optimization (DBPO) augments the pretrained backbone with a stochastic interface that provides exact conditional rollout likelihoods for PPO-style on-policy updates while preserving 1-NFE deployment. On the 12-task Diffusion Policy suite, DBP improves the average success rate from 0.79 to 0.83 while reducing inference from 100 NFEs to 1. Across 37 point-cloud manipulation tasks, DBP achieves an average success rate of 88.4%, surpassing the leading 1-NFE baseline OMP at 82.3%. DBPO further improves pretrained one-step policies through stable online fine-tuning on RoboMimic and D4RL. On a physical dual-arm UR5 platform, DBP achieves 123/150 successes (82%) with an average end-to-end latency of 9.5~ms, compared with MP1's 89/150 successes (59%) under the same setup.

Abstract:
Estimating high-fidelity 360 HDR environment lighting from a single limited-field-of-view (LFOV) LDR image is critical for photorealistic insertion of virtual objects in mixed reality. Existing diffusion-based methods hallucinate unobserved rear lighting from the full local scene context of the input, but lack physical illumination priors. By contrast, inverse rendering methods can decouple illumination from surface reflections of individual objects, yet their isolated per-object cues fail to capture coherent local scene context for omnidirectional lighting inference. Unifying these two complementary paradigms poses two core challenges: the severe spatial heterogeneity between standard perspective projections and geometry-invariant reflectance spaces, and the lack of a robust paradigm to guide diffusion models to jointly reason over physical reflection priors, coherent local scene context, and complex visual semantics. To address this, we propose ReCo, a physics-guided diffusion framework that elegantly redefines environment lighting generation, reformulating it as a physically-conditioned inpainting problem within an intermediate reflectance map space. By aligning the input background at the periphery and anchoring physical priors in the center, ReCo bridges physical constraints and semantic coherence, while inherently eliminating lateral seam artifacts plaguing direct equirectangular projection (ERP) generation. Finally, a dedicated rendering-aware module explicitly decouples latent features to reliably recover relative HDR intensities. Extensive experiments demonstrate that ReCo outperforms state-of-the-art methods, generating physically realistic, seamless 360 lighting with high-frequency details in real-world scenes.

Abstract:
Segment Anything Model 2 (SAM2) enables prompt-guided segmentation with cross-slice propagation, yet its adaptation to medical imaging often relies on idealized evaluation protocols. In realistic clinical scenarioscharacterized by unknown organ extents and sparse user prompts (e.g., a single click)achieving robust full-volume 3D segmentation remains a significant challenge due to error accumulation. To bridge this gap, we propose TIED-MedSAM2, a Test-time Inference Enhancement and Distillation framework. The framework operates in two stages: First, we stabilize long-range propagation using a suite of piggyback tunes, including BFS-based scheduling and multi-view voting, combined with a lightweight Meta-Prompt Correction Network that detects and rectifies early-slice failures. Second, we implement an organ-wise self-distillation strategy that uses enhanced predictions as pseudo-labels to internalize these self-correction behaviors into the model parameters. This allows the final distilled model to maintain high robustness while significantly reducing computational and interaction overhead. Experiments on BTCV and CHAOS-CT demonstrate that TIED-MedSAM2 consistently outperforms state-of-the-art SAM2-based baselines under strict, low-interaction settings. Overall, our work provides a practical and prompt-efficient paradigm for reliable 3D medical segmentation in real-world deployment scenarios.

Abstract:
Unsupervised cross-modal homography estimation remains a difficult problem in computer vision, mainly because of the substantial appearance discrepancies among different imaging modalities. Existing methods primarily rely on modality translation to reduce this gap and improve accuracy through iterative estimation, but imperfect translation often leaves residual inconsistency between the pseudo and real target domains. In addition, traditional iterative estimation pipelines rarely preserve historical correction information, making subsequent updates vulnerable to noise and prone to error accumulation. To address these issues, we propose a novel FMNet, Frequency calibration and Memory recurrent refinement Network for unsupervised cross-modal homography estimation. The proposed method contains two key components. First, a mid-frequency anchored calibration module is introduced to model the amplitude spectrum in separate frequency bands. By using the mid-frequency components across domains as anchors, the module calibrates low- and high-frequency responses, thereby alleviating residual domain discrepancies caused by imperfect modality translation. Second, a memory recurrent refinement decoder is designed for iterative estimation, where homography updates are formulated as a sequential recursive process which can improve the stability of the refinement process and suppress error drift. Extensive experiments on four cross-modal benchmarks demonstrate that FMNet achieves the SOTA performance among unsupervised methods.

Abstract:
Meteorological reanalysis encodes extreme weather through continuous, physically constrained fields, posing a fundamental challenge for visionlanguage models (VLMs) whose perceptual assumptions are shaped by natural images. Tropical cyclones exemplify this mismatch: critical properties such as intensity extrema, asymmetry, spatial extent, and physical impacts arise from field-level organization rather than object-centric visual cues. Existing approaches address this gap through text alignment or annotation, treating the problem as multimodal supervision rather than representation design. We introduce Physics-Aware Visual Abstraction (PAVA), a plug-and-play physics-aware representation and annotation interface that maps physical reanalysis fields to visually identifiable and semantically grounded perceptual abstractions for supervision and evaluation in vision--language reasoning. Building on PAVA, we construct OmniClimate-TC, a benchmark for tropical cyclone analysis spanning five classes of reasoning and nine tasks, with 243,890 physically grounded instruction-tuning pairs. Using PAVA-aligned supervision, we adapt VLMs and provide evidence that this representation design improves reasoning over tropical cyclone hazard fields. Our results position OmniClimate-TC as a benchmark for multimedia reasoning over structured geophysical fields, and highlight representation design as a key ingredient for physically grounded reasoning in scientific media.

Abstract:
VisionLanguage Models (VLMs) offer powerful multi-modal ability but also expose users to text-based privacy attacks where adversaries crawl online photos and query VLMs to extract sensitive attributes. Existing reversible adversarial example (RAE) methods protect images in purely visual tasks but fail in multi-modal settings, and current adversarial examples on VLMs rely on high-frequency noise that severely degrades visual quality. We propose CloakDiff, the first framework for reversible, high-fidelity privacy protection against text-based query attacks in VLMs. CloakDiff produces imperceptible adversarial examples by combining diffusion-based adversarial editing with an invertible network that embeds the original image for lossless recovery. It perturbs both pixel-space embeddings and manipulates latent cross attention maps to ensure strong cross-model and cross-prompt transferability while preserving global visual structure. To further enhance fidelity, we design EDM-Heuristic Sampling, a principled diffusion schedule for adversarial guidance. Experiments on multiple datasets and VLMs demonstrate that CloakDiff delivers multi-modal privacy preservation with high visual quality and reversibility.

Abstract:
Longitudinal medical image analysis, which tracks disease progression across sequential patient visits, plays a critical role in clinical prognosis and risk stratification. Despite their success in standard imaging tasks, existing medical image encoders lack dedicated mechanisms for capturing lesion-level temporal dynamics across visits. We introduce VITAL (Visually-Informed Temporal Anatomy-aware Learning), a framework for learning spatially grounded and temporally coherent chest radiograph representations from longitudinal sequences. VITAL leverages language-grounded anatomical encoding and temporally-aware contrastive learning to align visual features with clinical semantics at the region level, enabling fine-grained lesion localization within each visit and meaningful change modeling across visits. Extensive experiments demonstrate that VITAL achieves state-of-the-art performance on longitudinal clinical benchmarks, consistently outperforming both general-purpose and domain-pretrained encoders. Beyond longitudinal tasks, VITAL also generalizes effectively to a range of single-image and cross-modal tasks, suggesting its broader utility for chest radiograph representation learning.

Abstract:
The rapid deployment of Text-to-Image (T2I) diffusion models has raised urgent concerns about the generation of unsafe, copyrighted, and offensive content. Recently, steering-based concept unlearning methods have emerged as an efficient alternative to retraining, suppressing undesired concepts by injecting learned direction vectors into model representations. Despite their efficiency, these approaches rely on unconstrained directional updates that indiscriminately alter model behavior, often leading to over-suppression, cross-concept interference, and degradation of non-target generation quality. To address this issue, we propose RESET, a theoretically grounded framework for Representation steering via null-Space alignment in diffusion models. Moving beyond unconstrained heuristic steering, we steer representations under a null-space constraint to avoid interference with retained concepts. Specifically, we construct a preservation subspace induced by benign concept representations and restrict parameter updates to its orthogonal null space, thereby preventing interference with retained concepts. Within this constrained space, we derive a closed-form solution that selectively suppresses unsafe concepts while preserving general generation capability. Extensive experiments across object, style and identity benchmarks show that RESET consistently outperforms state-of-the-art methods, achieving more precise concept removal, improved robustness against red-teaming attacks, and superior preservation of unrelated content fidelity.

Abstract:
Modern text-to-image (T2I) models can now render legible, paragraph-length text, enabling a fundamentally new class of misuse. We identify and formalize the inscriptive jailbreak, where an adversary coerces a T2I system into generating images containing harmful textual payloads (\eg, fraudulent documents) embedded within visually benign scenes. Unlike traditional depictive jailbreaks that elicit visually objectionable imagery, inscriptive attacks weaponize the text-rendering capability itself. Because existing jailbreak techniques are designed for coarse visual manipulation, they struggle to bypass multi-stage safety filters while maintaining character-level fidelity. To expose this vulnerability, we propose \textscEtch, a black-box attack framework that decomposes the adversarial prompt into three functionally orthogonal layers: semantic camouflage, visual-spatial anchoring, and typographic encoding. This decomposition reduces joint optimization over the full prompt space to tractable sub-problems, which are iteratively refined through a zero-order loop. In this process, a vision-language model critiques each generated image, localizes failures to specific layers, and prescribes targeted revisions. Extensive evaluations across 7 models on the 2 benchmarks demonstrate that \textscEtch achieves an average attack success rate of 65.57% (peaking at 91.00%), significantly outperforming existing baselines. Our results reveal a critical blind spot in current T2I safety alignments and underscore the urgent need for typography-aware defense multimodal mechanisms.

Abstract:
Lifelong learning enables intelligent systems to acquire knowledge incrementally while preserving previously learned informationa hallmark of human cognition. However, most Spiking Neural Network (SNN) models lack the critical biological mechanisms that support this capacity, particularly sparse circuit reconfiguration and context-aware gating observed in the brain. Inspired by neuroscientific insights into synaptic rewiring and task-differentiated cortical modulation, we propose a novel framework for sparse structural continual learning in SNNs. Our approach integrates two core innovations: (1) a context-aware gating mechanism based on logistic regression that detects task-specific representations from spatio-temporal spike activity, and (2) an iterative sparse structure learning module that dynamically prunes and regrows task-specific connections in a biologically plausible manner. Together, these components facilitate efficient representation segregation, reduce catastrophic forgetting, and improve parameter efficiency. Experimental results on class-incremental CIFAR-100 benchmarks show significant gains in both continual learning performance and computational efficiency over state-of-the-art SNN baselines, which use only half the number of parameters, with up to 1.11% improvement in average accuracy and up to 2.67% increase in last accuracy. This work offers a biologically grounded pathway to scalable and adaptive lifelong learning in neuromorphic AI systems.

Abstract:
Food segmentation is essential for applications such as intelligent catering, dietary assessment, and food recommendation. However, existing benchmarks do not faithfully capture the dense inter-dish overlap, fine-grained class similarity, and extreme long-tail distributions of real-world dining scenes. To fill this gap, we introduce DishSeg24k, a large-scale dish-level segmentation benchmark with 24,096 images, 112,281 instances, and 278 fine-grained categories in real-world dining environments. Based on DishSeg24k, we further propose Food Expert-Adaptive Segmentation Transformers (FEAST) to address these challenges. FEAST models query-based decoding as a Markov Decision Process (MDP), treating each decoder-layer update as a sequential decision that enables exploration of uncertainty along dish boundaries. We further redesign the decoder with a reinforcement learning (RL)-guided Mixture-of-Experts (MoE) module, in which a decoupled dual-critic optimization scheme separates task-oriented query refinement from structure-aware expert routing. This design promotes expert specialization and prevents expert collapse under long-tail category distributions. Finally, extensive experiments on DishSeg24k demonstrate that FEAST outperforms the strongest baseline by 3.21% mIoU, 3.68% mDice, and 4.00% mAcc. We further validate the effectiveness of FEAST on FoodSeg103.

Abstract:
Building foundation models for medical imaging requires pooling data across institutions, yet privacy regulations prohibit centralized aggregation. Existing Federated Foundation Models either fine-tune natural-image models with poor medical-domain transfer, or train from scratch within a single modality, lacking the flexibility to unify tasks. We identify an under-explored challenge, \emphImaging Modality Heterogeneity, where clients operate under two structural regimes: \emphOverlapped (shared modalities with heterogeneous label distributions) and \emphNon-overlapped (fully disjoint modalities per client). We propose FM^2, a unified framework that trains the core backbone from scratch to preserve medical domain fidelity while optionally incorporating biomedical pretrained encoders for vision-language alignment. FM^2 equips each client with dual Mixture-of-Experts modules (a Class-wise MoE for personalized category knowledge and a Domain-wise MoE for shared cross-modality representations), coupled with a \emphHeterogeneous Modality Alignment (HMA) regularizer that explicitly aligns modality-specific expert parameters, admitting provable \gO(1/\sqrtT) convergence and generalization guarantees. FM^2 further incorporates \emphCaption-Enhanced Learning (CEL), where locally retained GPT-4o-generated captions serve as a textual semantic bridge enabling representation transfer across clients with disjoint modalities, and demonstrates extensibility to \emphFederated Medical VQA. Experiments on our MIMH benchmark (classification and CEL) and real-world medical VQA datasets confirm consistent superiority over state-of-the-art federated baselines and strong out-of-modality generalization across all three tasks.

Abstract:
Mobile graphical user interface (GUI) agents have demonstrated remarkable capabilities in automating complex tasks, yet they introduce critical safety risks because a single erroneous action can lead to irreversible consequences. Existing safety mechanisms are primarily reactive, lacking the ability to assess risks before execution. In this paper, we introduce SeerGuard, a consequence-aware safety framework designed to mitigate these risks through pre-execution instruction-level screening and action-level risk assessment. Specifically, the action-level assessment analyzes agent-proposed actions within current GUI states, anticipating likely outcomes to identify risks before they are executed. To enable these capabilities, we construct a unified safety-augmented world model (SAWM) via multi-task learning, integrating semantic next-state prediction with safety risk assessment. Extensive experiments demonstrate that SeerGuard generalizes effectively across diverse mobile GUI agents. On Qwen3-VL-8B-Instruct, it increases the safety-utility score from 0.191 to 0.596 at \omega=0.8 and reduces the risk-cost score from 0.347 to 0.130 at \alpha=0.8. Further analyses on our SAWM validate the effectiveness of the instruction-level screening, alongside the capability of action risk assessment and next-state prediction.

Abstract:
Multispectral filter array (MSFA) snapshot spectral imaging enables compact single-shot spectral acquisition for dynamic and real-time scenarios. However, the narrow-band filters in MSFA systems severely restrict photon throughput, producing photon-starved measurements with transmittance imbalance, illumination attenuation, and weakened structural details. Existing reconstruction methods mainly focus on spectral-spatial modeling from sparse mosaiced observations, but often overlook these photon-flux-induced degradations, resulting in suboptimal recovery and unreliable correction updates. To address this issue, we propose a Hierarchical Transmittance-Compensated Mamba framework (HTC-Mamba) for MSFA snapshot spectral reconstruction. A Local Transmittance Compensation Branch (LTCB) first estimates spatially adaptive gain priors from mosaiced inputs, while a Hybrid Pseudo-Panchromatic Generator (H-PPG) extracts global transmittance cues and structured guidance priors. These complementary priors are then hierarchically injected into a multi-scale Transmittance-Aware Mamba Backbone (TAMB) for degradation-aware reconstruction under photon-limited conditions. In addition, a MAP-inspired adaptive data-consistency module (MAP-CADC) adjusts correction strength based on measurement reliability and prior confidence, balancing reconstruction fidelity and noise suppression. Extensive experiments show that HTC-Mamba achieves superior reconstruction on simulated benchmarks in both quantitative and visual evaluations, while yielding favorable qualitative results on real MSFA measurements.

Abstract:
Multimodal large language models (MLLMs) enable interaction over both text and images, but their safety behavior can be driven by unimodal shortcuts instead of true joint intent understanding. We introduce CSR-Bench, a benchmark for evaluating cross-modal reliability through four stress-testing interaction patterns spanning Safety, Over-rejection, Bias, and Hallucination, covering 61 fine-grained types. Each instance is constructed to require integrated imagetext interpretation, and we additionally provide paired text-only controls to diagnose modality-induced behavior shifts. Our evaluation of 16 state-of-the-art MLLMs reveals a clear trade-off across reliability dimensions and highlights an overall performance degradation when shifting from unimodal to multimodal inputs. This suggests that models possess weak, non-transferable safety awareness. To address this, we propose U2M-Align, a two-stage training framework for improving cross-modal safety and reliability. Experiments across three model architectures show that U2M-Align improves all four reliability dimensions over unaligned baselines while preserving general capabilities, yielding a better overall balance. WARNING: This paper contains unsafe contents.

Abstract:
Composite image lighting adaptation aims to adjust the lighting of a composited foreground so that it harmonizes with the illumination of a target background. Recent diffusion models have significantly advanced this task by training on large-scale relighting pairs. However, existing methods often struggle to maintain both content consistency and lighting consistency in composite images. Moreover, the construction of relighting datasets and the corresponding training paradigm for lighting adaptation remain underexplored. To address these challenges, we propose HarmoLight, a multimodal-guided diffusion model for composite image lighting adaptation. Specifically, our model is guided by decoupled content and lighting prompts through a decoupled cross-attention mechanism, enabling more effective control over foreground appearance and illumination adaptation. To effectively derive appropriate lighting guidance, we further introduce a lighting representation distillation strategy that transfers lighting encoding knowledge from a large pre-trained model. In addition, to better facilitate the learning of diverse lighting transformations, we employ a curriculum lighting learning strategy over our specially designed datasets, where training samples are organized according to their lighting adaptation difficulty. Extensive quantitative and qualitative experiments on both synthetic datasets and real-world images demonstrate that our method consistently outperforms existing state-of-the-art approaches.

Abstract:
Resource constraints on UAV platforms have driven a paradigm shift in aerial tracking, from pursuing performance toward balancing accuracy with efficiency. Adaptive Transformer Trackers, which leverage input-dependent dynamic routing architecture, have emerged as a representative solution to this challenge. However, we reveal that behind this computation-on-demand flexibility hides a critical structural flaw: the Lipschitz singularity of computational path decisions, which has an unbounded local Lipschitz constant at discrete layer-skipping decision boundaries. This mathematical discontinuity renders adaptive tracking networks inherently unstable: tiny input perturbations can be amplified at the gating modules, causing dramatic changes in the inference topology. We formally characterize this singularity in the context of adaptive tracking architectures and, for the first time, identify it as a directly exploitable new attack surface. This insight reveals a previously overlooked and highly vulnerable topological path space attack surface. Unlike traditional adversarial attacks that target the output space, this new attack surface allows for the simultaneous manipulation of both the models semantic representation and its inference topology. Based on this, we propose the Adversarial Path-Inversion (API) framework. API generates imperceptible perturbations to precisely manipulate the gating decisions, forcing the inference onto altered computational paths. The severe inconsistency between the original and the inverted paths dismantles the representation capability of the model. Extensive experiments on state-of-the-art adaptive trackers demonstrate that API achieves superior perturbation stealthiness, more effective attack, and faster inference speeds. This work opens a new dimension for the security analysis of dynamic tracking networks and provides a theoretical warning for constructing robust adaptive tracking architectures in the future.

Abstract:
Automatic video dubbing remains limited by two competing constraints: hierarchical methods depend on brittle, multi-stage preprocessing that restricts data scalability and practical deployment, while holistic approaches operating on uncropped video suffer from weak temporal alignment and speaker-utterance ambiguity in multi-speaker settings. To this end, we propose CineDub, a unified diffusion-based model that achieves precise multi-speaker dialogue dubbing directly from uncropped videos, without face cropping or speaker diarization. Central to our approach is the Implicitly-Coupled Holistic Conditioning (ICHC) paradigm, where visual representations and a semantic-bundled transcription format are encoded independently by their respective holistic encoders, yet implicitly coupled through cross-modal training to resolve speaker ambiguity and enable precise multi-speaker multi-turn dialogue dubbing. Building on the unified temporal cues captured by holistic visual features, we further extend CineDub to joint speech-and-audio generation. We introduce an Ambient-to-Linguistic Curriculum Learning (ALC) to mitigate sub-task degradation, and a decoupled textual branch control mechanism to resolve cross-prompt interference during joint generation. Two benchmarks, CineDub-Multi for multi-speaker dialogue dubbing and CineDub-SA for video-to-speech-and-audio (V2SA) generation, are released to enable evaluation under realistic conditions. Experiments show that CineDub achieves state-of-the-art results on single-speaker dubbing and video-to-audio benchmarks while excelling in multi-speaker dubbing and joint speech-and-audio generation.

Abstract:
Recent advances in multi-modal retrieval-augmented generation (mRAG), which augments multi-modal large language models (MLLMs) with external knowledge, have shown that collective intelligence from multiple agents can outperform a single model through effective communication. Despite their strong performance, existing multi-agent systems incur substantial token overhead and computational cost, posing challenges for large-scale deployment. To address these issues, we propose a multi-modal multi-agent hierarchical communication graph pruning framework, termed M3Prune. M3Prune eliminates redundant communication edges both across and within modalities, improving the trade-off between task performance and token overhead. Specifically, M3Prune first performs intra-modal graph sparsification in the textual and visual modalities to identify task-critical communication links. It then constructs an inter-modal communication graph and sparsifies cross-modal connections while encouraging consistent cross-modal reasoning through a modality alignment score. Finally, it progressively prunes redundant edges to obtain an efficient hierarchical topology. Extensive experiments on both general-domain and domain-specific mRAG benchmarks show that M^3Prune consistently outperforms single-agent and strong multi-agent mRAG systems while significantly improving token efficiency.

Abstract:
Humans often identify subtle objects not just by searching for them, but by systematically ruling out irrelevant areas and focus only on regions with high probability. However, current Vision-Language Models (VLMs) mostly rely on ''positive-only'' attention that identifies important regions directly, leading to hallucination in complex, high-resolution scenes. We introduce Veto, an inference-time scaling agentic framework that mimics the human ''elimination process'' for fine-grained perception. Veto's core innovation is Negative Filtering, a strategy that progressively filters distracting regions to simplify the visual field. By leveraging a reasoning-driven scoring mechanism, Veto dynamically identifies and discards ''negative regions'' that do not contribute to the task, effectively enabling the well-designed Verifiable Visual Grounding stage to locate and validate intermediate predictions based on the processed visual evidence. Extensive experiments demonstrate that Veto gains leading performance against training-free methods across multiple high-resolution multimodal benchmarks. Notably, it achieves a substantial average performance gain of 7.5 percent on the Qwen2.5-VL-7B model, while consistently improving performance on general VLM and hallucination-related benchmarks. Furthermore, extensive ablation studies and cross-model evaluations confirm Vetos robust effectiveness and seamless transferability across various base architectures with native grounding capabilities.

Abstract:
3D building roof reconstruction remains a challenging task in urban modeling, particularly when processing rough roof models containing noise and structural imperfections. While existing methods have achieved significant progress, they often fail to maintain reconstruction quality when handling such inputs. This paper presents a unified reconstruction pipeline that transforms rough structured roof models into high-quality reconstructions by integrating 2D sketch optimization. Our approach begins by applying statistical priors to segment the input rough roof model into planar primitives with constrained orientations. We then extract 2D roof sketches and establish precise correspondences between sketch elements and planar segments. A structure-aware optimization process enforces geometric constraints to refine these sketches. The optimized parametric sketches guide the correction of planar primitives, which are subsequently fused into an enhanced 3D roof surface representation. Experimental evaluations demonstrate that our method outperforms current baseline approaches, achieving superior reconstruction accuracy while effectively handling noise and structural irregularities inherent in rough roof models. The proposed pipeline proves particularly effective in recovering sharp features and maintaining surface regularity from challenging input data.

Abstract:
Multimodal Large Language Models (MLLMs) have demonstrated remarkable capabilities on complex tasks through Chain-of-Thought (CoT) reasoning, yet at the cost of substantially increased inference overhead. While post-training quantization (PTQ) has emerged as a prevalent approach for reducing deployment costs, its impact on the reasoning capability of MLLMs remains largely unexplored. This paper presents a comprehensive empirical study on the reasoning capability of quantized MLLMs. Employing two representative weight-only quantization methods, AWQ and GPTQ, we systematically evaluate 10 models spanning the Qwen2.5-VL, Qwen3-VL, and InternVL3 families across 20 multimodal benchmarks, and analyze the performance--efficiency trade-offs introduced by quantization. We find that quantization incurs significantly greater degradation on multimodal reasoning tasks than on perception tasks, with the disparity becoming especially pronounced in long-chain reasoning scenarios. Mechanistic analysis attributes this degradation primarily to error accumulation during long-sequence decoding, reasoning trajectory drift, and systematic shifts in output patterns. Despite these adverse effects, quantization still yields substantial end-to-end speedups in reasoning-heavy, long-output scenarios. Finally, we empirically analyze two training-free factors that influence quantization performance, namely calibration set design and selective layer preservation, and discuss their implications for future multimodal quantization research.

Abstract:
Few-shot image generation aims to adapt a source generative model to a target domain using minimal training images. Most existing approaches introduce consistency constraints, such as instance-level or distribution-level loss functions, to directly align the distribution patterns of source and target domains within their latent spaces. However, these strategies often fall short. Overly strict constraints amplify the domain gap and cause distorted content, while overly relaxed constraints fail to leverage the source domain effectively. This limitation stems from inherent discrepancies in underlying distribution structures, compounded by the scarcity of target samples hindering accurate estimation. To address this, we propose Equivariant Feature Rotation (EFR), a novel strategy that aligns source and target domains within a self-rotated proxy feature space. We perform adaptive rotations within a parameterized Lie Group to transform features into an equivariant proxy space. These learnable rotation matrices bridge the domain gap by preserving intra-domain structural information without distortion, facilitating effective knowledge transfer. Comprehensive experiments on various datasets demonstrate that EFR significantly enhances generative performance in the target domain.

Abstract:
Open-World 3D Detection is essential for real-world applications such as autonomous driving, where perception systems typically need to accurately recognize and precisely localize objects beyond predefined categories. Although recent advances in vision-language models have significantly improved the semantic recognition of unseen objects, existing OW-3D methods still largely rely on heuristic algorithms for novel-object localization, resulting in low-quality 3D candidates and becoming a key bottleneck for downstream detection performance and practical deployment. To address this issue, we propose Diffuse-OW3D, a framework for improving candidate geometric quality in open-world 3D detection. Diffuse-OW3D consists of two key components: a Cross-Modality Validated Candidate Generator, which exploits geometric consistency between LiDAR and image cues to produce more reliable candidates, and an Annotated-Class Assisted Denoiser, which progressively refines these candidates through diffusion-based denoising. By further incorporating scene-level objectness priors and scale-aware geometric conditioning, the ACA Denoiser effectively corrects the systematic errors introduced by heuristic box construction. Extensive experiments demonstrate that Diffuse-OW3D produces more accurately localized and geometrically higher-quality novel-category candidates, and consistently improves downstream open-world 3D detection performance over previous state-of-the-art methods.

Abstract:
Weakly-supervised Whole Slide Image (WSI) classification requires aggregating thousands of patch embeddings into a slide-level prediction under bag-level supervision only. However, existing Multiple Instance Learning (MIL) methods operate in Euclidean feature spaces and overlook the directional geometry of patch descriptors, which naturally lie on a unit hypersphere after L_2-normalization. We propose GHH-MIL, which reformulates patch aggregation as a spherical harmonic expansion on the hypersphere, where the completeness of harmonics in L^2(\mathbbS^p-1) guarantees each impulse-measure patch an exact spectral representation. We prove first-degree harmonic bases are exactly equivalent to normalized Cartesian coordinates, resolving the numerical explosion and reducing spectral extraction to closed-form linear scaling with only p+1 multiplications per patch. Multi-head gated attention aggregates representations in spectral space, where sparsity is enforced by a learnable spectral filter with \tanh-constrained weights and softmax attention; the model is trained end-to-end with focal loss and label smoothing. Experiments on CAMELYON16, TCGA-NSCLC, and MOC demonstrate that GHH-MIL outperforms existing MIL methods, introducing a new paradigm grounded in hyperspherical harmonic analysis.

Abstract:
Specialized detectors for AI-generated images (AIGI) often achieve near-perfect accuracy on curated benchmarks, yet their performance degrades substantially in realistic, in-the-wild scenarios. In this work, we show that frozen features from modern Vision Foundation Models (VFMs), combined with a lightweight classifier, form a remarkably strong baseline for generalizable AIGI detection. Using representative modern encoders, including Perception Encoder, MetaCLIP 2, and DINOv3, we conduct a comprehensive evaluation across standard benchmarks, recent unseen generators, and challenging in-the-wild distributions. Across these settings, this simple baseline consistently matches or outperforms recent specialized detectors, with particularly large gains in realistic scenarios. We further investigate why this simple setup is so effective. Our analyses provide converging evidence that the strong forensic separability of modern VFMs is closely related to their exposure to synthetic web content during pre-training. In Vision-Language Models, this manifests as semantic alignment with forgery-related concepts, while in Self-Supervised Learning models it appears as implicit discrimination of generative distributions. Although a fully controlled pre-training study is beyond the scope of this work, multiple complementary analyses support this interpretation. We also identify important limitations. While modern VFMs are highly effective for global AIGI detection, they remain vulnerable to severe transmission degradation and perform poorly on pure VAE reconstruction and localized editing. Overall, our results suggest that progress in generalizable AIGI detection may depend more on preserving and leveraging strong pretrained representations than on increasingly complex task-specific forensic designs. Our code and pre-trained models are publicly available at ModelScopeand HuggingFace.

Abstract:
Pavement condition assessment is essential for road safety and maintenance. Existing research has made significant progress. However, most studies focus on conventional computer vision tasks such as classification, detection, and segmentation. In real-world applications, pavement inspection requires more than visual recognition. It also requires quantitative analysis, explanation, and interactive decision support. Current datasets are limited. They focus on unimodal perception. They lack support for multi-turn interaction and fact-grounded reasoning. They also do not connect perception with vision-language analysis. To address these limitations, we introduce PaveBench, a large-scale benchmark for pavement distress perception and interactive vision-language analysis on real-world highway inspection images. PaveBench supports four core tasks: classification, object detection, semantic segmentation, and vision-language question answering. It provides unified task definitions and evaluation protocols. On the visual side, PaveBench provides large-scale annotations and includes a curated hard-distractor subset for robustness evaluation. It contains a large collection of real-world pavement images. On the multimodal side, we introduce PaveVQA, a real-image question answering (QA) dataset that supports single-turn, multi-turn, and expert-corrected interactions. It covers recognition, localization, quantitative estimation, and maintenance reasoning. We evaluate several state-of-the-art methods and provide a detailed analysis. We also present a simple and effective agent-augmented visual question answering framework that integrates domain-specific models as tools alongside vision-language models. The dataset is available at: \urlhttps://huggingface.co/datasets/MML-Group/PaveBench.

Abstract:
Visual token compression is essential for improving the inference efficiency of multimodal large language models (MLLMs). However, existing training-free methods suffer from attention drift and degrade significantly at low retention rates. Training-based methods lack semantic guidance and incur high training costs, often requiring large-scale data to achieve generalization. To address these issues, we propose GAZE, a grounding-aware end-to-end token pruning framework. Specifically, GAZE injects foregroundbackground priors into a lightweight scorer via a localization-aware ranking loss, while employing dynamic budget allocation to enhance generalization. Furthermore, GAZE features a FocalSpatial dual-channel selection mechanism: the Focal channel focuses on semantically salient tokens, whereas the Spatial channel ensures structural coverage via grid-based priors. Both channels share a redundancy penalty to balance discriminativeness and diversity. Experiments show that GAZE requires only 70 minutes of training on a single RTX 4090, yet delivers competitive or superior performance across multiple benchmarks. Notably, GAZE retains 98% of full-token baseline accuracy at 30% visual-token retention and achieves up to 1.8× prefill speedup at 10% retention.

Abstract:
Large Vision-Language Models (LVLMs) achieve strong performance on many multimodal tasks, but object hallucinations severely undermine their reliability. Most existing studies focus on the text modality, attributing hallucinations to overly strong language priors and insufficient visual grounding. In contrast, we observe that abnormal attention patterns within the visual modality can also give rise to hallucinated objects. Building on this observation, we propose Segmentation-based Attention Entropy (SAE), which leverages semantic segmentation to quantify visual attention uncertainty in a semantic space. Based on SAE, we further design a reliability score for hallucination detection and an SAE-guided attention adjustment method that modifies visual attention at inference time to mitigate hallucinations. We evaluate our approach on public benchmarks and in real embodied multimodal scenarios with quadruped robots. Experiments show that SAE reduces hallucinations without additional training, improving LVLM reliability.

Abstract:
Visual Question Answering (VQA) models often suffer from severe generalization degradation under distribution shifts, as they excessively rely on statistical shortcuts while ignoring genuine causal evidence. Existing debiasing methods face two major limitations. First, in feature modeling, they frequently overlook the complex spatial topological dependencies between cross-modal entities. Second, regarding intervention mechanisms, they predominantly rely on manually predefined heuristic strategies, lacking the capability to automatically construct diverse interventional environments in an end-to-end manner. To overcome these limitations, we propose a novel framework termed Automatic Environment Construction and Invariant Learning (AECIL). Specifically, we design a Dynamic Causal Masking (DCM) mechanism that first captures the geometric and semantic correlations of multimodal entities via a structured graph network. It then employs differentiable discrete sampling to adaptively decouple features into causal features and spurious features, thereby automatically generating virtual interventional environments reflecting diverse distribution shifts. Building upon this, we further introduce a Joint Invariant Optimization (JIO) strategy, which integrates Counterfactual Consistency (CFC) and Invariant Risk Minimization (IRM) to compel the model to converge toward a robust cross-distribution causal mapping across varying interventional environments. Extensive experiments on mainstream datasets, including VQA-CP v2 and VQA v2, demonstrate that AECIL not only successfully overcomes the inherent in-distribution (ID) performance degradation bottleneck, but also achieves superior out-of-distribution (OOD) generalization performance compared to the benchmark debiasing models.

Abstract:
With the rapid advancement of Multimodal Large Language Models (MLLMs), the redundancy of visual tokens in images has become a critical bottleneck constraining inference efficiency. Existing resampling methods often suffer from detail loss and fail to dynamically adapt to user instructions. In this paper, we propose SG-MoE, a novel architecture designed for extreme vision token compression with robust cross-modal alignment. Specifically, we incorporate a semantic guidance mechanism that utilizes textual instructions to highlight relevant visual information. Subsequently, a heterogeneous Mixture-of-Experts (MoE) framework is employed to enrich multi-scale features and adaptively compress tokens based on image content. Furthermore, to compensate for potential loss during aggressive compression, we design a semantic-driven detail recovery strategy that restores fine-grained details risk being discarded. Extensive experiments demonstrate that our method achieves an aggressive compression ratio of \approx89% while maintaining 99.65% of the original performance across 9 benchmarks, even surpassing the uncompressed baseline on 5. This architecture offers an efficient and robust paradigm for constructing high-performance MLLMs.

Abstract:
Archaeological excavation reportscombining artifact illustrations, stratigraphic descriptions, and typological catalogsremain largely inaccessible to computational analysis, constraining the scale and reproducibility of cultural heritage research. Transforming these heterogeneous, unstructured documents into queryable databases requires solving two interrelated challenges: extracting structured records from complex multimodal layouts, and verifying extraction quality at corpus scale without exhaustive manual review. We present ArchFact, an end-to-end system integrating established computer vision and natural language processing techniques into a unified pipeline for archaeological report digitization. The extraction stage combines domain-adapted object detection, cross-modal matching, and multi-agent LLM frameworks to handle fragmented cross-page text and spatially separated figure-caption pairs. The verification stage introduces an assertion-driven quality control mechanism where evaluation criteria are iteratively refined through human-AI collaboration. Validated across Chinese and English archaeological corpora, the full system reduces per-catalog processing time by approximately 95% and expert verification workload by 92% compared to fully manual workflows.

Abstract:
Few-shot adaptation of vision-language models (VLMs) aims to surpass zero-shot performance with minimal labels while preserving out-of-distribution (OOD) robustness. However, existing methods share a critical yet overlooked assumption: that all feature directions can be corrected uniformly using statistics estimated from a tiny support set. This ignores feature prior asymmetry: directions aligned with the frozen text space inherit a strong semantic prior from large-scale pretraining and are inherently trustworthy, while residual directions must be estimated from noisy few-shot examples with no equivalent prior. Applying uniform corrections risks contaminating this reliable prior with noisy estimates. This insight aligns with the classic Wiener filtering framework, which optimally balances signal preservation and noise suppression based on directional reliability. To this end, we propose WienerCLIP for few-shot adaptation, a Wiener-inspired spectral filter that assigns adaptive gains based on directional priors. WienerCLIP constructs a semantic anchor subspace from frozen text weightsdirections here receive the Wiener-limit gain of 1, reflecting their strong pretraining prior. For residual directions, we estimate signal-to-noise ratios from the support set and apply a calibrated Wiener gain, which quantifies support-anchor alignment and acts as a prior confidence parameter. This yields stronger shrinkage when the prior is well-aligned with the support and gracefully relaxes to identity when alignment is low. The operator is closed-form, training-free, and plug-and-play with any adaptation head. Extensive experiments on 5 baselines, 11 in-domain, and 4 OOD benchmarks demonstrate consistent gains in the few-shot regime while preserving zero-shot OOD robustness with <0.1% runtime overhead.

Abstract:
Unmanned Aerial Vehicle (UAV) cross-view geo-localization, which matches UAV queries to the correct satellite references, is a core task in remote sensing and autonomous navigation. Most existing methods use only visual information and ignore the additional information available in text, including scene semantics and UAV-side metadata such as flight altitude and viewing angle. Their contrastive learning pipelines also rely on random sampling, where most negative pairs are easy and provide weak training signals. To address both limitations, we propose the Dynamic Hard Sample Mining and Multimodal (DHSMM) framework. DHSMM first extends SUES-200 and University-1652 with dense image captions that describe scene content, while UAV captions also include metadata such as flight altitude and viewing angle. Then, to fuse semantic information, DHSMM uses a pipeline with Stage 1 and Stage 2. Stage 1 learns strong visual representations, and Stage 2 fuses textual information through a text encoder and a fusion module to better distinguish visually similar locations. To improve weak training signals, DHSMM also includes a dynamic hard sample mining strategy that actively constructs batches of mutually confusing negatives, which improves the quality of contrastive training signals and boosts retrieval accuracy. Experiments show that DHSMM achieves leading bidirectional retrieval performance, and the ablation studies further confirm the value of both multimodal fusion and dynamic hard sample mining.

Abstract:
Recent generative models can synthesise piano audio that is musically plausible and perceptually convincing, yet generating performances that are perceived as truly \emphlive-like remains challenging. We argue that live-like piano generation is governed by two coupled dimensions: \emphperformance-side realism, which reflects the plausibility of expressive interpretation, and \emphaudio-side realism, which reflects the acoustic credibility of the produced sonic event. From this perspective, we propose \textscPRESONA, a phrase-level framework that factorises live-like piano generation into controllable symbolic performance generation and audio-guided realism alignment with explicit control preservation. Using aligned symbolic performances and recordings from MAESTRO, \textscPRESONA models expressive behaviour through semantically grounded controls and carries these behaviours into the final audio domain within a unified generation framework. We also develop a multi-level evaluation setup aligned with the proposed factorisation, including reference-space analysis, control validation, system-level comparison, subjective listening, and targeted ablations. Taken together, these results provide support for the proposed factorised view and indicate that it offers a practical way to study the relation between expressive controllability and audio-side realism in live-like piano generation.

Abstract:
Sign Language Production (SLP) represents a formidable cross-modal challenge, requiring the translation of spoken language into expressive sign motions, a process primarily driven by gloss-to-pose (G2P) generation. However, current G2P approaches are fundamentally bottlenecked by conventional self-attention mechanisms. The dense correlation matrices in these architectures predominantly attend to irrelevant regions, introducing severe computational redundancy and compromising generation accuracy. To overcome these structural limitations, we propose SRPro, a uniquely designed diffusion-based SLP framework featuring Semantic-aware Routing Attention (SRA) and Prototype Anchored Alignment (PAA). Specifically, SRA effectively eliminates non-informative queries by dynamically routing visual tokens into discrete, semantically aligned subspaces, confining attention exclusively to coherent regions. Leveraging these precise routing outcomes, PAA aggregates visual tokens into semantic-level representations and aligns them with textual gloss via learnable gloss-wise prototypes. Serving as stable semantic benchmarks, these prototypes enable robust cross-modal and cross-batch alignment, seamlessly bypassing the training instability commonly associated with diffusion models. Remarkably, SRPro achieves these gains without any extra inference parameters or computational overhead, consistently improving all evaluation metrics. Extensive experiments on PHOENIX14T and USTC-CSL demonstrate that SRPro establishes superior SOTA performance.

Abstract:
The effectiveness of edge-cloud collaboration for GUI grounding depends on autonomous requesting, where the edge agent selectively offloads complex tasks to the powerful cloud. However, in visually dense scenarios, lightweight edge agents often exhibit overconfident hallucinations, leading to a misalignment between confidence and accuracy that hinders reliable autonomous requesting. To address this, we leverage the observation that an agent's cognitive instability leads to significant latent drift under minute perturbations due to steep decision boundaries. We propose SPECTRA, a lightweight autonomous request framework for edge-cloud GUI grounding, comprising (1) \underlineSaliency-guided Targeted \underlinePerturbation and (2) \underlineEfficient \underlineCognitive \underlineTRajectory \underlineAnalysis. SPECTRA conducts a visual cognitive stress test by injecting masks into critical visual anchors and quantifies the topological divergence of the agent's high-dimensional cognitive trajectories during the prefill phase, avoiding inefficient output decoding. Experiments demonstrate that SPECTRA performs cloud request assessment without autoregressive decoding. Our GTA1-32B+InfiGUI-G1-3B and GTA1-32B+Holo1.5-3B maintain 93.44% and 95.60% of cloud-only performance with average request rates of 37.58% and 39.24%, respectively.

Abstract:
Self-supervised foundation models have recently shown strong potential for electroencephalogram (EEG)-based analysis. However, existing approaches struggle to capture the inherently multi-scale temporal structure of EEG signals, where local neural patterns and long-range dependencies jointly encode task-relevant information. This limitation hampers cross-scale representation learning and generalization across diverse downstream tasks. To address this challenge, we propose MSBraM, a Multi-Scale self-supervised Brain foundation Model designed to learn hierarchical EEG representations. MSBraM follows a two-stage pretraining framework. First, a multi-scale neural tokenizer discretizes raw EEG signals into semantic codes at different temporal resolutions via vector-quantized reconstruction. Second, the model is pretrained to predict masked codes using a curriculum multi-scale masking strategy, progressively integrating fine-grained local patterns with global temporal context. We pretrain MSBraM on over 2,400 hours of EEG data and evaluate it across 10 downstream tasks on 12 public datasets. Extensive experiments show that MSBraM achieves superior performance on other state-of-the-art pretrained models, demonstrating strong generalization and transferability. These results indicate that explicitly modeling multi-scale temporal dynamics is critical for effective EEG foundation models.

Abstract:
Image-based 3D visual grounding is critical for embodied agents, yet existing benchmarks suffer from loose text-observation alignment and neglect temporal ordering. We introduce SeqAlign3DVG, a novel benchmark dedicated to temporally ordered and strictly observation-aligned image-based 3D visual grounding. Unlike prior works using order-agnostic views or global point clouds, SeqAlign3DVG ensures all expressions are human-verified and strictly grounded in the provided RGB observations (single frames or ordered observation sequences). It comprises 9,622 single-view and 14,493 sequence samples featuring rich descriptions, complex relations, and multi-instance ambiguities. To tackle this benchmark, we propose a unified voxel-based pipeline featuring Relevance-Ordered Voxel Memory (ROVM) and Progressive Language-Voxel Fusion (PLVF). ROVM dynamically ranks and aggregates multi-view evidence via a conservative memory to mitigate noisy observations, while PLVF performs coarse-to-fine spatial-linguistic reasoning for precise disambiguation. Our approach achieves state-of-the-art performance under the depth-free protocol, significantly improving localization for targets defined by complex relations and appearance cues.

Abstract:
Multimodal Diffusion Transformers (MM-DiTs) have demonstrated remarkable text-to-image generation performance, surpassing traditional U-Net-based diffusion models. Nevertheless, their powerful generative capabilities also raise significant safety concerns, as they may generate sensitive or inappropriate content. While existing concept erasure methods aim to mitigate such risks, most require modifying model parameters, which are often architecture-specific and impractical for deployed larger models. Several tuning-free approaches face challenges when applied to advanced large-scale MM-DiTs due to their deeply embedded knowledge, broad semantic space, and context-dependent text encoders. To address these challenges, we propose to erase concepts by directly manipulating the model's internal representations. Our key insight, derived from an in?depth analysis of MM?DiTs block?wise generative roles, is that text-conditioned semantic representations are most salient in the middle blocks of MM-DiTs. Based on this, we extract representations of an unwanted concept and a desirable safe one from the middle block, construct a steering vector from their difference, and inject this single vector into consecutive early and middle blocks. By operating exclusively on the sparse text?branch tokens and leveraging the straight sampling trajectory of rectified flow, our method achieves effective concept erasure with negligible overhead and without any training. Extensive experiments across MM-DiT models demonstrate that our method achieves state-of-the-art performance in erasing diverse concepts, enables effective control over the final output, and remains robust to adversarial attacks.

Abstract:
Online understanding of who is talking to the camera wearer is a key capability for egocentric social interaction. However, existing talk-to-me (TTM) studies are commonly formulated as offline clip-level recognition, which is poorly aligned with online interaction and overlooks the diverse non-TTM speaking states that naturally arise in egocentric videos. In this paper, we revisit this problem by reformulating it as an online, frame-level prediction task. Instead of treating TTM as a binary problem against a single negative class, we model it in the presence of diverse and previously underexplored non-TTM states, such as talking-to-others, self-talking, and background conditions. To support this new formulation, we construct an Online TTM Dataset consisting of 406 egocentric video clips with approximately 900K annotated frames, each labeled with frame-level social interaction categories (e.g., background, TTM, talking-to-others, self-talking), by extending the Ego4D social interaction benchmark. In this benchmark, we evaluate five adapted baselines and develop a new model that integrates social cues across modalities. Experimental results show that our multimodal model, which jointly leverages audio, visual, and speech-semantic cues, achieves 75.5% frame-level F1 on TTM, outperforming strong baselines and enabling a systematic analysis of how different speaking states affect TTM recognition.

Abstract:
Cross-view UAV geo-localization, mapping Unmanned Aerial Vehicle (UAV) images to geo-referenced satellite maps, is crucial for autonomous navigation but remains challenging due to severe viewpoint discrepancies, scale variations, and repetitive structures. Most existing methods formulate this as a single-frame retrieval problem, which frequently yields visually plausible but geographically distinct candidates under local appearance ambiguity. While sequence-based filtering approaches attempt to exploit temporal coherence, they often fail under the unstable and noisy observations inherent in cross-view UAV matching. To address these limitations, we propose SMART, a Semantic-Motion Aware Robust Temporal aggregation framework that models the localization process as continuous spatio-temporal probability density estimation. SMART transforms discrete retrieval candidates into a continuous probability field using a Semantic-Spatial Aware State Initialization module to suppress semantically unreliable dense clusters. Then, SMART employs Kinematics-Driven Spatial-Temporal Propagation to align historical fields while explicitly modulating accumulated motion uncertainty. Finally, an Entropy-Guided Dynamic Temporal Aggregation mechanism adaptively weights frames based on their spatial compactness, emphasizing informative frames over ambiguous ones. Extensive experiments on four diverse benchmarks (i.e., SatLoc, VPAir, ALTO, AerialVL) demonstrate that SMART consistently outperforms state-of-the-art single-frame and sequence-based methods.

Abstract:
4D scene reconstruction requires robust handling of dynamic objects, yet existing streaming methods commonly assume static scenes, leading to pose drift and ghosting artifacts over long sequences. We introduce Stream4D, a training-free, motion-decoupled streaming reconstruction framework that can be directly integrated into causal Transformer backbones (e.g., StreamVGGT, Stream3R). Our key observation is that causal attention implicitly encodes rich temporal motion cues within the KV cache. We adaptively aggregate these cues from multiple attention layers and across temporal windows, and fuse them with depth priors to handle challenging cases such as occlusions and textureless regions, producing precise dynamic masks that decouple static and dynamic regions. We then apply a soft key masking scheme that leverages these dynamic masks to suppress interference from dynamic regions in causal attention, thereby improving camera pose estimation accuracy. Stream4D inherits the efficiency of causal attention and remains compatible with existing streaming architectures. On diverse dynamic-scene benchmarks, Stream4D significantly improves camera pose estimation and static-scene reconstruction quality, yields reliable dynamic masks, and effectively mitigates long-sequence ghosting.

Abstract:
Continual Test-Time Adaptation (CTTA) aims to enable deep models to adapt to non-stationary target distributions during inference without access to source data. Existing approaches often depend on fixed learning rates or manually defined thresholds, which can lead to instability or overfitting when target distributions change significantly. In this work, we propose AutoVote, an adaptive learning rate modulation method to address the aforementioned challenge. Specifically, we define parameter sensitivity based on the standard deviation of gradients, and further design the sensitivity reduction ratio (SRR) and volatility score (VS) to capture temporal changes in sensitivity. These components are fused into a correction factor to modulate the learning rate. Besides, to ensure robust and interpretable updates, we design a voting mechanism, which aggregates the judgments of all parameters and leverages normalized consensus to guide further updates. To detect data shift, we reuse the voting mechanism and trigger safe model resets. Extensive experiments on CIFAR10-C, CIFAR100-C, and ImageNet-C demonstrate that our approach achieves state-of-the-art performance under CTTA settings. Moreover, it exhibits strong stability under small-batch conditions and other realistic test-time adaptation scenarios, providing enhanced resilience to catastrophic forgetting.

Abstract:
Explainable aesthetic image cropping requires not only localizing a visually pleasing crop but also explaining why it is preferred. Existing crop-and-explain methods largely treat explanation as post-hoc text generation and overlook composition, a key aesthetic factor that links crop decisions with interpretable reasoning. In this paper, we reformulate explainable aesthetic image cropping as a structured crop-composition-explanation problem. To support this setting, we introduce COMEX, a new benchmark built through image expansion and an IO-reversal pipeline. COMEX contains 33,161 quadruples, each consisting of an expanded image, a crop box, a composition category, and a composition-grounded explanation, enabling joint learning of crop localization, composition understanding, and explanation generation. We further propose a two-stage SFT+GRPO framework, where supervised fine-tuning establishes the structured output protocol and basic cropping ability, and GRPO further improves crop quality, composition prediction, and explanation faithfulness. We benchmark 15 large vision-language models and existing cropping methods on COMEX, establishing a comprehensive testbed for composition-grounded explainable aesthetic cropping. Experiments on both COMEX and prior benchmarks demonstrate the effectiveness and transferability of our framework, with strong performance across evaluation metrics.

Abstract:
Point-of-Interest (POI) recommendation aims to predict users future visit locations and plays a critical role in location-based social networks. Recent works have introduced rich multi-modal POI data to unlock performance potential, yet two key challenges remain insufficiently addressed in real-world scenarios: (1) insufficient handling of missing modality scenarios, where incomplete spatio-temporal or semantic data leads to severe performance degradation; (2) negligence of inherent, unique characteristics of different POI modalities, which hinders the learning of high-quality POI representations. These flaws degrade recommendation accuracy, limit generalizability, and restrict practical applicability. To address these issues, we propose RoPOI, a robust POI recommendation framework via modality disentanglement and missing representation generation. Specifically, we design an information-theoretic modality disentanglement module to decompose each modalitys features into modality-invariant, shared and modality-specific, unique components, and build a fine-grained missing modality generation module to reconstruct missing features by integrating aligned general features from available modalities and implicit user-POI modality preferences. This synergistic design resolves performance degradation caused by missing modalities while preserving discriminative modality-specific information and eliminating redundant noise. Extensive experiments on real-world datasets show that RoPOI consistently outperforms state-of-the-art POI recommendation baselines.

Abstract:
Mixing-based semi-supervised learning has emerged as a strong and increasingly popular paradigm for modern medical image segmentation. It synthesizes intermediate training samples to bridge the empirical distribution gap between labeled and unlabeled data. However, when directly applied to anatomically structured 3D volumes, existing mixing-based approaches still leave a key methodological issue under-addressed: they exchange local regions across different subjects without explicitly ensuring that the mixed result preserves anatomically learnable structure after the exchange. This issue is particularly pronounced in multi-organ CT, where naive coordinate-based mixing can fragment organ continuity, disrupt essential boundary context, and weaken the structural learnability of the mixed sample. In this paper, we propose SLIP, a \underlineStructural \underlineLearning via \underlineImplicit-to-Explicit Alignment \underlineParadigm, which reformulates cross-subject mixing by turning its previously implicit structural and semantic assumptions into explicit alignment constraints. SLIP first improves structural learnability through a registration-assisted slab mixing strategy that establishes coarse anatomical correspondence and preserves axial continuity across subjects, then performs reliability-aware and text-anchored semantic alignment to explicitly aggregate same-class representations while separating different-class ones. Extensive evaluations on the Synapse and AMOS datasets demonstrate that SLIP consistently achieves state-of-the-art performance. Notably, it yields a substantial Dice improvement of up to 7 percentage points on challenging anatomical structures such as the pancreas, underscoring its robustness in delineating difficult structures.

Abstract:
Selecting first-person viewpoints within 3D scenes is important for applications such as virtual tourism, yet manual selection becomes impractical as scenes grow in scale. Existing methods address this task in indoor environments, relying on pre-labeled objects and structured layouts. These methods thus encounter challenges in outdoor landscapes, which often lack semantic annotations, exhibit irregular geometry, and span a much larger viewpoint space. We introduce ScapePhotographer, an efficient framework for in-scene view selection in unlabeled outdoor landscapes. Our approach first leverages a Vision Language Model (VLM) to automatically annotate the unlabeled scene. For each candidate view, we then compute lightweight feature maps based on depth and VLM-generated semantics instead of rendering costly RGB images. Nine photography-inspired metrics evaluate these feature maps directly, effectively filtering for compositionally and aesthetically promising views. A fine-tuned aesthetic assessment module then makes the final rating for this small set of candidates. Experiments on both synthetic and real-scan scenes show our framework outperforms photography-oriented, entropy-based, and reconstruction-oriented baselines, while achieving significant speedup over exhaustive rendering. A user study indicates that views selected by ScapePhotographer yield aesthetic images competitive with those from human experts.

Abstract:
Federated prompt tuning (FPT) enables collaborative adaptation of vision--language models (VLMs) using lightweight prompts. Existing methods often address heterogeneity and privacy through a split-prompt design under local differential privacy (DP), combining a shared prompt for global transfer with private prompts for local adaptation. However, a single shared prompt may over-smooth diverse transferable knowledge, weakening the balance between personalization and generalization. Multi-expert prompts (MEPs) can better capture this diversity, but enlarge the communicated space, increasing DP noise and communication cost while making robust expert composition more difficult. We propose FedSEPT, a privacy-preserving Federated Subspace-decomposed Expert Prompt Tuning. Specifically, we employ Subspace-decomposed Expert Modeling (SEM) to parameterize multiple prompt experts with shared low-rank factors, a fixed public basis, and private residuals, thereby confining communication and DP perturbation to a compact factor space while enabling direct server aggregation in a common coordinate system. We further design Instance-aware Expert Fusion (IEF), which adaptively combines semantically complementary experts via on-device routing and performs efficient logit-level fusion using cached expert-specific text features. Extensive experiments on 11 heterogeneous benchmarks show that, under the same privacy constraints, FedSEPT achieves a better trade-off between local adaptation and global generalization than strong baselines.

Abstract:
Federated Prompt Learning (FPL) is a communication-efficient paradigm for collaboratively fine-tuning Vision-Language Models (VLMs). However, FPL suffers from performance degradation under data heterogeneity. While existing methods primarily improve personalization capability, the aggregated global model still suffers from degradation due to imbalanced local knowledge. To tackle this challenge, we propose Federated prompt learning via Text-Guided Augmentation (FedTGA), a novel framework that leverages the shared text modality of VLMs as a semantic bridge for privacy-aware cross-client knowledge sharing. Through Text-bridged Prototype Estimation (TPE) and Attribute-based Diversity Sharing (ADS), FedTGA enables both inter-class knowledge complementation and fine-grained diversity sharing. Specifically, the TPE module utilizes cross-modal structural similarity to collaboratively estimate visual prototypes of unfamiliar classes without exposing local prototypes, and ADS leverages textual attributes to aggregate and transfer fine-grained visual variations across clients to enhance intra-class diversity. Extensive experiments demonstrate that FedTGA serves as an effective and orthogonal module, consistently and significantly boosting the global performance of various FPL methods across diverse non-IID settings.

Abstract:
Diagnostic reports provide rich semantic cues that can be leveraged to substantially enhance medical image segmentation. Existing methods typically adopt a one-to-one image-text matching paradigm, while in clinical practice, multiple images may share similar or even identical reports, forming an inherent one-to-many relationship, which limits the ability to capture shared semantics and leads to suboptimal lesion localization. To this end, we propose a novel multi-modal medical image segmentation framework (O2MSNet) that explicitly models the one-to-many correspondence between images and diagnostic reports. Specifically, we propose a text-guided IoU-weighted one-to-many cross-modal contrastive learning strategy to associate images sharing the same report at the image level while enforcing keyword-guided pixel-level alignment, thereby overcoming the limitation of one-to-one matching and improving fine-grained semantic consistency. We further introduce a Keyword Reconstruction Constraint to reconstruct masked diagnostic entities, enabling explicit modeling of key semantics (e.g., lesion quantity and location) and alleviating the lack of entity-level supervision. Moreover, we design a Progressive Hierarchical Cross-modal Interaction module to progressively fuse multi-scale visual features with textual semantics, enhancing cross-modal interaction and representation learning. Extensive experiments on QaTa-COV19 and MosMedData+ demonstrate that our method outperforms state-of-the-art approaches.

Abstract:
\beginabstract While Multimodal Large Language Models (MLLMs) have demonstrated exceptional general reasoning, they often fail in personalized scenarios due to three fundamental challenges: (i) data noise and heterogeneity in raw user traces, (ii) the entanglement of soft aesthetic preferences with hard objective constraints, and (iii) the inherent difficulty of high-dimensional subjective alignment. To address these issues, we propose PersonaRL, a unified reinforcement learning framework for subjective multimodal personalization. We first design a Multi-Agent Data Process Pipeline that employs specialized agents for preference analysis and query rewriting, systematically transforming scattered user inputs into standardized training signals. To ensure precise alignment, we introduce a Structured User Profile Constructor that leverages task-adaptive templates to partition user context into a multi-granularity representation, effectively integrating explicit attributes with implicit preference. Building on this, we develop a specialized Unified Reward Model to capture multifaceted user tastes and employ GRPO to fine-tune the policy model. This enables the policy to align with complex user preferences through online reward-guided optimization rather than static offline preference learning. To rigorously evaluate our framework, we establish a Unified Personalized Evaluation and construct Multi-task PersonaBench, a benchmark covering open-set personalized outfit recommendation, personalized conversation, and image evaluation. Extensive experiments show that PersonaRL significantly outperforms strong baselines across all tasks. \endabstract

Abstract:
Multi-modal image fusion aims to integrate complementary information from different sensors while preserving structural fidelity and semantic consistency. Existing methods are largely domain-specific and generalize poorly to unseen modality pairs because transferable fusion cues remain entangled with domain-dependent appearance statistics, requiring task-specific retraining for each new scenario. To address this, we propose FuseAnywhere, a "Train Once, Fuse Anywhere" framework trained only on infrared-visible data that generalizes zero-shot to unseen medical, multi-focus, and multi-exposure fusion tasks. The core idea is to separate shared cross-domain information from modality-specific content and progressively reorganize visual features under structured semantic guidance. To achieve this, FuseAnywhere (1) preserves modality discrepancies to stabilize shared representation learning, (2) restructures holistic pre-trained vision foundation model priors into disentangled shared and modality-specific guidance, (3) reorganizes visual features through semantic-guided interaction, and (4) enhances reconstruction with a frequency-aware fusion stage to improve structural robustness and high-frequency detail preservation. Extensive experiments demonstrate strong in-domain performance and robust zero-shot cross-domain generalization. Even when trained on a single infrared-visible source domain, FuseAnywhere achieves average relative improvements of 6.74% on three in-domain benchmarks and 4.28% on five cross-domain benchmarks.

Abstract:
Knowledge graph completion (KGC) aims to infer missing entities or relations from incomplete graph structures, and has evolved into multimodal knowledge graph completion (MMKGC), where entities are associated with multiple modalities such as text and images. Traditional representation learning approaches follow the embedding-based paradigm and may struggle when relation-specific evidence is limited. Meanwhile, LLM-based reasoning methods typically linearize graph structures into textual prompts, which obscures structural topology and neglects vital visual information. While vision-language models (VLMs) excel at multimodal reasoning, they cannot natively interpret structured graph topology, particularly when it comes to knowledge graphs where nodes and edges carry complex semantics. To bridge this gap, we propose ViSR-KGC, a visual subgraph reasoning approach for KGC. It integrates three complementary capabilities to capture semantic correlations: identifying global topology dependencies via representation learning, analyzing local multimodal evidence using VLMs, and providing necessary commonsense knowledge inherent in pre-trained models. Based on learned multimodal embeddings, our framework first extracts a compact and query-aware subgraph from the MMKG. Then, this subgraph is transformed into a visually interpretable image using a layout strategy selected through empirical comparison. Finally, the visualized subgraph, entity images, textual descriptions, and candidate answers are combined into a unified prompt, enabling the VLM to infer the missing entity. Experimental results on two real-world MMKG datasets demonstrate that ViSR-KGC consistently outperforms traditional embedding-based models and state-of-the-art multimodal baselines. This highlights the unique advantages of jointly leveraging VLMs' visual perception and internal knowledge for structured reasoning, thereby paving the way toward deeper investigation into VLMs' capacity for semantic graph analysis. The full version including Supplementary Materials is available at \urlhttps://arxiv.org/abs/2608.05833.

Abstract:
Large Multimodal Reasoning Models (LMRMs) have recently shown strong capabilities in complex visual understanding and multi-step inference. However, they remain highly vulnerable to multimodal hallucinations over extended reasoning trajectories. In this work, we study this failure as a state-dynamics problem: as reasoning unfolds, hidden states become progressively less grounded in visual evidence and increasingly dominated by parametric priors retrieved through Feed-Forward Networks (FFNs). Our analysis reveals that this visually ungrounded drift accumulates over long reasoning chains and shows consistent association with FFN residual updates. Based on this view, we propose Adaptive Fine-grained Residual Gating (AFRG), a training-free controller that regulates harmful FFN updates during inference. Specifically, AFRG measures how each FFN residual update changes alignment with the most relevant visual tokens and uses an adaptive bounded gating function to suppress updates that pull hidden states away from the image-supported evidence. Across four hallucination benchmarks and three general multimodal benchmarks on Qwen3-VL and InternVL3.5, AFRG reduces MMHal hallucination rate from 0.34 to 0.26, improves CHAIR_I by 3.8 points, and raises MME Total by 38.3 points while preserving general multimodal capability. These results highlight FFN-level prior regulation as a promising inference-time direction for faithful multimodal reasoning.

Abstract:
The rapid deployment of low Earth orbit (LEO) satellite constellations has greatly expanded Earth observation (EO). But the massive volume of EO data and stringent timeliness requirements remain major bottlenecks for latency-sensitive EO applications such as fire detection and flood monitoring. Existing studies improve LEO data delivery through compression, forwarding, and onboard processing, yet still treat imagery largely as homogeneous traffic, limiting their ability to prioritize image content that contributes most to downstream EO analytics performance. A key challenge lies in the fact that content contribution is application-dependent and not directly observable, yet must be used to guide joint decisions on quality, forwarding, and compression under dynamic LEO connectivity. To address this challenge, we present PrismSpace, an end-to-end adaptive satellite imagery delivery framework that explicitly incorporates downstream EO application performance into the delivery pipeline. PrismSpace estimates cross-application image contribution for black-box EO applications, uses this signal to jointly adapt image quality and forwarding under dynamic LEO connectivity while suppressing redundant regional deliveries, and preserves informative content through a semantic-adaptive codec. Experiments on our prototype show that PrismSpace consistently outperforms baselines in both delivery timeliness and downstream EO analytics performance across diverse network settings.

Abstract:
Traditional Text-based Person Search (TPS) is typically limited to matching static appearance attributes, severely neglecting dynamic action information. The Text-based Person Anomaly Search (TPAS) task bridges this gap, requiring models to locate micro-level specific abnormal behaviors while matching macro-level appearance of pedestrians. However, current TPAS methods face two limitations: external explicit pose estimators are fragile in unconstrained surveillance scenarios, and implicit learning encounters visual decoupling failure under pixel-level entanglement, causing dominant appearance information to easily swallow and contaminate subtle action features. Moreover, performing contrastive optimization on hard negative samples in conventional Euclidean spaces induces severe shortcut learning. To address these, we propose the Lightweight Action Inversion and Riemannian rectification network (LightAIR). First, it introduces textual semantic priors as anchors via a lightweight action inversion operator to extract pure action features, thereby overcoming visual-inherent coupling. Then, it employs orthogonal null-space projection to constrain appearance features within the orthogonal complement space of action features, guaranteeing strict forward decoupling. Finally, we designed a gradient rectification module that computes Riemannian gradient to constrain the backpropagation trajectory, forcing the gradient flow to update along tangent space that preserves decoupling properties. Extensive experiments on both TPAS and TIPR datasets demonstrate that LightAIR outperforms existing SOTA.

Abstract:
Diffusion-based 3D hand pose estimation methods have achieved considerable success, yet they model kinematic structure agnostic to joint-specific uncertainty caused by varying occlusion severity, failing to selectively leverage reliable joints while allowing errors to spread. To overcome this, we propose a novel two-stage framework, FieldDiff, that integrates uncertainty-aware voting with hierarchical diffusion refinement. The key insight is to explicitly quantify joint-wise uncertainty and leverage it as a guidance signal to control information flow across the kinematic structure. To this end, AUVF (Anisotropic Uncertainty Voting Field) first learns an implicit voting field where each point predicts per-joint offsets with axis-wise anisotropic uncertainties, and aggregating votes yields an initial pose alongside joint-level uncertainties that reflect occlusion severity. Subsequently, UGDR (Uncertainty-Guided Denoising Refinement) uses these uncertainties to modulate inter-joint message passing in a hierarchical uncertainty-guided denoising process. At the coarse level, average uncertainties determine reliable anchors to guide holistic structural updates. At the refinement level, axis-wise uncertainties independently control information exchange along each coordinate dimension. Finally, a smoothing level enforces kinematic consistency through skeleton-based constraints. By selectively amplifying contributions from low-uncertainty joints while attenuating those from high-uncertainty ones, this hierarchy progressively refines the pose while suppressing error propagation. Extensive experiments on three challenging benchmarks demonstrate state-of-the-art performance.

Abstract:
Error detection in egocentric procedural videos is critical for ensuring consistent outcomes without deviations in AR assistants and embodied robotics. While existing methods learn static or dynamic step-specific prototypes from normal executions, they remain limited in detecting diverse, unforeseen errors in open-world scenarios. We revisit the characteristics of step errors in procedural activities and find that they can be primarily categorized into two types: (i) The segments that are semantically distinct from all the predefined normal steps are likely to indicate errors. (ii) The erroneous executions may also closely resemble normal steps in both activity semantics and motion patterns, yet contain subtle deviations. To comprehensively detect these two types of errors, in this paper, we propose a dual memory-augmented coarse-to-fine prototype learning (DM-CFPL) framework. Specifically, to solve the error (i), we introduce a memory-based coarse-grained prototype learning strategy, which explores cross-modal clues from step-label text descriptions and informative visual contexts. Thus, the learned semantics-enhanced coarse-grained prototypes facilitate the identification of erroneous steps according to prototype-matching degrees during inference. Furthermore, to address the error (ii), a proxy task, namely memory-augmented procedural activity anticipation, is devised to learn fine-grained prototypes that help to well forecast future activities based on historical observations. Then, in the inference phase, the steps that are poorly reconstructed by the fine-grained prototypes will exhibit higher anticipation errors when compared to the expected activities. Extensive experimental results on three public data sets demonstrate that the proposed DM-CFPL framework achieves remarkable advancements over the competing methods.

Abstract:
Prompt inversion, as a typical reverse engineering technique, enables text-to-image (T2I) diffusion models to generate the desired target images without extensive prompt engineering. However, existing prompt inversion methods suffer from significant limitations: (1) gradient-based methods are unstable and uninterpretable, often resulting in generated images with severe artifacts; (2) gradient-free methods yield human-readable prompts but still fail to preserve visual fidelity due to the lack of fine-grained detail alignment. We contend that the limitations stem from treating prompt inversion as a sufficient condition for reverse engineering, ignoring the critical role of the latent noise that encodes structural information. Consequently, we propose Dualin (Dual inversion), a two-stage method that jointly recovers both the semantic prompt and latent noise of the target image. In the first stage, we integrate vision-language model, CLIP and large language model to invert a faithful, human-interpretable hard prompt. In the second stage, unconditional DDIM inversion reconstructs the exact latent noise of the target image, guaranteeing the consistency at the structural information level. Theoretically, we prove that the inverted noise enables flexible image editing without re-optimization. Extensive experiments on diverse datasets demonstrate that Dualin simultaneously generates high-quality inverted prompts and achieves state-of-the-art image fidelity. Additionally, Dualin can establish a robust foundation for the precise and controllable image editing.

Abstract:
Most digital watermarking techniques embed an external payload that is semantically unrelated to the host image. Such a paradigm inevitably introduces perceptible artifacts and remains vulnerable to realistic image manipulations, including geometric transformations (e.g., cropout and resizing) and modern AI-based editing, which can easily distort or remove the embedded message. In this work, we propose GraphMark, a novel content-aligned watermarking framework that constructs a graph-based watermark signal from the semantic structure of the image and explicitly embeds this signal back into the host image. Specifically, GraphMark partitions the image into patches, identifies semantically similar patch pairs, and organizes them into a sparse semantic graph that serves as the watermark signal. The graph is then encoded into the image through a novel Attentive EdgeConv layer, which aggregates features among semantically related nodes while preserving high visual fidelity. During decoding, watermark extraction is formulated as a graph recovery task that predicts the embedded pairwise relationships from the potentially edited image. Unlike fingerprinting methods that directly extract signals from the original image, GraphMark introduces a structured, content-dependent watermark that is only recoverable after the embedding process. By aligning the watermark signal with the image semantics rather than injecting an arbitrary payload, GraphMark achieves a more favorable trade-off between robustness and imperceptibility, significantly outperforming existing watermarking methods.

Abstract:
Animating a single anime illustration into a controllable video is of significant artistic value and remains challenging. Existing image-to-video models typically synthesize motion at the pixel level, lacking an understanding of physical dynamics. Meanwhile, explicit modeling approaches are often limited to the deterministic refinement of simulation results, failing to bridge the gap between physical reality and artistic style. To address this, we propose DynaCartoon, a framework for generating stylized cartoon videos guided by physics-based motion priors. Given a single illustration and user-specified forces, our method first explicitly models the geometric and attribute information of the scene. Coupled with a physics simulator, it derives force-driven motion sequences to serve as 3D dynamic priors. To overcome the visual rigidity of simulation results, we design a sketch-guided video cartoonization model. By introducing the Background Dynamic Enhancement (BDE) and Foreground Deformation Compensation (FDC) modules as two decoupled structural control branches, this model injects a dynamic atmosphere to the background and compensates for the deformations absent in physical solvers, synthesizing the artistic tension characteristic of cartoons. Experimental results demonstrate that compared to existing methods, DynaCartoon achieves significant improvements in motion plausibility and visual quality.

Abstract:
Autism Spectrum Disorder (ASD) is a complex neurodevelopmental disorder for which early and accurate diagnosis is critical to improving long-term developmental outcomes. However, existing ASD recognition methods are often constrained by the scarcity of diagnostic text data, forcing them to rely mainly on visual analysis and limiting their ability to model clinically meaningful semantic reasoning. To address this challenge, we propose UniAR, a unified framework enhanced by multi-granularity prompt learning for robust ASD recognition under heterogeneous data variations. Specifically, UniAR leverages a large multimodal model to generate hierarchical diagnostic descriptions at the word, phrase, and sentence levels, compensating for the lack of paired clinical reports. To align the generated semantics with visual evidence, we further design a Mixture-of-Experts-based Multi-Scale Alignment Module, which dynamically matches vector-quantized visual prototypes with semantic representations at corresponding granularities. Extensive experiments on four benchmarks covering brain MRI and facial expression scenarios show that UniAR consistently outperforms existing state-of-the-art methods, achieving average accuracies of 75.9% on MRI benchmarks and 91.6% on facial benchmarks, while improving average Accuracy on MRI benchmarks by 1.5 percentage points and average Accuracy on facial benchmarks by 1.2 percentage points over baselines.

Abstract:
Multi-modal large language models (MLLMs) integrate heterogeneous inputs to solve tasks that require joint reasoning across modalities, such as multi-modal question answering (MQA). Beyond conventional prompt-based inference, multi-agent frameworks can elicit complementary evidence and aggregate diverse reasoning traces to better address challenging queries. Existing multi-agent methods are largely inspired by individual-level cognitive mechanisms (e.g., fast-slow thinking and metacognition), but often under-exploit social cognitive factors that govern how groups reconcile disagreements. Moreover, while debate-style strategies can improve logical consistency, their final summarization is frequently heuristic, making it prone to overlooking factual details. We propose CRATER, a consensus-driven multi-agent framework for MQA that aligns discordant perspectives through structured interactions among heterogeneous MLLM agents. CRATER proceeds in three stages. First, in Collaborative Multi-Agent Initialization, a discussion group generates candidate opinions from diverse perspectives and modalities to cover a broad space of viewpoints. Second, Multi-Dimensional Scaling (MDS) Projection embeds these opinions into a low-dimensional geometry, facilitating the identification of polarized stances. Third, polarized opinions are reconciled via an iterative Monotonic Concession Re-discussion procedure, which progressively narrows disagreements and establishes consensus. Experiments on public MQA benchmarks spanning general capability evaluation, text recognition, and multidisciplinary domains show that CRATER consistently outperforms strong baselines. Additional analyses validate the contributions of each core component.

Abstract:
With the malicious abuse of face-swapping technology, facial identity information is at serious risk of leakage. Restoring the original facial identity from manipulated images has become a core demand for proactive face protection and a highly challenging research task in the field of multimedia forensics. Most existing related works are limited to the detection or disruption of face-swapping operations, which merely achieve authentication purposes but fail to effectively recover the original facial information. Furthermore, these methods often exhibit insufficient robustness when confronting quality degradation and composite attacks during real-world dissemination. To address the aforementioned issues, this paper proposes a Generative Identity Encoding-based Robust Face Recovery Network, abbreviated as GIER-Net. Departing from the traditional approach of directly embedding identity information in the pixel space, GIER-Net models identity representations in the latent space of an invertible generative framework to mitigate statistical inconsistencies and enable natural and authentic face restoration. It introduces a redundant identity encoding module (RIEM) integrated with facial region priors, which effectively preserves identity information under scenarios of local tampering and composite degradation. Additionally, a structured Refine mechanism composed of Attack-Aware Encoder (AAEnc) and Refine Inverse Module (RIM) is designed to explicitly capture perturbations induced by face-swapping attacks and adaptively correct latent space deviations during the inversion process. Comprehensive experiments conducted on various face-swapping methods and real-world degradation scenarios demonstrate that the proposed GIER-Net consistently outperforms state-of-the-art approaches in terms of high-fidelity face restoration, robustness against complex perturbations, and low statistical detectability. This work thus offers a novel and effective solution for recovering original facial identities from face-swapped images in practical multimedia scenarios, advancing the research frontier of active face protection in multimedia forensics.

Abstract:
While video foundation models excel at single-shot generation, real-world cinematic storytelling inherently relies on complex multi-shot sequencing. Further progress is constrained by the absence of datasets that address three core challenges: authentic narrative logic, spatiotemporal text-video alignment conflicts, and the "copy-paste" dilemma prevalent in Subject-to-Video (S2V) generation. To bridge this gap, we introduce MuSS, a large-scale, dual-track dataset tailored for multi-shot video and S2V generation. Sourced from over 3,000 movies, MuSS explicitly supports both complex montage transitions and subject-centric narratives. To construct this dataset, we pioneer a progressive captioning pipeline that eliminates contextual conflicts by ensuring local shot-level accuracy before enforcing global narrative coherence. Crucially, we implement a cross-shot matching mechanism to fundamentally eradicate the S2V copy-paste shortcut. Alongside the dataset, we propose the Cinematic Narrative Benchmark, featuring a visual-logic-driven paradigm and a novel Anti-Copy-Paste Variance (ACP-Var) metric to rigorously assess continuous storytelling and 3D structural consistency. Extensive experiments demonstrate that while current baselines struggle with continuous narrative logic or degenerate into trivial 2D sticker generators, our MuSS-augmented model achieves state-of-the-art narrative effectiveness and cross-shot identity preservation.

Abstract:
Multi-modal talking avatar synthesis aims to generate realistic talking videos from a reference portrait and speech. Despite rapid progress in diffusion-based methods, existing approaches still struggle to jointly achieve accurate lip articulation, human-preferred motion aesthetics, and efficient inference. We observe that phonetic accuracy and motion aesthetics arise from fundamentally different sources and should be addressed at complementary stages rather than learned implicitly by a single generator. Based on this insight, we propose VibeAvatar, which disentangles these two objectives through a Phonetic Kinematics Adapter (PKA) that converts recognition-oriented speech features into phonetic-kinematic conditions at the conditioning stage, and an Aesthetic Motion Policy (AMP) that optimizes a flow-consistent stochastic sampling policy via Group Relative Policy Optimization (GRPO) at the post-training stage. With a lightweight flow-based motion generator operating in a compact 1D warp-based latent motion space, VibeAvatar achieves state-of-the-art results in articulation, aesthetics, and efficiency on both objective metrics and user studies, while generating a 10-second 512\,px video in under 10 seconds with only ～3\,GB VRAM.

Abstract:
Embodied mobile manipulation requires language, visual observations, three-dimensional scene structure, and action feasibility to be aligned before execution. We study open-vocabulary target grounding with few-shot manipulation in local household workspaces and present an embodied multimodal grounding framework that integrates active multi-view Semantic 3D Gaussian Splatting (Semantic-3DGS), reachability-aware base positioning, and a diffusion-based vision-language-action policy. A task-driven local Semantic-3DGS serves as a shared interface across active sensing, language-conditioned 3D localization, obstacle-aware scene reasoning, base preparation, and semantic conditioning of the action model. To preserve pretrained action priors, the 3D semantic cues are injected only into the late action-expert blocks. In expanded 50-trial real-robot evaluations against representative vision-language-action (VLA) approaches, the full system achieves 60% long-horizon success compared with 40% for PointVLA and 28% for DexVLA, and reaches 74% success in heavily cluttered manipulation compared with 52% for the single-view variant and 46% for PointVLA. It also maintains 75% success under a 75cm height shift and eliminates photo-induced false grasps. These results indicate that explicit, refreshable 3D semantic grounding can improve robustness under clutter, occlusion, viewpoint variation, and embodiment constraints.

Abstract:
Explainable Multimodal Sentiment Analysis (EMSA) is a burgeoning research area aimed at advancing reliable sentiment reasoning and providing explainable predictive rationales. Preliminary progress in EMSA remains hindered by the scarcity of high-quality data. We introduce ERAS, the first Chinese benchmark with concept-level supervision for explainable multimodal sentiment analysis. ERAS contains 3,686 curated video segments annotated with sentiment polarity and intensity. We further annotate each sample with three semantic concepts as process-level supervision for explainable reasoning. Based on ERAS, we propose CG-DAPO, a concept-guided reinforcement learning framework for MLLMs that leverages concept-level rewards to generate evidence-grounded explanations. We further establish a comprehensive benchmark covering MSA methods across both regression and classification tasks, extending to MLLMs under zero-shot inference, supervised fine-tuning, and reinforcement learning settings. Extensive experiments show that CG-DAPO achieves the strongest overall results on ERAS, outperforms state-of-the-art fusion-based baselines, and delivers strong performance against both closed-source and open-source MLLMs. We believe ERAS and CG-DAPO will facilitate future research on explainable multimodal sentiment analysis.

Abstract:
The rapid growth of Artificial Intelligence-generated content (AIGC) is reshaping the landscape of video production and circulation. Unlike traditionally produced videos, AIGC videos often exhibit greater uncertainty in visual details, narrative coherence, and content expression, which may introduce new risks for populations that are particularly sensitive to inappropriate or misleading content, such as children. While extensive research has investigated video safety and violation detection, most existing approaches focus on detecting general violations from an adult-oriented perspective, remaining insufficient for identifying the fine-grained, implicit, and context-dependent risks that children may encounter when viewing AIGC videos. This work focuses on child-oriented AIGC video reviewing. We first construct CAVSR, a benchmark of 605 real-world videos collected from multiple platforms, and develop a hierarchical risk taxonomy comprising 6 top-level categories and 26 fine-grained labels to support systematic evaluation of childrens viewing risks. Then, we propose QVRS-E, a knowledge- and experience-augmented video reviewing framework that combines multi-agent collaboration with expert and experiential knowledge to support targeted evidence acquisition and fact-grounded reviewing decisions. Extensive experiments demonstrate that our method significantly enhances the reviewing of child-related risks integrated with vision-language models, and yields more robust review reports.

Abstract:
\beginabstract Video human pose estimation benefits from temporal context but remains vulnerable to occlusion and motion blur. Existing fusion methods assess frame importance in a single forward pass and do not revisit reliability as representations improve. We propose ReliaPose, a closed-loop reliability-aware temporal fusion framework that combines spatial confidence from an independently parameterized heatmap head with auxiliary supervision and temporal conformity between frame features. These signals asymmetrically modulate temporal attention: uncertain queries are amplified to gather context, whereas reliable keys and values are strengthened. Across shared-parameter refinement rounds, conformity is recomputed against the refined target representation, neighboring confidence maps remain fixed, and the target-frame confidence is refreshed after each non-final round. On PoseTrack2017, PoseTrack2018, and PoseTrack2021, ReliaPose achieves 89.0, 87.2, and 86.5 mAP, improving over prior best results by 1.0, 2.3, and 1.8 points, respectively. Ablations validate the complementary reliability signals and closed-loop refinement. \endabstract

Abstract:
Autoregressive (AR) diffusion models offer a promising framework for sequential generation such as video synthesis by combining diffusion modeling with causal inference. Although they support streaming generation, existing AR diffusion methods struggle to scale efficiently. In this paper, we identify two key challenges in hour-scale real-time human animation. First, most forcing strategies propagate sample-level representations with mismatched diffusion states, causing inconsistent learning signals and unstable convergence. Second, historical representations grow unbounded and lack structure, preventing effective reuse of cached states and severely limiting inference efficiency. To address these challenges, we propose Neighbor Forcing, a diffusion-step-consistent AR formulation that propagates temporally adjacent frames as latent neighbors under the same noise condition. This design provides a distribution-aligned and stable learning signal throughout the AR chain. Building upon this formulation, we introduce a structured ConvKV Memory mechanism that compresses the keys and values in causal attention into a fixed-length representation, enabling constant-memory inference and effectively unbounded video generation without short-term motion-frame memory. Experiments show improved convergence, hour-scale generation quality, and inference efficiency over existing AR diffusion methods. SoulX-LiveAct enables hour-scale real-time human animation at 20 FPS on two NVIDIA H100 or H200 GPUs while attaining strong lip-sync accuracy, human animation quality, emotional expressiveness, and inference efficiency.

Abstract:
Deep learningbased methods have shown promising results in chest X-ray (CXR) image classification. However, dataset bias has been widely reported in medical image analysis tasks, where spurious correlations between bias attributes (i.e., non-diagnostic attributes) and disease labels are prevalent. Existing studies have demonstrated that deep neural networks tend to learn from shortcuts under such bias, leading to predictions that overly rely on bias attributes spuriously correlated with the target classes. A common strategy to address this issue is to leverage bias information, such as bias labels or predefined bias types, to guide debiasing. However, in clinical practice, bias types are often difficult to define in advance, and bias labels are costly and labor-intensive to obtain. Therefore, training a debiasing model without bias information remains a significant challenge. To address this challenge, we propose an Adversarial and Correlation-Aware Feature Augmentation framework for debiased CXR image classification. First, we present a novel strategy that facilitates intrinsic feature learning (i.e., feature that inherently defines the target class) by exploring the correlations between bias-aligned and bias-conflicting samples (i.e., samples with and without spurious correlations). To achieve this goal, we introduce a Transformer-based encoder module to capture inter-sample correlations at the batch level. By enabling cross-sample interactions, the proposed module promotes the learning of class-shared intrinsic features while suppressing the model's reliance on bias attributes. However, due to the presence of dataset bias, bias-conflicting samples are often scarce, which limits the effectiveness of such correlation modeling. To alleviate this issue, we perform adversarial feature augmentation on disentangled bias features to generate more diverse bias-conflicting samples, thereby improving the reliability of the correlation learning process. Furthermore, motivated by the observation that shallow layers encode fewer bias features, we introduce multiple debiasing experts at different network depths and aggregate the debiasing knowledge learned by shallow experts into the deepest expert via reverse self-distillation. Extensive experiments demonstrate the effectiveness of the proposed framework on debiased CXR image classification task without bias information.

Abstract:
Unmanned aerial vehicles (UAVs) have emerged as powerful embodied agents. One of the core abilities is autonomous navigation in large-scale three-dimensional environments. Existing navigation policies, however, are typically optimized for low-level objectives such as obstacle avoidance and trajectory smoothness, lacking the ability to incorporate high-level semantics into planning. To bridge this gap, we propose ANWM, an aerial navigation world model that predicts future visual observations conditioned on past frames and actions, thereby enabling agents to rank candidate trajectories by their semantic plausibility and navigational utility. ANWM is trained on 4-DoF UAV trajectories and introduces a physics-inspired module: Future Frame Projection (FFP), which projects past frames into future viewpoints to provide coarse geometric priors. This module mitigates representational uncertainty in long-distance visual generation and captures the mapping between 3D trajectories and egocentric observations. Empirical results demonstrate that ANWM significantly outperforms existing world models in long-distance visual forecasting and improves UAV navigation success rates in large-scale environments.

Abstract:
The prevailing assumption in the multimedia community is that scaling will eventually solve the hallucination problem in Vision-Language Models (VLMs): larger models will generate more accurate outputs and better verify them. We challenge this assumption with both empirical evidence and a theoretical argument. Through systematic experiments across three VLM scales (3B, 7B, 72B parameters) on document understanding tasks, we demonstrate that while generation accuracy improves substantially with scale (+26pp from 3B to 72B), verification quality actually degrades: error detection rate drops from 43% to 8% at the 7B scale, with the smallest model proving the best error detector. We observe that model confidence is anti-predictive of verification quality: false acceptances tend to carry higher confidence than correct rejections, undermining confidence-based filtering strategies. We explain these findings through the generation-verification asymmetry thesis: generation compresses input evidence into outputs, while verification requires decompressing outputs back to evidence, a capability that current transformer architectures lack and that scaling does not provide. This asymmetry is amplified in VLMs, where visual evidence undergoes additional lossy compression through vision encoders. As a proof-of-concept, we show that chain-of-thought reasoning provides zero FAR improvement, while evidence-citing verification (forcing models to locate and quote input evidence before judging) reduces FAR by up to 19pp at the 72B scale, albeit at the cost of higher false rejection, suggesting that verification-native architectures, not scale alone, are the path forward. We conclude by proposing five open problems that frame verification as a first-class research challenge for the multimedia community.

Abstract:
Automated visual monitoring of biodiversity produces unprecedented volumes of imagery, yet rare speciescritical for ecological assessmentremain severely underrepresented, often with fewer than ten training samples. While synthetic augmentation via generative models offers a principled remedy, naively mixing generated and real images introduces domain shift and variable fidelity, both of which degrade learned representations. We present QuACS (Quality-Aware Cross-Domain Self-Distillation), a unified framework built on a simple yet powerful principle: decouple what to learn from how much to trust. QuACS operates in two stages. Stage~I employs LLM-guided diffusion to synthesize diverse tail-class samples. Stage~II integrates them through three synergistic mechanisms: (1) multi-center self-distillation that decouples domain-specific styles from semantic content, (2) quality-aware weighting that suppresses low-fidelity samples via discriminator uncertainty, and (3) hierarchical part attention that captures localized discriminative cues via learnable queries. On two long-tailed bird recognition benchmarks with imbalance ratios up to 293:1, QuACS achieves 88.93% and 86.84% Top-1 accuracyoutperforming state-of-the-art SSL methods by 7.120.0% absolute while maintaining comparable model complexity. These results establish quality-aware cross-domain learning as a principled solution for fine-grained recognition under extreme scarcity, enabling scalable biodiversity monitoring in real-world deployments.

Abstract:
Most existing visual CAPTCHAs remain spatially solvable: the required information is exposed by static appearance, local structure, and interface state. This assumption is weakened by advances in multimodal large language models (MLLMs) and Graphical User Interface (GUI) agents, which exhibit strong visual perception, reasoning, and browser interaction capabilities. We propose Motion Vision CAPTCHA (MVCAP), a hierarchical motion-based CAPTCHA framework in which target semantics are instantiated as motion-defined foreground structures and become recoverable only through temporal segregation from a dynamically evolving background. Built on this shared principle, MVCAP is instantiated in three perceptually progressive levels: coherent motion, structural motion, and biological motion. To evaluate this framework, we introduce MVCAP-Bench, a browser-based benchmark with 600 live CAPTCHA instances, together with a matched foreground-only control benchmark, MVCAP-Bench-FG. We evaluate humans, Browser Use agents, native computer use agents, and a supplementary offline VQA setting derived from the same instances. Results reveal a substantial human--agent gap: on the full MVCAP-Bench, human accuracy reaches 99.6%, whereas the best GUI agent achieves only 16.8%, close to the six-way chance level. The foreground-only control further shows that the key difficulty comes from dynamic background camouflage rather than answer format or browser interaction alone. These findings identify a measurable human--agent perception gap and position MVCAP-Bench as a benchmark for studying motion-defined perception in current agents.

Abstract:
With the proliferation of multimedia and social sensing technologies, massive volumes of time series data are continuously collected from platforms capturing emotional, behavioral, and infrastructural signals, making anomaly detection increasingly critical. Current time series anomaly detection (TSAD) methods typically learn representations in geometrically unconstrained spaces, causing two critical issues: reconstruction-based methods indiscriminately encode contaminated anomalies, and augmentation-based methods learn class boundaries without meaningful inter-class geometry. We propose HyperAD, a framework that imposes hyperspherical geometry on the latent space to learn structured representations for anomaly detection. HyperAD employs angular prototypes and an information gate to decouple reconstruction and classification into complementary subspaces on the unit hypersphere, and adaptively corrects pseudo-anomaly labels based on the evolving angular relationships among prototypes. Experiments on five benchmarks demonstrate superior detection performance.

Abstract:
Vision-language models (VLMs) process large numbers of visual tokens, resulting in substantial inference latency and memory overhead. This has motivated extensive research on visual token compression. While training-free strategies rely on heuristic metrics and suffer significant performance degradation under high compression ratios, many training-based methods introduce external compression modules that force the VLM backbone to adapt, incurring substantial retraining cost and compromising VLMs' priors. Effective visual token compression hinges on strong information encoding, a capability already present in pretrained VLMs but underutilized by existing approaches. Motivated by this, we propose VisCo, a training-efficient self-compression framework that reuses the pretrained VLM itself as an intrinsic compressor. VisCo is a parameter-sharing autoencoder that compresses visual information using a small set of memory tokens and transfers hierarchical information from encoding to decoding. Experiments show that VisCo surpasses prior methods across all evaluated compression ratios, with larger gains under more aggressive compression, and remains stable even in the extreme single-token setting. Moreover, when combined with the original visual tokens, the learned memory tokens can even improve the base model, suggesting that VisCo captures complementary representations beyond compression.

Abstract:
Event cameras have shown great potential for autonomous driving due to their high temporal resolution and wide dynamic range. However, most existing event-based object detectors rely on LSTM-based temporal aggregation, which remains insufficient in highly dynamic scenes. In particular, conventional LSTMs depend on a single global memory to model temporal dependencies, making them vulnerable to abrupt variations in target-related event patterns and thus limiting robustness. Moreover, their implicit temporal modeling is not well suited to capturing motion dynamics explicitly, which often leads to inaccurate object localization. To address these issues, we propose EDeMM, a novel event-based object detection framework with enhanced memory and motion modeling capabilities. Specifically, we design a Double-Phase LSTM (DPLSTM) consisting of Outer and Inner recurrent units to jointly capture global temporal dependencies and local temporal variations, thereby improving memory robustness under dynamically changing inputs. We further introduce a Motion-Enhancement Module (MEM) to explicitly strengthen motion trend representation and improve localization accuracy in highly dynamic scenarios. Extensive experiments on the GEN1 and 1Mpx datasets demonstrate that EDeMM achieves superior detection performance and robustness under challenging real-world conditions.

Abstract:
Current Multimodal Large Language Models (MLLMs) have advanced rapidly and follow two main paradigms: one involves concatenating visual and textual tokens for input into Large Language Model (LLM), while the other injects visual knowledge into Feed-Forward Networks (FFNs) of the LLM. However, the former introduces a large number of redundant visual tokens, substantially increasing computational overhead. The latter reduces sequence length but removes images from the input stage entirely, which may hinder the activation of image-related knowledge. Moreover, naively injecting visual knowledge into the FFNs could induce activation distribution shifts due to cross-modal semantic discrepancies. To address these challenges, we develop Beyond Token Stacking (BTS), a novel visual enhancement framework for MLLMs. Specifically, we propose a Multi-scale Visual Modulation (MVM) mechanism that injects multi-scale visual information at the input stage via modulation without increasing the token sequence length. Furthermore, we develop an Adaptive Fused-feature Calibration (AFC) module that performs per-channel affine calibration to decouple visual injection from distribution alignment in the FFN layers, improving multimodal fusion stability while preserving language priors. Experiments show that our approach achieves state-of-the-art performance.

Abstract:
Contemporary embodied multimedia systems demand agents capable of seamlessly bridging high-dimensional sensory streams, complex natural language instructions, and precise physical interactions. While Vision-Language-Action (VLA) models show promise in this direction, their deployment in safety-critical, real-time scenarios like autonomous driving is constrained by inefficient processing of redundant visual tokens, weak 3D spatial grounding, and an opaque semantic-to-control gap. We propose HiST-VLA, a hierarchical and spatio-temporal VLA framework specifically designed to handle the rigorous multimodal requirements of embodied systems. Our core contributions are: 1) A dynamic token sparsification mechanism that dynamically prunes redundant visual tokens based on self-similarity attention, serving as an efficient cross-modal alignment layer for high-throughput multi-view video streams. 2) A multi-stage, semantics-aligned hierarchical planner that bridges granular language commands with precise motion primitives, incorporating confidence-aware regularization for transparent uncertainty quantification. Extensive evaluation on the NAVSIM v2 benchmark demonstrates state-of-the-art performance, achieving an EPDMS of 88.6 on the open-loop Navtest setsurpassing prior SOTA methods by 1.5 points. Crucially, we provide in-depth analyses of cross-modal token efficiency, interpretability, and real-world deployment feasibility, establishing a robust paradigm for trustworthy embodied multimedia intelligence.

Abstract:
Unified Multimodal Relation Extraction (UMRE) aims to identify intra-modal and cross-modal relations between textual entities and visual objects. However, existing UMRE studies still encounter two critical issues: ignoring inherent aleatoric uncertainty causes noise propagation, and deep-seated heterogeneity between distinct modal distributions hinders alignment. To address these issues, we propose the Uncertainty-Guided UMRE Network (UG-UMRE). Specifically, we design an Uncertainty-Driven Unimodal Augmentation (UDUA) module, which models features as Gaussian distributions based on the Variational Information Bottleneck. By incorporating an uncertainty-aware self-supervised contrastive learning mechanism, UDUA effectively filters out noise while maintaining semantic consistency. Furthermore, we introduce the Joint Aleatoric Uncertainty Alignment (JAUA) module as a global semantic pre-calibration mechanism. JAUA leverages probabilistic distribution consistency to construct a shared latent space, eliminating the distributional gap by synchronizing cross-modal statistical properties, thereby laying a robust foundation for fine-grained interaction. Experiments on three benchmark datasets (UMRE, MORE, and MNRE) demonstrate that UG-UMRE achieves state-of-the-art performance. Further analysis validates the pluggable and effective performance of the proposed UDUA and JAUA modules.

Abstract:
\beginabstract In recent years, vision--language models (VLMs) and vision--language--action (VLA) models have become mainstream choices for autonomous-driving decision making. However, many methods still encode key-object localization as ordinary text tokens, which weakens visual localization and directly compromises planning stability. We investigate a new formulation that aims to improve object-localization reliability while preserving planning interpretability. Trajectories remain in numeric text form to keep the planning process interpretable, whereas object localization is routed through Visual Reference Tokens (VRTs) rather than coordinate strings, mitigating token activation dispersion in the language channel. Based on this design, we propose VRT-3D-VLM, which injects depth-aligned 3D positional information into visual tokens and employs an object-centric decoder to jointly regress continuous 2D boxes and ego-frame 3D boxes. Although trained only on nuScenes, VRT-3D-VLM demonstrates strong stability in both open-loop planning and NeuroNCAP closed-loop driving. In particular, it outperforms current state-of-the-art models on the Frontal and Stationary closed-loop subsets. Ablation studies show that pure text grounding leads to token activation dispersion and degraded planning, whereas depth-enhanced VRT grounding yields more reliable localization and safer driving behavior. \endabstract

Abstract:
Instruction following is emerging as a defining capability of multimodal embedding models: for a fixed query, changing the instruction should change which candidate is preferred. Existing benchmarks do not isolate this property, because they rarely keep the query and candidate set fixed while varying only the instruction. This limitation is especially acute in video retrieval, where a single query naturally admits multiple globally plausible matches along different semantic axes, and instruction following requires the model to reweight which aspect of relevance should dominate. We present IF-Bench, an evaluation benchmark for instruction following in video embedding models built on a simple principle---controlled preference flips: fix the query and candidate set, vary only the instruction, and require the preferred candidate to reverse. To instantiate this setting at scale, we develop a video-semantic pipeline that automatically discovers structurally valid preference-flip instances from publicly available videos, enforcing premise plausibility, controlled semantic contrast, contrastive instruction validity, and anti-leakage constraints. The resulting benchmark contains 4,000 instances across six major content categories, covering both cross-axis preference flips and harder within-axis fine-grained distinctions. We further introduce two complementary metrics, Flip Accuracy and Soft Flip Score, to measure both the correctness and confidence of preference reversals. Experiments on nine state-of-the-art embedding models show that even the strongest model achieves only 47.9% Flip Accuracy. Strong retrieval performance does not reliably translate into instruction-conditioned preference control, revealing a systematic capability gap that existing benchmarks fail to expose.

Abstract:
Multimodal Aspect-Based Sentiment Analysis (MABSA) predicts sentiment toward a specified aspect from paired text and images. In realistic multimodal posts, however, the two modalities are often asymmetric in reliability and may even provide conflicting evidence for the queried aspect. Correlation-based fusion is therefore vulnerable to dataset-specific shortcuts and unreliable modality cues. To address this problem, we propose CMF-LLaVA, a causally inspired modality-aware extension of LLaVA for MABSA. The core of the model is a Causal Disentanglement and Attribution (CDA) module that performs four steps: (1) reduces readily exploitable label cues through adversarial regularization, (2) derives modality-effect surrogates by contrasting aspect-conditioned representations with learnable null references, (3) reduces aspect--context dependence using HSIC, and (4) performs reliability-aware FiLM fusion. Semantic baseline regularization constrains the null references to induce high-uncertainty predictions, preventing arbitrary latent offsets. The resulting effects are interpreted as intervention-inspired representation contrasts rather than formally identified causal quantities. We further introduce the Modality Attribution Dataset (MAD), a human-annotated benchmark with substantial inter-annotator agreement (Krippendorff's \alpha=0.81), to evaluate agreement between model attributions and human judgments. On Twitter-15, Twitter-17, and MASAD, CMF-LLaVA outperforms strong MABSA and MLLM baselines and exhibits smaller performance drops under modality conflict, spurious visual cues, and aspect ambiguity. These results support causally inspired effect modeling for robust and interpretable MABSA.

Abstract:
This paper provides an in-depth analysis of the limitations of existing evaluation metrics in long-text image description tasks, revealing that the score dilution effect caused by training bias is the core reason why fine-grained errors are difficult to detect. To address this issue, we propose the v-PPL metric, a visual-grounded perplexity that achieves a decomposed evaluation of image-text alignment quality through token-level quantification of visual contribution---explicitly measuring the degree to which each textual token is grounded in visual evidence. Experimental results demonstrate that v-PPL effectively suppresses interference from non-critical tokens, accurately captures subtle visual deviations in long texts, and reaches new heights in terms of sensitivity and objectivity. This research paves a new way for fine-grained visual understanding evaluation in complex scenarios.

Abstract:
3D Gaussian Splatting (3DGS) enables real-time, high-quality novel view synthesis, but its models routinely exceed 600 MB because each scene stores many primitives with high-dimensional view-dependent attributes. We present LeanGS, a compact 3DGS framework that jointly removes redundant primitives and reduces per-Gaussian appearance storage without post-hoc compression. For primitives, we interpret alpha blending as a rendering-attention mechanism and use an energy-retention objective to derive scene-adaptive pruning budgets. For appearance, we introduce Factored View-Dependent Color, which replaces 48 spherical-harmonic coefficients with 5 scalar weights over a shared view basis, reducing per-Gaussian attributes from 62 to 22. Across Mip-NeRF 360, Tanks & Temples, and Deep Blending (13 scenes), LeanGS shrinks standard 3DGS models from 675 MB to 24.88 MB on average with 90% fewer Gaussians and only a 0.06 dB PSNR drop. It also remains compatible with post-hoc quantization, reaching 9.2 MB when further compressed.

Abstract:
Bias in pre-trained vision-language models often leads to unfair and unreliable predictions due to spurious correlations in large-scale training data. Existing debiasing methods are typically static or stage-isolated, separating training-time rebalancing from deployment-time correction and thus failing to reuse signals about subgroup failure across stages. In this paper, we propose BiasFAR, a feedback-reuse framework for bias mitigation in multimedia recognition. The key idea is to treat subgroup feedback as a shared control signal across training and deployment rather than a stage-specific diagnostic. Specifically, BiasFAR uses model-driven and vision-language alignment feedback to adaptively reweight subgroups during training, and employs a re-inference mechanism that combines sample-specific confidence with a feedback-derived class-level subgroup-risk prior to refine predictions for bias-sensitive samples at test time. Under this view, adaptive re-weighting and selective re-inference are not independent, but two operational forms of the same closed-loop feedback mechanism. To remain practical for large pre-trained backbones, BiasFAR adopts a parameter-efficient fine-tuning strategy by optimizing only lightweight projection heads, thereby avoiding substantial computational overhead. This closed-loop design enables cross-stage bias mitigation from optimization to deployment, improving robustness in diverse multimedia scenarios. Extensive experiments on Waterbirds, CelebA, BAR, CIFAR-10C, and MetaShift show that BiasFAR achieves competitive performance against representative state-of-the-art debiasing methods while maintaining favorable efficiency.

Abstract:
Open-Vocabulary Fine-Grained Action Recognition (OV-FGAR) aims to recognize actions with subtle inter-class differences in an open setting. It faces two core challenges: capturing fine-grained discrepancies among similar actions and achieving zero-shot generalization, along with limitations such as scarce annotations, ambiguous semantic alignment, and complex spatio-temporal modeling. Existing methods focus either on closed-set fine-grained recognition (fixed similar categories) or open-vocabulary coarse-grained recognition (no subtle variations), failing to address the coupled challenges of fine-grained discrimination and open-vocabulary zero-shot generalization. To this end, we propose Hierarchical Action Parsing with Differentiable Sampling (HAP-DS), a new OV-FGAR framework with two complementary components: Segment-to-Frame Hierarchical Parsing (SFHP) and Soft-Hard Sampling Fusion (SHSF). Specifically, SFHP decomposes complex actions into compact sub-action segments, including Segment Decoupling and Frame Association (modeling sub-action structures via bidirectional cross-attention) and Differentiable Sampling Quota Allocation (adaptive sampling resource scheduling based on query scores). In addition, SHSF integrates soft sampling (stable gradients) and hard sampling (practical inference), enabling end-to-end optimization and using Straight-Through Estimation (STE) to bridge training-inference gaps. Extensive experiments on three OV-FGAR benchmarks demonstrate that our HAP-DS outperforms state-of-the-art methods, validating the effectiveness of our hierarchical parsing and differentiable sampling strategies.

Abstract:
All-in-one image restoration aims to handle diverse degradations within a unified framework. Existing methods commonly encode heterogeneous degradation conditions in a shared latent space, where degradation-related cues and scene content can remain entangled. We characterize the resulting challenge as dual ambiguity: semantic ambiguity in channel-wise modulation and spatial ambiguity in restoration responses, which can lead to content corruption and residual artifacts. To mitigate this issue, we propose DAR-Net, a Dual-Ambiguity Rectification Network for all-in-one image restoration. DAR-Net first introduces a Degradation Archetype Representation (DAR) module to construct a structured degradation state through simplex-constrained archetype mixture modeling. Based on this state, a Semantic Ambiguity Rectification (SeAR) module generates degradation-aware prompts to improve channel-wise conditioning in the decoder. A Spatial Ambiguity Rectification (SpAR) module further regularizes degradation-aware and complementary features toward orthogonal response subspaces, reducing spatial interference between removal and preservation cues. Extensive experiments on standard benchmarks show that DAR-Net achieves the best overall performance under both 3-degradation and 5-degradation settings, improving the average PSNR over the strongest competitor by 0.14 dB and 0.34 dB, respectively.

Abstract:
3D Gaussian Splatting (3DGS) has recently gained significant popularity for efficient 3D reconstruction, prompting many platforms to offer commercial 3D reconstruction services based on user-uploaded images. This untrusted-input reliance exposes a severe security risk within the core pipeline of 3DGS. By uploading images with malicious perturbations, attackers can trigger uncontrolled Gaussian growth, leading to GPU memory exhaustion or even Denial-of-Service (DoS). However, existing attacks cause limited damage and struggle in complex 3D scenes due to hardware constraints. To address this, we propose H-poison, an efficient 3DGS adversarial attack framework centered on manipulating 3DGS systems to trigger exponential growth of Gaussian primitives. Specifically, we first formulate a Gaussian-weighted Total Variation (GW-TV) objective to synthesize needle-like texture perturbations, which reliably induce unnecessary Gaussian densification across arbitrary directions and step sizes. We subsequently design a Composable 3D-Driven Attack Module (C3DAM) to inject reusable 3D adversarial textures into 2D images. The cascaded C3DAMs bridge the 2D-3D domain gap by simultaneously outputting multi-view consistent adversarial images, optimizing scene-specific proxy models, and refining a universal 3D noise model during cross-scene attacks. Extensive white-box and black-box experiments demonstrate that H-Poison achieves state-of-the-art performance in both victim-side damage and attacker-side cost, while preserving imperceptible perturbation and high reconstruction quality.

Abstract:
Diffusion models have demonstrated superior performance in image super-resolution, yet balancing perceptual quality, pixel-level fidelity, and inference speed remains a significant challenge. Most existing approaches prioritize efficiency by leveraging Variational Autoencoders (VAEs) for latent-space generation; however, this paradigm introduces an inherent information bottleneck that limits pixel-level fidelity, particularly when recovering strongly semantic details such as small text. To address this trade-off, we propose POD-SR, a high-fidelity, pixel-space one-step diffusion framework. POD-SR utilizes a frequency-decoupled pixel-space architecture to circumvent the VAE bottleneck, effectively preserving high-frequency details. Furthermore, to maintain restorative quality in a single step, we introduce two key modules: a Degradation-Aware Graph Perception Network (D-GPN) and an Asymmetric Generative Prior Alignment (AGPA) module. Specifically, D-GPN models local structural anomalies to dynamically predict image-specific starting timesteps, providing an adaptive alternative to static timesteps. Concurrently, AGPA realigns the restorative trajectory with the model's intrinsic generative prior to enhance structural faithfulness. Extensive evaluations confirm that POD-SR achieves competitive performance in fidelity metrics while maintaining competitive inference efficiency.

Abstract:
3D Gaussian Splatting (3DGS) has become a core representation for immersive multimedia on the web, powering volumetric video, dynamic avatars, and interactive 3D experiences. While recent advances in compression and streaming bring 3DGS content closer to end users, existing web viewers naively port CUDA-based pipelines whose synchronization and memory patterns do not generalize across GPU architectures, causing rendering failures on non-NVIDIA and mobile devices. We present WebSplatter, a WebGPU-native rendering framework that eliminates these failure modes through a wait-free radix sort and a hardware-accelerated rasterization pipeline with opacity-aware quad sizing. Evaluated across eleven devices spanning five GPU architectures, WebSplatter is the first web-based 3DGS renderer that achieves cross-device reliability, rendering every benchmark scene on every tested device without failure. WebSplatter delivers 1.1 to 2.5 times speedups over the surviving baselines and near-lossless fidelity to the native CUDA renderer, with an average \DeltaPSNR~0.07\,dB.

Authors: Haitian Li (Beijing University of Technology), Yanghao Zhou (Beijing Institute of Technology), Heyan Huang (Beijing Institute of Technology), Liangji Chen (Shanghai Film Academy), Xu Liu (Hefei University of Technology), Dian Jin (Hefei University of Technology), Jiajun Xu (Inkeverse Group Limited), Jingyun Liao (Hunan University), Tian Lan (Beijing Institute of Technology), Ziqin Zhou (University of Adelaide), Yiming Chen (Tsinghua University), Yueying Liu (Beijing University of Technology), Yu Bai (Beijing Academy of Artificial Intelligence), Changsen Yuan (Beijing University of Technology), Jinxing Zhou (OpenNLPLab), Xian-Ling Mao (Beijing Institute of Technology), Xuefeng Chen (Inkeverse Group Limited), Yousheng Feng (Inkeverse Group Limited)

Abstract:
In recent years, Multi-Talker Audio-Video Generation (MTAVG) models have shown promising performance on fundamental metrics such as lip-sync and audiovisual alignment. However, these metrics remain insufficient for assessing cinematic expressiveness in scene-level generation. In multi-character scenes, generation models must go beyond audiovisual realism to convey coherent character performance and other higher-level cinematic qualities. To fill this gap, we introduce MTAVG-Bench 2.0, a benchmark for diagnosing failure modes of cinematic expressiveness in multi-talker audiovisual generation. Unlike prior settings that mainly focus on the quality of basic multi-turn dialogue, MTAVG-Bench 2.0 targets short-drama and scene-level generation, and establishes a high-level failure taxonomy spanning acting, atmosphere, and cinematography. Based on this taxonomy, we construct 11.6K diagnostic QA instances, together with subsets for short-drama-level assessment and temporal localization of failure modes, to systematically evaluate the ability of omni large language models to diagnose high-level audiovisual failures. Experimental results show that commercial omni models such as Gemini substantially outperform other evaluators, yet even the strongest models continue to struggle with complex failures in our benchmark. These results demonstrate that MTAVG-Bench 2.0 provides a systematic benchmark for failure diagnosis in cinematic multi-talker audiovisual generation. The code and data are available at GitHub and Hugging Face.

Abstract:
Zero-shot 3D visual grounding aims to localize specific objects based on textual descriptions and 3D visual input, playing an important role in many fields such as embodied intelligence. However, the effectiveness of existing methods is significantly hindered by the ambiguous query text and deficient viewpoints. To address these issues, we propose TDVR, a training-free reasoning framework that disambiguates the input text and infers accurate viewpoints for zero-shot 3D visual grounding. First, we construct semantic 3D scene graph from the detected instances in the 3D point cloud.Subsequently, we put the original query, appearance and spatial relationship descriptions into the LLM for fusion, thereby disambiguating the initial input.Given the complexity of linguistic expressions, we leverage chain-of-thought reasoning to generate the structured representation of disambiguated query.Then taking the scene graph and structured query as input, we get the optimal view via viewpoint reasoning to solve the problem of missing viewpoints during grounding.Based on the obtained optimal viewpoint,we further discriminate the distracting objects,enabling the model with the ability to distinguish similar instances.After that, we match the category text and appearance images with the query by computing the similarity of feature vectors.Finally the target object was identified by integrating the viewpoint score,confusion score,category score,and appearance score.Compared with previous methods, our TDVR has stronger capabilities in viewpoint reasoning, similar object discrimination, and ambiguous query understanding.Experimental results on the public ScanRefer dataset show that our method outperforms the existing state-of-the-art methods by 15.25% and 14.46% in Acc@0.25 and Acc@0.5 respectively, demonstrating the effectiveness of our TDVR in addressing ambiguous query text and deficient viewpoints.

Abstract:
Current autoregressive generation of scalable vector graphics (SVG) often struggles to maintain structural coherence and geometric fidelity, leading to part omission, layout drift, and repetitive strokes. This limitation arises from their purely text-conditioned decoders that decode SVGs directly from prompts without visual references. To address this, we present \emphFetch, Frame, and Focus (F^3SVG), a retrieval-augmented framework for decoder-side visual guidance injection in text-to-SVG generation. Specifically, given a prompt, Fetch retrieves relevant reference SVGs via CLIP-based text similarity search, decomposes each reference into ordered path groups, and extracts \emphscene tokens and \emphdetail tokens. Frame compresses the scene tokens into a \emphPanoramic Memory that provides persistent compositional guidance, while Focus encodes the detail tokens into \emphSpotlight Memory slots that preserve group-specific structure and are accessed through a lightweight token-conditioned router. Both memories are injected only into the late layers of a frozen autoregressive decoder, preserving single-pass decoding and the backbone architecture. Furthermore, we conduct systematic ablations to identify principles for incorporating these memory modules into cutting-edge models like OmniSVG. Extensive experiments on MMSVGBench show that F^3SVG consistently improves upon OmniSVG across quality metrics while adding only 257M trainable parameters (3.3% of a frozen 7.6B backbone), and remains competitive with proprietary closed-source models like GPT-5.4. These results suggest that retrieval-guided memory injection is an effective, lightweight, and architecture-preserving route to stronger autoregressive SVG generation.

Abstract:
Medical vision-language models (MVLMs) are increasingly used as perceptual backbones in radiology pipelines and as the visual front end of multimodal assistants, yet their reliability under real clinical imaging workflows remains underexplored. Prior robustness evaluations often assume clean, curated inputs or study isolated corruptions and pixel-level perturbations, overlooking routine acquisition, reconstruction, display, and delivery operations that preserve clinical readability while subtly shifting image statistics. To address this gap, we propose CoDA, a chain-of-distribution framework that constructs clinically plausible pipeline shifts by composing acquisition-like shading, reconstruction and display remapping, and delivery and export degradations. Under masked structural-similarity constraints, CoDA jointly optimizes stage-composition families and within-stage parameters to induce failures while preserving visual plausibility. Across brain MRI, chest X-ray, and abdominal CT, CoDA substantially degrades the zero-shot performance of CLIP-style MVLMs, with ablations showing that chained compositions are consistently more damaging than any single stage. We also evaluate multimodal large language models (MLLMs) as technical-authenticity auditors of imaging realism and quality rather than pathology. Proprietary MLLMs show degraded auditing reliability and persistent high-confidence errors on CoDA-shifted samples under a single-call self-challenge protocol, while the medical-specific MLLMs we test exhibit clear deficiencies in medical image quality auditing. Motivated by these findings, we introduce a post-hoc repair strategy based on teacher-guided token-space adaptation with patch-level alignment to improve zero-shot classification accuracy on archived CoDA outputs. Overall, our study characterizes a clinically grounded threat surface for MVLM deployment and shows that lightweight alignment improves robustness in deployment.

Abstract:
The prevailing video retrieval evaluation suffers from fragmented benchmarks that obscure true model generalization. To diagnose this limitation, we conduct a structured assessment across 16 established public datasets, revealing critical capability gaps, particularly pronounced deficiencies in fine-grained, temporal, and partially relevant video retrieval. The discoveries directly motivate the holistic design of our framework: V-SynFlow synthesizes 1.55 million high-quality, multi-task video-text pairs explicitly targeting the uncovered deficiencies, while the Modality Pyramid curriculum trains the General Video Embedder (GVE) by leveraging inherent task hierarchies for progressive learning. GVE achieves leading zero-shot performance across diverse scenarios, including the highest overall UVRB macro averages. Our analysis further demonstrates that conventional benchmarks poorly correlate with holistic retrieval capability, and identifies partially relevant retrieval as a pivotal yet overlooked real-world challenge. This work establishes a practical path to universal video retrieval by tightly coupling evaluation diagnostics with structured training on diverse synthesized data.

Abstract:
Online video understanding requires models to proactively decide when to respond and which information to preserve from inherently redundant continuous visual streams. However, current online Video Large Language Models (Video-LLMs) either operate in a reactive, query-driven manner with limited autonomy, or rely on active mechanisms that demand costly instruction tuning yet still yield suboptimal performance. Moreover, these existing paradigms often face challenges in effectively reducing the dense spatiotemporal redundancy of continuous video streams. To address these limitations, we propose Reference State Verification (RSV), a generalized, training-free paradigm that transforms pre-trained offline Video-LLMs into proactive online assistants through semantic state comparison. Building upon RSV, we first introduce Attention Distribution Divergence (ADD), utilizing the relative deviation of cross-modal attention as a robust verification signal for autonomous response triggering. Furthermore, we present Verification-guided Spatio-Temporal Compression (VSTC), which re-purposes these verification signals as a free byproduct to achieve multi-granular redundancy elimination, including semantic-guided frame dropping and fine-grained spatial token pruning. Extensive experiments demonstrate that our plug-and-play approach significantly enhances the inference efficiency and proactive responsiveness of various Video-LLMs without requiring any architectural modifications or parameter fine-tuning.

Abstract:
Radiology report generation (RRG) has recently benefited from large language models, which substantially improve report fluency. However, clinically faithful generation remains challenging because current supervision is still imposed mostly at the report level. This creates a granularity mismatch: radiology reports are composed of disease-grounded findings, while existing methods are trained mainly with whole-report objectives. To address this problem, we propose Graph-Supervised Hierarchical Clinical Alignment, which reformulates image-report supervision as a hierarchical clinical alignment problem. Our method structures this alignment as a disease-conditioned process, where supervision is decomposed into two levels: Disease-Centric Alignment for fine-grained disease-specific correspondence, and Global Clinical Semantic Alignment for report-level semantic coherence. A clinical knowledge graph is used as a training-time-only structural prior that defines disease-specific supervision units and their clinical relationships, introducing no additional overhead at inference. Because standard contrastive alignment could produce false negatives when studies share overlapping pathologies, we combine instance-conditioned discriminative matching with disease-conditioned soft regularization, enabling fine-grained yet clinically consistent cross-modal representations. Experiments on MIMIC-CXR, IU-Xray, and COV-CTR show that our method consistently improves performance on both conventional and clinical metrics. Notably, our 3B model surpasses several prior systems with larger 7B/13B backbones, suggesting that improving supervision structure, rather than increasing model size, can be more effective for RRG.

Abstract:
Surgical Visual Question Localized-Answering (VQLA) aims to predict semantically accurate answers to surgical questions while simultaneously localizing the relevant regions in surgical images, supporting advanced surgical navigation and educational applications. Despite prior efforts, most existing approaches predict answers and their corresponding locations separately using simple loss functions (e.g., cross-entropy loss for answer prediction and GIoU loss for bounding box regression), without considering the semantic consistency between them, which may result in inaccurate predictions. To address this limitation, we propose a novel surgical VQLA framework that effectively enhances answer-location consistency via a newly introduced reverse question learning module. Specifically, this module is designed to project predicted answers, their corresponding locations, and associated visual contexts back into the original question embedding space, and models their joint dependencies to impose a reverse semantic constraint. This constraint effectively regularizes the VQLA training process and enforces consistency between predicted answers and their locations, thereby improving prediction reliability and accuracy. Furthermore, we introduce a contrastive learning-based global-local alignment module to align global images with question-relevant local patches, enabling the model to better focus on fine-grained visual cues thus improving its VQLA performance. Comprehensive experiments on two publicly available porcine video datasets, i.e., EndoVis-18-VQLA and EndoVis-17-VQLA, along with one in-house collected patient video dataset, i.e., SGH-VQLA, demonstrate that our method achieves superior surgical VQLA performance compared with state-of-the-art approaches.

Abstract:
\beginabstract Infrared small target detection plays an important role in many practical applications, yet its deployment often requires not only high detection accuracy but also strict efficiency in terms of computation and storage. Although structured pruning has shown promising results in general vision tasks, the direct application of existing pruning methods to IRSTD often leads to severe performance degradation at high pruning ratios. In this paper, we revisit structured compression for IRSTD from the perspective of representation organization and propose a unified framework termed Channel Polarization and Structured Pruning (CPSP). The proposed method first introduces a channel polarization mechanism during dense training to reduce channel-wise mixed encoding of target and background information, thereby encouraging clearer target-dominant and background-dominant channel roles and improving the pruning-friendliness of the network. Based on the resulting disentangled representations, we further design a global-budget disentanglement-aware structured pruning strategy that jointly considers generic channel importance, target-related responsibility, and structural compression gain, enabling more reliable cross-layer pruning decisions under aggressive compression budgets. Extensive experiments on multiple representative IRSTD backbones and benchmark datasets show that CPSP consistently achieves a better balance between detection accuracy and model efficiency than conventional structured pruning methods, providing an effective framework for lightweight deployment of IRSTD models. \endabstract

Abstract:
Bird's-Eye-View (BEV) representations provide a unified spatial space for autonomous driving perception. However, existing multimodal BEV methods typically fuse camera and LiDAR features through rigid operations such as concatenation or cross-attention, entangling modality-specific noise with task-relevant semantics. We propose G-DRIVE, a robust multimodal BEV framework that disentangles camera and LiDAR representations into a shared subspace and two modality-exclusive subspaces under orthogonality constraints. A Geometric Context Prior distills stable structural cues from LiDAR to guide spatially adaptive expert routing at each BEV location, enabling robust fusion without relying on fragile visual signals. Comprehensive experiments on nuScenes and Dur360BEV demonstrate that G-DRIVE achieves state-of-the-art performance.

Abstract:
In digital pathology, multi-magnification multiple instance learning (MIL) has shown strong potential for whole slide image (WSI) analysis by integrating complementary information across scales. Existing methods have introduced a variety of specialized modules for combining multi-scale features, but often overlook spatial relationships during fusion, which may connect spatially mismatched regions and dilute discriminative evidence. Therefore, the key challenge lies not merely in feature combination, but in establishing reliable cross-scale correspondence and interaction. To address these issues, we propose MPFusion-MIL, a multi-magnification MIL framework with explicit cross-scale correspondence modeling and localized fine-to-coarse fusion. Specifically, a Fine-to-Coarse Concentric Patch Division module is introduced to preserve cross-scale correspondence, a Context-Guided Patch Representation Enhancement module strengthens coarse-level screening features with surrounding context, and a Precise Spatial Fusion Strategy injects localized high-magnification evidence into aligned low-magnification representations for precise fusion. Experiments on three benchmark datasets show that MPFusion-MIL achieves average accuracy and F1-score of 94.6% and 95.1%, respectively, outperforming state-of-the-art MIL methods. Ablation studies confirm that the gains arise from improved correspondence, stronger coarse-level feature enhancement, and more precise cross-scale fusion.

Abstract:
Consolidating diverse subtasks into a unified Multimodal Information Extraction (MIE) model frequently leads to negative transfer, stemming from gradient conflicts between perception-heavy and reasoning-heavy tasks. While Mixture-of-Experts (MoE) models mitigate this via dynamic routing, they compromise structural efficiency and increase inference latency. To address this, we introduce Null-Space Constrained Multimodal LoRA (NS-MLoRA). Instead of routing activations, NS-MLoRA decouples the optimization landscape into shared and modality-specific orthogonal subspaces. We enforce orthogonality to the retained opposing gradient subspaces using a Projected Adaptive Update (PAU) mechanism, addressing the theoretical mismatch between gradient projection and adaptive optimizers. By applying these constraints strictly during training, NS-MLoRA preserves the linear reparameterization property of standard LoRA. Post-training, the orthogonal adapters can be directly merged into the backbone weights, achieving MoE-level task isolation with no additional inference overhead. Evaluations on seven MIE benchmarks demonstrate that NS-MLoRA consistently outperforms state-of-the-art MoE baselines in both performance and deployment efficiency.

Abstract:
Multi-task learning (MTL) improves efficiency and performance by jointly learning multiple related tasks through shared representations. However, inter-task interference can impair the generalization of these representations, leading to negative knowledge transfer. Knowledge distillation (KD) has emerged as a promising approach to address this issue. Existing MTL-KD methods, however, typically rely on task-specific teacher models trained on downstream datasets, incurring high computational costs and exhibiting limited generalization. To overcome these limitations, we propose a CLIP-based multi-task distillation framework, CMKD, which leverages the pre-trained CLIP model as a general teacher to enhance MTL model generalization without additional task-specific teacher training. CMKD achieves this through two complementary distillation paths. The shared representation path employs three parallel projection operators to align feature dimensions and semantic representations between CLIP and the shared MTL encoder, facilitating more effective knowledge transfer. The task-specific semantic path aligns CLIPs image-text matching logits with the students task outputs to transfer cross-modal prior knowledge. Experimental results demonstrate that CMKD consistently improves MTL performance across diverse datasets.

Abstract:
Multimodal large language model (MLLM) agents are increasingly used as personal assistants for long-running tasks. Their utility depends on continuity: agents must retrieve and use earlier evidence across dialogue, files, and workspace state. However, agents can generate plausible answers even when access to that history has degraded, causing outcome-only evaluation to overestimate true evidence use. We present MIRAGE (Multimodal Interaction Retrieval, Attribution, and Grounding Evaluation), a controlled study of historical evidence use under conversation-state variation in multimodal personal agents. MIRAGE holds evidence objects, questions, and scoring fixed while varying only conversation state, and evaluates whether an agent can determine answerability, recover the correct source, and answer from it. Across seven frontier and open-weight multimodal backbones, we find that: 1)~pre-compaction depth and post-compaction continuation form distinct, non-monotonic failure regimes rather than a single degradation curve; 2)~open-weight models rely heavily on context continuity and are reluctant to spontaneously switch to tool-mediated retrieval when provenance fails; and 3)~retrieval pressure improves source attribution in deep pre-compaction states for tool-compliant models, but consistently regresses after compaction, where stored evidence has already degraded. These findings show that historical evidence use should be evaluated under state variation, rather than inferred from outcome-only correctness.

Abstract:
The proliferation of multimodal fake news, which combines misleading text with suggestive imagery, presents a significant societal challenge, demanding robust detection methods. Existing approaches have evolved from simple feature fusion to sophisticated cross-modal attention networks to capture inter-modality relationships. However, they still face critical limitations: 1) a failure to consult external knowledge, leaving them vulnerable to evolving narratives, and 2) an inability to represent the heterogeneous evidence patterns of multimodal samples, leading to overconfident errors on ambiguous cases. To address these gaps, we propose the Evidence-Driven Multimodal Experts (EDME), a framework that emulates human fact-checking by integrating retrieval-based evidence with Dirichlet-profile decision-making. EDME first employs a Multimodal Evidence Retriever to fetch relevant real and fake news articles, providing crucial external context. This evidence is then leveraged by specialized Evidence-Enhanced Multimodal Experts, whose judgments are dynamically aggregated by an Evidence-Driven Expert Decision mechanism that composes evidence-conditioned categorical profiles. Extensive experiments on three benchmark datasets demonstrate that EDME significantly outperforms state-of-the-art baselines, improving accuracy by up to 1.2% on the Weibo dataset.

Abstract:
Existing generalizable human reconstruction methods often struggle with feature misalignment in sparse-view scenarios or rely on heavy latent tokens that incur prohibitive computational overhead. While pixel-aligned features and latent tokens offer complementary strengths in local fidelity and global coherence, effectively unifying them within a single framework remains challenging. In this work, we propose PT-Human, a point-based framework that integrates Pixel-aligned features and lightweight multimodal Tokens through an iterative refinement process. By leveraging the global guidance of tokens to rectify local feature misalignments, our method effectively unifies these complementary representations for high-fidelity reconstruction. Specifically, we first design a Multimodal Transformer to establish a unified token space, distilling initial image evidence into point-wise tokens. We then propose a neighborhood-aware iterative process for the stage-wise refinement of 3D geometry and appearance. By interleaving Gated Multi-view Attention (GMA) for cross-view feature aggregation and Gated Neighborhood-aware Attention (GNA) for geometric consistency, the framework simultaneously rectifies local feature misalignments and maintains global structural coherence via latent tokens. Experiments demonstrate that PT-Human effectively models localglobal interactions, leading to superior reconstruction fidelity across diverse subjects.

Abstract:
Current image fusion methods assume high-quality inputs and degrade under low-resolution conditions, and recent text-guided approaches also rely on single-type prompts that cannot handle distinct fusion objectives or apply guidance adaptively across spatial locations. To address these issues, we propose a novel Dual-text Guided Super-Resolution and Fusion (DGSRF) framework that jointly optimizes super-resolution and fusion tasks with two specialized text guidance streams. We design a new Interactive Gating Multimodal Cross-fusion (IGMC) module that performs cross-modal mixed attention between infrared and visible features, injects descriptive text semantics through token-importance-weighted cross-attention with content-dependent gating, and applies cross-modal gating fusion to selectively absorb complementary information. We further devise a new Hybrid Instruction Conditioned Modulation (HICM) module that combines a learnable instruction library with scene-aware selection and external instruction fusion to generate spatially-adaptive modulation parameters, enabling each spatial location to be modulated according to both local content and instruction semantics. IGMC and HICM are cascaded at every stage of a unified SR-fusion backbone, with a fusion head that predicts pixel-adaptive weight maps to produce the final fused output. Experiments on three benchmark datasets with multiple SR backbones demonstrate that DGSRF outperforms state-of-the-art fusion methods in terms of fusion quality.

Abstract:
Insufficient lighting poses a significant challenge in low-light object detection. Existing enhancement-then-detection approaches prioritize perceptual fidelity, exacerbating the conflict between enhancement and detection goals. Current enhancement-for-detection approaches are insensitive to inter-class similarity, primarily due to pixel distribution shift and residual semantic confusion. In this paper, we propose YOLO-LOD, a novel framework that follows a shallow-to-deep, decomposition-aggregation paradigm for end-to-end low-light object detection, aimed at striking a balance between visual quality and detection robustness. Specifically, we design a Hierarchically-Distilled Illumination Estimator (HDIE) built upon a multi-scale parallel architecture, which incorporates a customized Multi-distillation Feature Extraction Block-n (MFEB-n) to implement hierarchical pixel adjustment from shallow to deep layers. We address semantic confusion through two dedicated modules: Retinex-based decomposition and Wavelet-Attention Feature Aggregation (WAFA). Retinex-based decomposition achieves the disentanglement of a low-light image into illumination-invariant and illumination-specific components. The WAFA integrates the distinct frequency information of wavelet transforms with a gated attention mechanism for discriminative and complementary feature aggregation, thereby effectively mitigating inter-class confusion and boosting overall detection performance. Extensive experiments demonstrate that our approach consistently outperforms state-of-the-art methods across five public datasets, highlighting its robustness and effectiveness in diverse challenging scenarios.

Abstract:
Referring Video Object Segmentation (RVOS) aims to segment specific object instances throughout a video based on a referring language description. The prominent challenge lies in aligning static text with dynamic visual content. However, prevailing `detect-then-filter' frameworks tend to evaluate each object candidate in isolation, failing to incorporate complex contextual dynamics described in the query. In this paper, we propose PARSE-VOS, a novel, training-free framework powered by a Large Language Model (LLM), for a hierarchical, coarse-to-fine reasoning across text and video domains. Our approach begins by parsing the natural language query into structured semantic commands. Next, we introduce a spatio-temporal grounding module that generates spatio-temporal trajectories for all relevant objects, guided by the parsed semantics. Finally, a hierarchical identification module selects the correct target through a two-stage reasoning process: it first performs coarse-grained reasoning about object spatial states and motion by incorporating video context to narrow down candidates; if ambiguity remains, a fine-grained visual-semantic verification stage is conditionally triggered to disambiguate. The final output is an accurate segmentation mask for the target object. PARSE-VOS achieves state-of-the-art performance on three major benchmarks: Ref-YouTube-VOS, Ref-DAVIS17, and MeViS.

Abstract:
Modern financial markets have evolved into complex multimedia environments where high-frequency visual liquidity structures, continuous numerical streams, and sparse semantic narratives interact stochastically. Traditional reinforcement learning and generative agents often fail in this landscape because they lack physical constraints and rely on reactive intuition rather than deliberative planning, leading to hallucinated arbitrage loops. To address this, we introduce the Holographic Financial World Model (HoloFWM), a unified generative framework that treats market dynamics as a Multimodal World Simulation problem. HoloFWM utilizes an Inverse Holographic Perception module with a Dual-Stream Selective State Space Model to synchronize asynchronous Limit Order Book imagery, executive audio prosody, and textual news into a unified latent manifold. The framework employs a Riemannian Rectified Flow dynamics model with a Zero-Curl constraint that regularizes the learned dynamics toward a conservative vector field, providing an economic-consistency bias aligned with the No-Arbitrage principle. Furthermore, a System 2 Adjoint Planner performs test-time trajectory optimization that dynamically scales compute based on market entropy. Extensive experiments across six global markets in 2024 demonstrate that HoloFWM significantly outperforms state-of-the-art baselines, achieving a Sharpe Ratio of 2.45 on DJIA-30 and demonstrating superior robustness against tail risks with a Max Drawdown of -5.4%.

Abstract:
Rapid advances in artificial intelligence generated content (AIGC) have enabled increasingly realistic synthetic and manipulated speech, making generalizable audio deepfake detection (ADD) essential for multimedia security. Existing ADD systems are often formulated as binary classifiers and may overfit to attack-specific artifacts, leading to degraded performance under unseen attacks and domain shifts. To address this issue, we propose MP-ADD, a framework for generalizable audio deepfake detection that integrates Group-Augmented Cross-Layer Representation Learning (GCRL) and Multi-Prototype Authenticity Modeling (MPAM). GCRL combines semantically grouped audio augmentation with dynamic fusion of Wav2Vec2 layers to improve representation robustness. MPAM learns multiple authenticity prototypes to model the multimodal distribution of bonafide speech, together with a structural consistency constraint and two-stage decoupled optimization that stabilize feature learning and limit spoofing gradients from updating the prototypes. Evaluations of MP-ADD on ASVspoof 2019 LA, ASVspoof 2021 LA and DF, and In-the-Wild yield an average EER of 2.31% and an In-the-Wild EER of 6.08%, demonstrating competitive cross-domain generalization.

Abstract:
Traditional event extraction mainly relies on language predicates (e.g., hit and steal) to characterize events, yet this text-centered paradigm does not fully capture the embodied physical world, where events, especially in embodied security-critical scenarios, are primarily triggered by visual objects rather than language cues. With this in mind, we propose a new Visual-object Triggered Video Event Extraction and Localization (VTrig-E^2L) task for Embodied Security-critical Scenarios, aiming to extract and localize structured event tuples (i.e., [trigger-subject and its box, event type, trigger-object and its box]), where trigger-subject and trigger-object refer to the subject and object at the initial event-triggering moment, facilitating early warning. Furthermore, this paper reveals two key challenges in the proposed task: 1) how to effectively incentivize the unified reasoning of explicit language thinking and implicit visual thinking; and 2) how to highlight implicit visual thinking beyond language priors. To tackle these challenges, this paper proposes a new EagleEye approach, i.e., EagleEye acts like an eagle, using implicit visual thinking to scan the scene, spot triggers, and localize events. Specifically, EagleEye incorporates two key components: the Unified Language-Visual Thinking Module (U-LVT) and Causal-guided GRPO Module for Implicit Visual Thinking Enhancement (C-GRPO), to address the above two challenges, respectively. Extensive evaluations on our constructed CoTEvent benchmark show the advantages of EagleEye over several advanced Video-LLMs for the VTrig-E^2L task. These results justify the importance of language-visual thinking, especially implicit visual thinking, in VTrig-E^2L and demonstrate the effectiveness of EagleEye in highlighting it.

Abstract:
Hyperspectral image super-resolution (HSI-SR) reconstructs high-resolution HSI from complementary observations. Existing task-specific methods assume predefined input morphologies and lack flexibility for diverse or unknown spatial configurations. We propose U-HSR, a unified paradigm handling aligned, unregistered, and unpaired spatial morphologies without prior morphology knowledge. Specifically, the spatial morphology prompt generator (SMPG) uses similarity-driven statistics to classify inputs and synthesize textvisual prompts for morphology-aware representation. This representation jointly guides the proposed task-oriented correction and enhancement block (TCEB) and prompt-guided Mamba fusion module (PMFM), enabling adaptive joint optimization. TCEB employs a dual-MoE architecture for feature-specific extraction and spatial correction to achieve precise cross-modal alignment. Subsequently, PMFM applies a prompt-guided spatial-attention mechanism for selective information aggregation, enhancing reconstruction fidelity. Extensive experiments demonstrate that U-HSR outperforms state-of-the-art task-specific methods across diverse spatial morphologies and real-world acquisition settings.

Abstract:
Procedural materials provide editable, resolution-independent representations essential for virtual asset authoring. Yet existing methods mainly address appearance reconstruction or text-guided generation, where physical attributes, perceptual cues, and subjective impressions are entangled in a single description, causing ambiguous generation targets and limiting semantic control over executable procedural graphs. To address this gap, we propose SCHEMA (Semantic-Cognitive Hierarchical Engine for Material Authoring), a two-stage framework that separates semantic grounding from structural recovery. We support this decomposition with a four-level description framework spanning physical, perceptual, impressionistic, and emotional dimensions, and use it to construct a validated dataset of 529,558 image-description-code triplets, comprising 12,424 curated base triplets and 517,134 purified augmented triplets. A diffusion-based aligner first maps hierarchical descriptions to a rendered material appearance that preserves material structure while serving as a unified visual anchor, and a vision-language generator then translates this appearance into executable Blender procedural code. On the VLMaterial Blender test split, SCHEMA improves CLIP similarity by ?8.6% and DINOv2 structural similarity by ?20% over VLMaterial, while achieving 97.87% code executability on out-of-distribution data. In two blinded pairwise studies (n = 27 and n = 33), participants favored the structured hierarchy and the full pipeline on affective match and overall quality under the tested conditions.

Abstract:
Large Language Models (LLMs) have achieved remarkable success across a wide range of tasks. However, fine-tuning LLMs for Gloss-Free Sign Language Translation (GFSLT) remains a challenge. In this paper, we investigate how to effectively adapt LLMs to the GFSLT task. We show that there are two key issues that need to be solved: (1) the inherent distributional gap between visual feature inputs and text feature inputs makes it difficult for LLMs to interpret visual inputs; and (2) existing approaches typically concatenate visual and textual features in an autoregressive framework, which leads to the model overemphasizing textual inputs and deprioritizing visual cues, as LLMs are pretrained predominantly on text-centric data. To address the first challenge, we propose a simple yet effective method named Filtered Pseudo-Gloss CTC Pretraining, which leverages filtered pseudo-gloss sequences generated from text sequences to supervise the training of the visual backbone. To tackle the second issue, we introduce a Visual-Prioritized Distillation training strategy. Specifically, we define a visual-only prediction path in which text inputs are masked, and the model is required to generate the target sequence relying solely on visual inputs. To guide this path, the outputs from the standard visual-textual prediction are then distilled into the visual-only prediction path, encouraging the model to prioritize visual features. Comprehensive experiments and qualitative analyses demonstrate the effectiveness of the proposed model. The proposed SignLlama achieves very competitive performance on multiple datasets for GFSLT tasks, without using any extra modalities or external sign language datasets for pretraining.

Abstract:
Amodal segmentation aims to recover complete object shapes, including occluded regions, serving as an essential technique for user-centric multimedia authoring and object-level visual manipulation. Existing methods typically rely on informative prompts, such as bounding boxes or dense visible masks, which heavily degrade the user experience and interaction efficiency in real-world multimedia applications. While recent interactive paradigms (e.g., the Segment Anything Model) support lightweight point-based interactions, they often perform direct mask regression. Crucially, the opaque nature of these direct-regression models offers no visual explainability regarding how occluded structures are inferred, conflicting with the growing demand for interpretable multimedia systems. To address these limitations, we propose VELA, an end-to-end VElocity-driven Level-set Amodal segmentation method that enables explicit and transparent contour evolution driven by simple point clicks. VELA constructs an initial level set function from visual features and the user's point input, which then progressively evolves into the final amodal mask under the guidance of a shape-specific motion field predicted by a fully differentiable network. This mechanism learns to generate evolution dynamics at each step, ensuring that the spatial reasoning process is geometrically grounded, topologically flexible, and visually explainable to the user. Extensive experiments on COCOA-cls, D2SA, and KINS benchmarks demonstrate that VELA outperforms existing methods that use bounding-box or dense visible-mask prompts while requiring only a single-point prompt, validating the effectiveness of explainable geometric modeling for interactive multimedia tasks.

Abstract:
Radar-camera 3D object detection has shown great promise for autonomous driving, yet the inherent sparsity and noise of radar point clouds continue to hinder the detection performance. Existing methods typically adopt single-branch radar enhancement, which struggles to balance feature completeness and suppress noisy responses. To address this issue, we propose a dual-branch radar feature refinement framework that enhances radar representations from complementary perspectives. Specifically, a density-guided branch employs a statistically estimated density prior for sparse feature enhancement, while an intensity-guided branch utilizes a physically derived intensity prior for noise suppression. To offer richer semantic context, we further employ a camera-prior mask generated by spatial attention to enrich radar foreground features. Since radar refinement can further amplify the inherent cross-modal inconsistency, we design a deformable bidirectional alignment fusion module for cross-modal guided geometric alignment and adaptive feature fusion through bidirectional learnable offsets. In addition, a reliability-aware module generates reliable scores based on low-level visual cues to ensure robustness across diverse imaging conditions. Extensive experiments on the nuScenes and View-of-Delft( i.e., VoD) datasets demonstrate the effectiveness of the proposed framework. Compared to the baseline, our method improves mAP by 5.6% and NDS by 6.6% on the nuScenes dataset.

Abstract:
Detecting and segmenting tiny UAVs with event cameras in long-range scenarios requires recovering slender spatio-temporal trajectories despite complex background interference. However, target events are extremely sparse and often form thin, broken, and locally missing trajectories in the XYT space, while existing event-frame, voxel, and point-cloud methods typically make predictions on observed events or fixed grids. This discrete formulation is poorly matched to thin trajectories with gaps and missing segments. To address this challenge, we propose RidgeField-Net, a continuous field modeling framework for tiny UAV detection and segmentation in long-window event streams. Instead of predicting foreground labels only on observed events or discrete voxel locations, RidgeField-Net models the foreground as a continuously queryable implicit spatio-temporal tube probability field, enabling recovery of trajectory responses at unobserved coordinates. We further introduce the Local-Conditioned Implicit Tube Field (LITF), which preserves local structural cues of weak targets through sparse spatio-temporal encoding and continuous-coordinate querying. In addition, Patch-wise Skeleton Consistency (PSC) together with a volumetric sparsity constraint improves centerline consistency and suppresses spurious foreground responses in background regions. Experiments show that RidgeField-Net achieves stronger detection and segmentation performance than competing methods, while recovering more continuous tiny UAV trajectories under complex interference, weak responses, and long temporal windows.

Abstract:
Cross-view geo-localization (CVGL) remains challenging under limited field-of-view (FoV) conditions. A key reason is that restricted FoV reduces stable geometric structural information, making the model more susceptible to latent confounders that introduce spurious correlations between environmental context and localization target. In this work, we formalize this problem using a structural causal model and show that the standard observational learning objective is inherently biased; we then derive a causal intervention strategy to eliminate this bias. Building on this insight, we propose a feature-space diffusion network framework with causal intervention. Specifically, we decompose ground-view features into high-frequency structure-biased components and low-frequency context-biased components via frequency-domain decomposition and apply a batch-wise intervention strategy to block the confounding path. The resulting intervention-aware features guide a diffusion network that progressively refines noisy satellite features and aligns them with ground-view representations in latent space. In addition, we construct two multi-weather robustness benchmarks, CVUSA-MW and CVACT-val-MW, covering 15 weather conditions to support systematic robustness evaluation. Experiments on CVUSA and CVACT show that our method achieves state-of-the-art retrieval accuracy across multiple FoV settings, while demonstrating superior robustness under adverse weather and cross-domain scenarios.

Abstract:
Humans naturally map exocentric demonstrations to egocentric motion, and embodied agents likewise need this ability to retrieve view-invariant correspondences in the real world. In this paper, we propose the real-world cross-view video retrieval (RCV) task and establish the E2E-VR benchmark to simulate the bidirectional retrieval process of observation and imitation under realistic distractors. Methodologically, we introduce the salience-guided actional dual-encoder (SADE) architecture with salience-guided action tokenization (SAT) to extract sparse action evidence tokens from spatiotemporal token sequences and construct motion tokens via temporal differences computed on the same spatial patch across frames. To align these tokens and obtain viewpoint-agnostic dynamic cues by suppressing viewpoint-specific style, we design atomic action matching (AAM) through an interactive token-level matching scheme. This enables robust retrieval via partial evidence alignment, even under large cross-view appearance shifts, temporal misalignment, and strong distractors. To further enhance cross-view invariance, we introduce Tri-Objective Alignment (TOA), which combines multi-positive aggregation, bidirectional ranking consistency, and intra-modal cohesion regularization. Extensive experiments on E2E-VR show that SADE consistently outperforms strong baselines under bidirectional retrieval with realistic ambiguity and distractors on RCV tasks.

Abstract:
Micro-gesture recognition demands the detection of fleeting, spatially localized movements that are frequently overwhelmed by dominant static appearances and background noise. While Multimodal Large Language Models (MLLMs) excel at general video understanding, they inherently struggle with subtle kinematics and often rely on static posture priors. To this end, we propose GMoT, a Gated Motion-Aware Tokenization module that explicitly distills sparse kinematic evidence into a compact sequence prior to temporal modeling. GMoT dynamically spotlights action-relevant regions via spatially weighted pooling, extracts adjacent-frame temporal differencing to capture precise motion energy, and adaptively fuses these cues into the visual stream using a conservatively initialized semantic gate. To transition from simple classification to evidence-grounded reasoning, we further introduce a progressive reward-guided policy refinement paradigm, supported by a semi-supervised annotation pipeline that generates anatomically focused captions. Beyond achieving the best Top-1 accuracy among the compared methods on iMiGUE (67.32%) and SMG (73.11%), improving the Qwen3-VL-8B baseline by +6.80 and +3.11 points, our framework introduces Body-Region Grounding (BRG) Recall as an anatomical-grounding proxy conditioned on correct predictions, together with an overlapping-label cross-domain transfer protocol between iMiGUE and SMG. Extensive evaluations demonstrate that our GMoT-augmented model improves in-domain accuracy, retains clear gains under label-preserving corruptions, and improves accuracy-oriented cross-domain transfer under explicit small-split caveats while maintaining high anatomical grounding in its generated rationales. Code is available on GitHub.

Abstract:
Early Action Prediction (EAP) aims to classify ongoing actions from early partial observations, which is challenging because actions unfold dynamically over time. Estimating action progress provides a temporal anchor to align partial observations within the full action sequence, enabling more consistent modeling of action evolution. Leveraging this insight, we propose a novel paradigm termed "predict progress then complete future", which decomposes actions into temporal phases to facilitate a structured representation of action progression. Specifically, a monotonic phase predictor estimates the current action progress by deriving a monotonically ordered phase distribution, and further generates phase-level representations. Building upon this, we introduce a transition-aware phase completion mechanism that infers unobserved future phases based on phase-wise transition prototypes, reconstructing the evolution trajectory of action. Finally, a reliability-driven classifier emphasizes informative phases for robust prediction. Our method demonstrates clear improvements over existing methods on three benchmark datasets, with particularly significant gains at extremely low observation ratios.

Abstract:
Scene-aware human motion prediction requires balancing macroscopic global navigation with fine-grained local interactions. Existing methods often compromise by either heavily downsampling scenes at the expense of crucial geometric details, or processing the entire dense scene, which introduces massive computational overhead and perceptual noise. To overcome these limitations, we propose Aegis (Adaptive Egocentric Interactive Scene Modeling), a novel hierarchical framework designed to bridge the gap between global navigation and local fine-grained interaction. Aegis operates in a coarse-to-fine manner: first, a lightweight Sparse Global Trajectory Planner (SGTP) generates a coarse future path from sparse scene representations to provide long-range guidance. Subsequently, an Adaptive Egocentric Scene Extractor (AESE) dynamically distills motion-relevant sub-regions from the dense global point cloud. By adaptively adjusting the perception radius based on the human's instantaneous velocity and directional intent, AESE preserves critical local geometric features while filtering out global spatial noise. Finally, a Progressive Motion Decoder (PMD) autoregressively refines the coarse trajectory into high-fidelity motions. Extensive experiments demonstrate that Aegis achieves competitive prediction performance, effectively mitigating long-term trajectory drift while maintaining comparable pose accuracy and scene consistency.

Abstract:
Fine-Grained Ship Classification (FGSC) is crucial for maritime safety monitoring, traffic management, and anti-smuggling operations. However, due to the insufficiency of explicit target semantics, current mainstream vision-based approaches inevitably encounter the visual confusion problem when deployed in open-environment applications. Recently, Multimodal Large Language Models (MLLMs) have provided a transformative paradigm for enhancing target semantic understanding in FGSC, offering the potential to optimize visual representations through generation and alignment based on target textual knowledge. However, through careful inspection, we reveal that semantic description hallucinations and coarse alignment granularity are two significant challenges to reliably combine MLLMs with FGSC. To break aforementioned limitations, this paper introduces Anti-Confusion Semantic Generation and Alignment (ASGA), an efficient MLLMs-based FGSC approach containing two innovative modules, termed Multi-Agent Collaboration-driven Knowledge Extraction (MCKE) and Schrdinger Coupling-based Multimodal Alignment (SCMA). Specifically, MCKE seeks to automatically generate consistent and discriminative structured attribute priors for ship visual entities through multiple agents, whereas SCMA overcomes the limitations of traditional point-to-point hard matching by computing a global optimal coupling measure with soft alignments. Extensive experiments validate that our proposed ASGA method achieves remarkably superior performance with strong few-shot capacities over current competitive approaches, reaching 96.28% and 88.99% accuracy on two widely-used FGSC benchmarks, respectively.

Abstract:
Instruction-based image editing uses a planner-renderer pipeline: a vision-language model (VLM) converts the instruction into an edit plan; a diffusion model then executes it. Training such systems with only final-image rewards is inefficient because a poor edit does not reveal whether additional optimization should emphasize the planner or the renderer, and even planner-dominant cases remain difficult to localize within a free-form reasoning trace. We present \method, a reinforcement learning framework for dual-\vadjust\penalty-10000\linebreak level credit assignment in this two-stage setting. Across modules, multi-plan multi-render rollouts estimate between-plan and within-plan reward variability for soft module routing, while rollout mean rewards provide hardness estimates for an adaptive curriculum. Within the planner, a four-field structured reasoning output enables a prefix-gated reward and token-level advantage reweighting, turning outcome-level feedback into localized supervision. Experiments on five benchmarks show that \method outperforms a Joint~RL baseline with the same backbone, data, reward model, and rollout budget, with the largest gains on reasoning-intensive edits.

Abstract:
Vision-and-Language Navigation (VLN) is a challenging task that requires an agent to navigate through photorealistic environments following natural-language instructions. Recent research underscores the potential of harnessing large language models (LLMs) for VLN due to their powerful reasoning capabilities. Despite their strengths, a profound discrepancy persists between LLMs' immense magnitude and the scarcity of VLN data, fundamentally precluding them from precisely comprehending the physical world and thus bottlenecking their navigation performance. Though data augmentation is a promising way for scaling up the datasets, generating high-fidelity VLN environments remains an open challenge. To cope with this issue, we introduce GeoDream, a geometry-aware generative world model designed to synthesize VLN environments that exhibit both diversity and strict physical coherence, aimed at bolstering the spatial grounding of navigation agents. Our framework consists of two stages: the first trajectory stage, which maintains inter-viewpoint coherency by recursively warping 3D spatial priors via a point-cloud-based technique; and the second viewpoint stage, which guarantees intra-viewpoint and wraparound consistency by grounding visual synthesis in spherical projective manifolds. Moreover, we leverage the ControlNet-based diffusion architecture for visual rendering, ensuring every dreamed pixel is both photorealistic and strictly anchored to the underlying geometric priors. Finally, to achieve cross-modal coherence, we finetune a vision-language model to produce new instructions semantically aligned with the synthesized trajectories. Experimental results demonstrate that our synthesized data yields a 6.3% SPL improvement over the baseline, surpassing current state-of-the-art methods by a margin of over 2.0%. Beyond quantitative gains, qualitative evaluations also confirm that our new environments exhibit superior structural coherence and physical plausibility.

Abstract:
Post-training reinforcement learning (RL) algorithms are widely adopted to align large visionlanguage models (LVLMs) with human intent and the demands of visual reasoning tasks. However, existing RL-based alignment methods are often resource-intensive and encounter mismatches between training objectives and inference-time distributions. To bridge this gap, we propose a novel test-time alignment approach that leverages trajectory-guided structured sampling for dynamic inference-time refinement, enabling LVLMs to better align with visual grounding and maintain logical consistency. Our approach begins with curating a reasoning memory bank via a trajectory learning algorithm, which decomposes complex question solving into ordered sequences of predefined reasoning patterns. It subsequently achieves inference-time alignment by first collecting trajectories from reasoning memory bank to establish a global structural reasoning prior, and then using an iterative Markov Chain Monte Carlo (MCMC) algorithm for localized multi-objective refinement of the reasoning trace. Experiments across multiple multimodal reasoning datasets demonstrate that our approach significantly improves accuracy without incurring prohibitive inference overhead. These results establish trajectory-guided test-time sampling as a scalable and effective alternative to traditional RL-based post-training alignment, particularly for complex visual reasoning tasks.

Abstract:
Continual Test-Time Adaptation (CTTA) aims at maintaining stable and accurate inference in continually shifting environments, where source data is unavailable and target domain distributions evolve over time. Existing methods mainly focus on prediction confidence to guide model adaptation, which easily causes semantic bias as incorrect predictions with high confidence are taken as correct, and leads to representation drift as low-confidence samples lack proper supervision. It is crucial to avoid miscalibrated predictions and unreliable pseudo-labels, which are not well explored by previous domain adaptation methods. In this work, we propose Divide and Correct (DAC), a lightweight CTTA framework that partitions target samples into four distinct quadrants using an online data subdivision protocol based on confidence and prototype misalignment. For low-confidence yet prototype-consistent samples, DAC calibrates ambiguous posteriors via a prior-aware alternating normalization driven by a running class prior estimated from reliable anchors. For confident but prototype-inconsistent samples, DAC suppresses harmful updates using temperature-scaled self-training and a prototype-margin objective that pulls features away from confusable negatives. Extensive experiments on standard and non-i.i.d. CTTA benchmarks demonstrate that DAC achieves state-of-the-art performance and remains stable under temporally correlated label shifts.

Abstract:
Recent generative video models achieve impressive visual quality but remain constrained by limited physical consistency and controllability. Existing video generation methods provide minimal physical control, and single-image-to-3D conversion approaches often suffer from object interpenetration. Furthermore, physics-based scene-level 3D generation methods exhibit spatial misalignment, stylized artifacts, and inconsistencies with the input data, restricting their use in realistic interactive video synthesis. We propose PhysOmni, a training-free framework that converts a single image into a physically consistent and controllable video through holistic scene-level 3D reconstruction. By representing the full scene geometry in a unified spatial coordinate system, PhysOmni resolves object penetration and alignment ambiguity. Unlike prior methods, this formulation enables accurate scene-level multi-object interactions and introduces richer, complex control types for advanced mechanics-based manipulation. By decoupling simulation from rendering, PhysOmni bypasses latency-heavy priors, achieving real-time physical interaction previews paired while preserving photorealistic visual fidelity. Experimental results demonstrate that PhysOmni substantially outperforms prior methods in physical fidelity, spatial coherence, and controllability.

Abstract:
Image steganography, particularly coverless image steganography (CIS), has emerged as a powerful means for covert communication without modifying the cover image, thereby evading conventional passive steganalysis. Existing active defense strategies against CIS primarily rely on traditional image processing operations such as rotation, cropping, and adding noise. However, modern CIS algorithms exhibit strong robustness against these operations. Moreover, such defensive manipulations frequently degrade image visual quality, potentially arousing suspicion during transmission. To address these limitations, we propose CIS-Guard, a novel preemptive defense algorithm based on adversarial perturbations. To the best of our knowledge, this is the first work dedicated to defending against CIS algorithms. Unlike traditional steganalysis methods that only detect the presence of hidden information, CIS-Guard proactively injects adversarial perturbations into stego-images to disrupt the correct extraction of secret information at the receiver. Furthermore, we introduce the visual callback loss (VCL) to enhance the visual quality. Experimental results demonstrate that CIS-Guard significantly reduces the extraction accuracy of state-of-the-art CIS methods while maintaining superior visual quality.

Abstract:
Modern DNNs face fluctuating resource constraints in edge deployments across diverse multimedia applications. Existing dynamic networks struggle with parameter redundancy from storing multiple sub-networks and lack flexibility due to predefined widths. To address this, we propose a layer-wise decoupled dynamic network paradigm via hierarchical dynamic filter generation. Our core innovation, intra-layer decoupling, dynamically generates structurally diverse and arbitrary-width convolutional filters within each layer from a compact shared parameter space. Our decoupled generation mechanism eliminates the need for multiple sub-networks by integrating structural adaptability into the generation process, enabling low memory cost and smooth capacity transitions. We further introduce a unified training scheme that jointly optimizes multiple sub-networks via in-place knowledge distillation and rank-aware parameter regularization, ensuring consistent performance and training stability across all width configurations. Extensive experiments on Kinetics-Sound and CREMA-D (under a unified robustness protocol) demonstrate strong accuracy-efficiency trade-offs among dynamic models, with the dynamic filter generation achieving 10× parameter reduction while maintaining competitive performance across varying widths, supporting efficient dynamic inference for multimedia scenarios.

Abstract:
Recently, multimodal typographic adversarial attacks have raised critical safety concerns about Large Vision-Language Models (LVLMs). By embedding semantically coherent but misleading text into images, these attacks reveal severe vulnerabilities of LVLMs. Existing defenses primarily rely on text filtering or input purification, which fail to capture cross-modal inconsistencies and often disrupt visual integrity. To this end, we propose SAGE (Scene-Aware Guard Engine), a novel defense framework that fundamentally shifts from adversarial content removal to reasoning-level semantic control. Specifically, we introduce a scene-aware semantic consistency detector that identifies adversarial text that contradicts visual context under the query. Subsequently, an explicit semantic separation mechanism is employed to isolate misleading textual signals from visual reasoning. Furthermore, we develop a self-reflective optimization module that adaptively refines defense strategies through failure-driven feedback. Extensive experiments across diverse attack methods, models, and benchmarks demonstrate that SAGE significantly improves defense success rates while maintaining competitive performance on benign samples, exhibiting outstanding robustness and generalizability.

Abstract:
Reinforcement learning, particularly Group Relative Policy Optimization (GRPO), has emerged as an effective framework for post-training visual generative models with human preference signals. However, its effectiveness is fundamentally limited by coarse reward credit assignment. In modern visual generation, multiple reward models are often used to capture heterogeneous objectives, such as visual quality, motion consistency, and text alignment. Existing GRPO pipelines typically collapse these rewards into a single static scalar and propagate it uniformly across the entire diffusion trajectory. This design ignores the stage-specific roles of different denoising steps and produces mistimed or incompatible optimization signals. To address this issue, we propose Objective-aware Trajectory Credit Assignment (OTCA), a structured framework for fine-grained GRPO training. OTCA consists of two key components. Trajectory-Level Credit Decomposition estimates the relative importance of different denoising steps. Multi-Objective Credit Allocation adaptively weights and combines multiple reward signals throughout the denoising process. By jointly modeling temporal credit and objective-level credit, OTCA converts coarse reward supervision into a structured, timestep-aware training signal that better matches the iterative nature of diffusion-based generation. Extensive experiments show that OTCA consistently improves both image and video generation quality across evaluation metrics.

Abstract:
\beginabstract While CLIP excels at zero-shot recognition, its accuracy degrades sharply under common corruptions such as noise, blur, weather effects, and digital distortions. To mitigate this issue, existing test-time adaptation (TTA) methods adapt CLIP during inference, but they are largely developed and evaluated under isolated-corruption protocols, where each corruption type is treated as an independent test condition. These methods can fail when handling mixed online test streams, in which corruption types can change frequently across batches and adaptation errors can accumulate over time. Motivated by this limitation, we introduce Hierarchical Mixed Corruption Streams (HMCS), a protocol family that systematically evaluates CLIP-TTA under progressively broader within-family, family-pair, and global mixing scenarios. We further propose \oursfull (\ours), a CLIP-TTA framework that decouples batch-local adaptation from persistent non-parametric state while coordinating them through shared online statistics. Built on a batch-local adaptation step, \ours improves long-horizon stability through filtered memory updates, shift-aware stability control, and quality-aware prediction fusion. Extensive experiments on CIFAR-10-C, CIFAR-100-C, and ImageNet-C-5K show that \ours consistently outperforms strong CLIP-TTA baselines, suggesting that under mixed corruption streams with frequent corruption changes, effective adaptation depends not only on the current batch but also on preserving the reliability of accumulated state. \endabstract

Abstract:
Large Vision-Language Models (LVLMs) have achieved strong performance in visual understanding, but hallucination remains a critical challenge. Existing alignment methods such as Direct Preference Optimization (DPO) often suffer from unstable optimization under off-policy settings due to the mismatch between model policies and fixed preference data. To address this issue, we propose Dual-Direction Preference Optimization (DPO), an adaptive distribution alignment framework for stable preference learning in LVLMs. DPO introduces an Adaptive Distribution Contraction (ADC) module to align model generation distributions with high-quality regions and an Entropy-guided Hard Sample Discovery (EHSD) strategy to identify hallucination-prone samples for targeted correction. Experiments on MMHalBench, ObjectHalBench, AMBER, and POPE Adv demonstrate that DPO consistently outperforms DPO-based methods on 7B and 13B LVLMs with only 3.1K preference samples, reducing hallucination rates by 4050% and achieving state-of-the-art performance in MLLM hallucination mitigation.

Authors: Xueyang Zhou (Huazhong University of Science and Technology), Zijia Wang (Huazhong University of Science and Technology), Guiyao Tie (Huazhong University of Science and Technology), Sizhe Zhang (Huazhong University of Science and Technology), Junran Wu (Huazhong University of Science and Technology), Hecheng Wang (Huazhong University of Science and Technology), Xu Yongtian (Huazhong University of Science and Technology), Zhichao Ma (Huazhong University of Science and Technology), Yan Zhang (Huazhong University of Science and Technology), Xiangyu Zhang (Huazhong University of Science and Technology), Yongchao Chen (Tsinghua University), Pan Zhou (Huazhong University of Science and Technology), Lichao Sun (Lehigh University)

Abstract:
As household robots move from controlled demonstrations into messy, unpredictable homes, trust-related failures in physical-world interactions can cause serious real-world harm, including safety hazards. To enable risk-grounded evaluation of embodied trustworthiness, we analyze widely reported robot failures and abnormal household incidents, distilling 9 representative failure modes that capture recurring violations of human trust. We ground our benchmark in 7 real-world household scenarios and systematically instantiate these failure modes to construct 150 fine-grained interactive evaluation tasks in EB-Habitat. We evaluate embodied trustworthiness along five dimensions: Truthfulness, Safety, Fairness, Robustness, and Privacy, and introduce HoRoTrust as a unified benchmark for interactive household assessment. Experiments on 19 state-of-the-art MLLMs reveal pervasive weaknesses, particularly in privacy protection and proactive risk avoidance. We further find that general capability does not positively correlate with trustworthiness, while explicit reasoning traces provide little benefit. Overall, HoRoTrust offers an interactive, risk-grounded platform for diagnosing trust failures in household embodied agents.

Abstract:
Vision-language model (VLM)-based game agents have shown considerable promise. However, although these agents generally perform well in the early stages of a game, their success rates decline substantially as task difficulty increases. As a result, enabling an agent to solve advanced levelsor even complete an entire gameremains a formidable challenge. In fact, human players often consult external walkthroughs to overcome difficult levels. For VLM-based agents, however, the multimodal, unstructured, and often redundant nature of such walkthroughs presents significant obstacles. In this paper, we propose Walkthrough Agent, a novel framework that enables agents to adaptively identify relevant clues from redundant walkthroughs and, through reasoning, derive appropriate actions for successful execution, thereby completing more challenging game levels. Specifically, the framework consists of three modules: 1) Gather, which automatically collects raw multimodal walkthroughs; 2) Reasoner, which employs a three-stage strategylocating, reasoning, and actionto progressively identify relevant clues and infer appropriate actions; and 3) Executor, which directly applies the generated action commands to interact with the game environment. Extensive experiments across three game environments demonstrate that our framework substantially improves success rates on challenging levels. Moreover, using only 1,000 verified samples to fine-tune Qwen3-VL-8B, our method substantially outperforms the same-backbone RAGEN baseline and achieves competitive performance relative to closed-source models.

Abstract:
Current generative mural restoration models prioritize visual continuity over historical authenticity, which causes ungrounded hallucinations and relies on synthetic datasets lacking real-world complexity. To address these issues, we introduce a novel evidence-reasoning-driven paradigm that models digital restoration as a traceable, consensus-based decision process. Under this paradigm, we first construct a new benchmark for the Kizil Grottoes, providing authentic degradations, expert-verified ground truths, and multimodal evidence chains. Furthermore, we propose Mural-Reason, an evidence-grounded framework using Multimodal Retrieval-Augmented Generation (MRAG) to synthesize unaligned text-image evidence into structured restoration plans. To execute these plans without structural drift, our Evidence Reasoning Adapter (ERA) employs uncertainty-aware spatial fusion and adaptive temporal scheduling. Experiments demonstrate that our method significantly improves upon baselines in both perceptual quality and historical fidelity.

Abstract:
Magnetic Resonance Imaging (MRI) provides rich diagnostic information through multi-contrast scans, but acquiring all contrasts significantly extends scan time. To break the acceleration limits, we tackle the highly ill-posed compound inverse problem of reconstructing all target contrasts from a highly undersampled and partially missing input contrasts. In this work, we propose a Missing-aware Latent Diffusion Model (MissLDM), which jointly resolves contrast synthesis and k-space reconstruction within a disentangled latent space. Specifically, MissLDM comprises three core components: (1) a Contrast Disentangled Network (CDN) that encodes available inputs into a contrast-invariant latent representation and contrast-specific codes; (2) a Multi-Contrast Invariant Latent Diffusion Model (MILDM) that learns a generative prior over the invariant latent space, enabling the sampling of plausible latent representations from Gaussian noise; and (3) a Measurement-Consistent Refinement (MCR) module that enforces data consistency during reverse diffusion. Furthermore, we collect uBrain, a raw k-space aligned multi-contrast MRI dataset comprising 28,563 2D slices, to evaluate reconstruction quality under realistic clinical conditions. Extensive experiments on uBrain and multiple public benchmarks demonstrate that MissLDM outperforms state-of-the-art methods in both reconstruction quality and computational efficiency.

Abstract:
Learning animatable head avatars from a short monocular video is attractive for expression-driven and pose-driven rendering, yet recovering fine subject-specific appearance from such input remains difficult. Existing Gaussian head avatar methods are efficient and practical, but many still learn point-wise appearance with weak constraints under monocular ambiguity, which often leads to over-smoothed results. This work presents HiFiAvatar, a monocular head avatar framework that improves appearance reconstruction by introducing a frozen subject-specific visual prior. The method first aggregates pretrained visual features extracted from the training video into a compact identity and texture dictionary. Each Gaussian primitive then queries this dictionary from canonical space, fuses the retrieved cue with a learnable point embedding, and predicts its color through a lightweight surface-space decoder before standard Gaussian rendering. Because the prior is fixed during optimization, it acts as a stable regularizer that narrows the appearance solution space and reduces over-smoothing. Experiments on four benchmark datasets show that the proposed method achieves the best average results on three datasets and the best structural similarity and perceptual quality on the fourth, while preserving the efficiency of Gaussian avatar rendering. These results demonstrate more faithful recovery of high-frequency facial appearance from monocular video.

Abstract:
Portrait sketching is not only a foundational training for art professionals but also a crucial way for self-learners to hone their modeling capabilities. However, providing objective, multi-dimensional feedback remains a significant challenge for learners lacking professional guidance, as self-assessment often fails to identify specific structural or lighting defects due to subjectivity. Existing automated assessment methods frequently overlook the ``global-governing-local" cognitive logic essential in art evaluation and are prone to overfitting on small-scale datasets. To address these issues, this paper constructs the first fine-grained, multi-attribute diagnostic assessment dataset for portrait sketches (FG-SHPD). This dataset comprises 1,339 high-quality images with consensus-based multi-dimensional annotations across five fine-grained attributes: composition, structure, light, details \& texture, and artistic appeal, effectively breaking the data bottleneck in this field. Building upon this, we propose a novel intelligent scoring framework based on the Star-shaped Graph Interaction Network (SGIN). First, a domain-specific self-supervised pre-training strategy tailored to sketching characteristics is introduced to bridge the domain gap in feature extraction. Second, a star-shaped topology centered on global features is designed, forcing the assessment of local attributes to be routed and calibrated through the global context, which mimics the professional evaluative mindset of ``global-governing-local." Finally, an ordinal regression mechanism is incorporated to capture the ordinal semantics of discrete scores, enabling a joint assessment of multi-attribute fine-grained diagnosis and overall continuous scores. Experimental results demonstrate that the proposed method significantly outperforms existing aesthetic assessment baselines, achieving an SROCC of 0.8841 in total score prediction and an average QWK of 0.8296 across multi-dimensional attributes. This work provides an objective and efficient pedagogical assistant for self-learners in sketching. The source code is publicly available at \urlhttps://gitee.com/yongzhenke/sgin.

Abstract:
Multimodal intent recognition (MIR) aims to infer user intentions by jointly modeling linguistic, visual, and acoustic cues. However, prevailing text-dominated fusion paradigms suppress the information expression of non-verbal signals, leading to misinterpretation when linguistic information is ambiguous or insufficient. To overcome these limitations, we propose AMG-CPS, an Adaptive Modality Gating and Contrastive Prototype Shaping Network that moves beyond the text-dominance and enhances multimodal understanding in MIR. Since the non-verbal information density varies drastically across samples, the Adaptive Modality Gating module performs hierarchical feature filtering, adaptively retaining informative features while suppressing redundancy. Furthermore, to enhance the discriminative distribution of the embedding space, we introduce a Contrastive Prototype Shaping module that aligns multimodal representations with learnable intent prototypes rather than relying solely on the text modality. It enforces intra-class compactness and cross-modal semantic consistency, forming well-separated and semantically coherent class distributions. Within two mechanisms, AMG-CPS fundamentally allows the model to dynamically adapt its reliance on diverse modalities and coherently interpret intent across varying contexts. Extensive experiments on two challenging benchmarks, MIntRec and MIntRec2.0, demonstrate that AMG-CPS outperforms state-of-the-art methods for robust intent recognition.

Abstract:
Continual face forgery detection (CFFD) requires detectors to learn emerging forgery paradigms without forgetting previously seen manipulations. Under strict memory budgets, existing replay-based CFFD methods either retain sparse historical faces, limiting forgery-cue coverage and raising privacy concerns, or use detector-dependent perturbations tied to past decision boundaries. To improve the representational capacity of replay while avoiding raw-face storage, we introduce Distribution-Discrepancy Condensation (DDC) to directly condense real-to-fake distribution discrepancies and Manifold-Consistent Replay (MCR) to perform distribution-level replay with current real faces. Specifically, DDC models the real-to-fake discrepancy via a surrogate factorization in characteristic-function space and condenses it into a tiny bank of distribution discrepancy maps. MCR synthesizes replay samples through variance-preserving composition of these maps with current-stage real faces, yielding samples that reflect previous-task forgery cues while remaining compatible with current real-face statistics. On two CFFD benchmarks, our framework substantially reduces forgetting and achieves the best overall results. As a plug-in replacement, our replay also consistently improves five replay-based detectors.

Abstract:
Eye-based emotion recognition enables eyewear devices to perceive users emotional states and support emotion-aware interaction. However, deploying such functionality on their resource-limited embedded hardware remains challenging. Time-to-first-spike (TTFS)-coded spiking neural networks (SNNs) offer a promising solution due to their extremely sparse and energy-efficient computation, where each neuron emits at most one binary spike. While prior works have primarily focused on improving TTFS SNN training algorithms, the role of network architecture has been largely overlooked. This is particularly critical, as spike timing in TTFS SNNs is tightly coupled with architectural design, and eye-based emotion recognition requires compact yet highly efficient networks. In this paper, we propose TNAS-ER, the first neural architecture search (NAS) framework tailored to TTFS SNNs for eye-based emotion recognition. TNAS-ER presents a novel ANN-assisted search strategy that leverages a ReLU-based ANN counterpart to guide architecture optimization and stabilize training of the TTFS SNN. TNAS-ER employs an evolutionary algorithm, with weighted and unweighted average recall jointly defined as fitness objectives for emotion recognition. Extensive experiments demonstrate that TNAS-ER achieves high recognition performance with significantly improved efficiency. Furthermore, we evaluate TNAS-ER on neuromorphic hardware, confirming its superior energy efficiency and strong potential for real-world applications.

Abstract:
The rapid evolution of video generation has shifted the paradigm from pure text-driven to multi-conditional controllable generation, with reference images now widely adopted as conditional inputs to achieve superior spatiotemporal consistency. While these reference images serve as powerful visual anchors that significantly enhance controllability, their impact on safety remains largely unexplored. In this work, we reveal the visual anchoring effect: by enforcing consistency, the mechanism prevents the generated content from drifting away from the original harmful intent, thereby eliminating the model's natural safety escape route from harmful to benign content. Consequently, visual anchors inherently increase the safety risk---this is the price of consistency. Building on this insight, we propose Decoupling Intent via Visual Anchors (DIVA), a training-free multimodal jailbreak framework for video generation that exploits this vulnerability. DIVA decouples harmful intent into a static visual anchor image and a dynamic motion text prompt, and employs dual-criteria selection to balance attack stealthiness with semantic preservation. Extensive experiments across various leading commercial platforms and mainstream open-source video generation models demonstrate that DIVA achieves a substantially higher Attack Success Rate than existing text-only methods. To facilitate future research, we additionally contribute TI2VSafetyBench, the first safety benchmark for multi-conditional video generation.

Authors: Fangyu Hao (Beijing University of Post and Telecommunications), Jiayu Yang (Beijing University of Post and Telecommunications), Yifan Zhu (Beijing University of Post and Telecommunications), Zijun Yu (Beijing University of Post and Telecommunications), Qicen Wu (Beijing University of Post and Telecommunications), Yunlong Wang (Beijing University of Post and Telecommunications), Jiawei Li (bupt), Yulin Liu (Beijing University of Post and Telecommunications), Xu Zeng (Beijing University of Post and Telecommunications), Guanting Chen (Beijing University of Posts and Telecommunications), Shihao Li (Beijing University of Post and Telecommunications), Zhonghong Ou (Beijing University of Posts and Telecommunications), Meina Song (Beijing University of Posts and Telecommunications), Mengyang Sun (Tsinghua University), Haoran Luo (Nanyang Technological University), Yu Shi (Peking Union Medical College Hospital), Yingyi Wang (Peking Union Medical College Hospital)

Abstract:
Lung cancer clinical decision support requires precise reasoning across complex, multistage oncology workflows, yet existing multimodal large language models (MLLMs) struggle with guideline-constrained staging and treatment decisions. We formalize three oncological precision treatment (OPT) tasksTNM staging, treatment recommendation, and end-to-end clinical decision supportand introduce LungCURE, the first standardized multimodal benchmark comprising 1,000 clinician-labeled real-world cases from over 10 hospitals. We propose LCAgent, a multi-agent framework that ensures guideline-compliant decision-making by suppressing cascading reasoning errors across the clinical pathway. Experiments reveal substantial differences in complex medical reasoning across large language models (LLMs) under precise treatment requirements and show that LCAgent, as a simple yet effective plug-in, improves reasoning performance in real-world medical scenarios.

Abstract:
Text-guided visual token pruning mitigates the computational burden of Large Vision-Language Models by retaining only the image patches that are relevant to the query. Most existing paradigms rely on cosine similarity to assess the relevance of visual tokens. However, our analysis reveals that this raw similarity signal carries systematic nuisance baselines that misrank candidates before any selection takes place. To overcome these limitations, we propose TEDPrune, a training-free pruning pipeline. First, Double-Centered Correction debiases the raw similarity matrix to isolate query-relevant regions. Second, Energy Distance formulates token retention as a distribution-matching problem and refines the debiased pool with a set-dependent criterion. This two-stage approach yields a refined subset that aligns with relevant areas and restores the coverage lost to redundancy. Furthermore, Double-Centered Correction functions as an independent module that seamlessly improves existing text-guided methods, providing a robust baseline for future text-guided pruning. Extensive experiments across multiple LVLM backbones demonstrate that TEDPrune achieves a strong overall trade-off between computational efficiency and generation quality under aggressive pruning.

Abstract:
Despite the success of Vision-Language Models (VLMs), misleading charts remain a significant challenge due to their deceptive visual structures and distorted data representations. We present ChartCynics, an agentic dual-path framework designed to unmask visual deception via a ``skeptical'' reasoning paradigm. Unlike holistic models, ChartCynics decouples perception from verification: a Diagnostic Vision Path captures structural anomalies (e.g., inverted axes) through strategic ROI cropping, while an OCR-Driven Data Path ensures numerical grounding. To resolve cross-modal conflicts, we introduce an Agentic Summarizer optimized via a two-stage protocol: Oracle-Informed SFT for reasoning distillation and Deception-Aware GRPO for adversarial alignment. This pipeline effectively penalizes visual traps and enforces logical consistency. Evaluations on two benchmarks show that ChartCynics achieves 74.43% and 64.55% accuracy, providing an absolute performance boost of ～29% over the Qwen3-VL-8B backbone, outperforming state-of-the-art proprietary models. Our results demonstrate that specialized agentic workflows can grant smaller open-source models superior robustness, establishing a new foundation for trustworthy chart interpretation.

Abstract:
Multi-turn GUI agents enable complex task completion through sequential decision-making, but suffer from severe context inflation as interaction history accumulates. Existing strategies either sacrifice long-term context via truncation or compromise spatial structure through token pruning. In this paper, we propose Coordinate Compression Policy Optimization (CCPO), an efficient policy optimization framework that couples visual compression with policy optimization for multi-turn GUI agents. CCPO introduces Coordinate-Aware Spatial Compression (CASC), which aggregates coordinates from multiple rollouts to capture target-relevant regions and progressively narrow historical attention around key visual areas. This mechanism ensures that the model retains high-frequency interaction details while discarding static background information. From interactions across rollouts, CASC adaptively constructs attention boundaries that concentrate computation on the most informative regions of the scene. This mechanism creates a virtuous cycle where focused inputs facilitate precise predictions, subsequently tightening the spatial boundaries for future steps. We further design a Distance-Based Advantage that provides fine-grained learning signals based on distance rather than binary correctness, improving both grounding accuracy and compression quality. Extensive experiments demonstrate that CCPO achieves state-of-the-art performance across four benchmarks with up to 55% token compression and 3.8x training speedup.

Abstract:
In medical visual question answering (VQA), hallucinations of vision-language models (VLMs) may lead to confident but incorrect responses, raising the risk of diagnostic errors. Existing hallucination detection methods uniformly estimate the reliability of VLM outputs from response consistency or visual evidence. However, such uniform verification across questions ignores question-specific characteristics, resulting in missed overconfident errors and false alarms from over-verification. We present PROVE (Proof-guided Regime-aware Operator VErification), a black-box detector that adapts verification strategy to the evidential structure of each question. PROVE classifies questions into three verification regimes based on what kind of visual proof they demand, activates a regime-specific subset of five complementary operators, and adjusts operator importance per question through a lightweight calibration layer conditioned on deterministic question-answer features. PROVE uses question-specific evidence to reweight operators and produce a calibrated risk score. Evaluated on 8,048 test samples across three medical VQA benchmarks and four frontier VLMs, PROVE achieves 0.821 AUROC, outperforming the strongest baseline by +0.159, with consistent gains across all models and benchmarks.

Abstract:
Multimodal table understanding (MTU) has achieved strong performance on clean tables, but does not generalize well to real-world photographic scenarios. The core challenge is that wild table images often suffer from both content degradation (e.g., shadows, textures, blur), which impairs the readability of textual and numerical content, and structure degradation (e.g., paper bending, warping, perspective distortion), which corrupts row-column alignment. To address this challenge, we decouple wild table understanding into content modeling and structure modeling. We build a controllable degradation pipeline for disentangled diagnosis of content and structure degradation and for wild-style data synthesis, and also construct a real-world benchmark of wild table images for evaluation in realistic settings. Building on this foundation, we propose WildTab, a content-structure decoupled framework that combines Content-Consistent Visual Distillation for robust degraded-content perception with a Structure-Aware Deformable Projector that reorganizes distorted visual evidence to recover more reliable row-column correspondence before projection to the LLM. WildTab is further optimized with a progressive three-stage training paradigm that sequentially enhances content robustness, structure-aware perception, and downstream task understanding. Extensive experiments on both synthetic and real-world benchmarks show that WildTab consistently outperforms existing general-purpose MLLMs and specialized MTU models, highlighting that robust wild table understanding requires both accurate perception of degraded content and reliable preservation of table structure under geometric distortion.

Abstract:
Continuous sign language recognition is fundamentally constrained by spatiotemporal biases: insufficient adaptability to heterogeneous motion scales and poor spatial semantic discriminability. To explicitly disentangle how a sign is performed from where it occurs, we propose STAR, a causality-inspired Spatiotemporal Adaptive Rectification framework. Temporally, the Temporal Dynamic Rectifier (TDR) employs a mixture-of-experts mechanism to adaptively capture diverse motions, ranging from rapid finger articulations to broad arm trajectories. Spatially, the Spatial Causal Rectifier (SCR) leverages causal counterfactual intervention to suppress non-causal background cues, focusing strictly on semantic regions like hands and faces. Extensive experiments on PHOENIX14, PHOENIX14-T, and CSL-Daily demonstrate that STAR achieves new state-of-the-art performance without additional inference latency, ensuring robust semantic grounding and resilience to contextual interference.

Abstract:
Infrared--visible image fusion seeks to retain structures shared by the two modalities while preserving their complementary cues. The appropriate fusion rule, however, is not uniform across spatial locations or feature scales. Regions in which the modalities agree mainly require structure preservation, whereas regions with large cross-modal differences benefit from stronger complementary enhancement. We therefore introduce a coarse-to-fine framework with two coupled modules: an Uncertainty-Aware Dynamic Prompt Router (UDPR) and a Task-Tailored Bidirectional Vision Mamba (TT-BVM) block. UDPR predicts hierarchical uncertainty maps and uses them to route prompts across scales. TT-BVM uses the routed prompt and uncertainty map to separate common and discrepancy features, then models the two branches with dedicated bidirectional state-space blocks. The model is trained with an uncertainty-conditioned reconstruction loss and auxiliary terms that regularize routing and branch coordination. When trained only on MSRS, it obtains the best score in 16 of 20 metric entries across four public infrared--visible fusion benchmarks. The same model also performs competitively on three medical fusion benchmarks without retraining and improves semantic segmentation and object detection on fused images. These results support the uncertainty-guided design, and the mechanism analysis shows stable routing and branch behavior in the evaluated setting.

Abstract:
Existing embodied AI benchmarks often focus on the final stage of the pipeline (e.g., instruction following or question answering), implicitly assuming perfect perception and overlooking the upstream challenges of active exploration and representation building. They are also typically confined to room-scale settings with short, object-centric instructions, and thus fail to capture the complexity of building-scale tasks that require reasoning over global layouts across multiple rooms and floors. To address these gaps, we introduce ERNav, the first unified benchmark that integrates Exploration, Representation, and Navigation into an end-to-end embodied pipeline. ERNav requires an agent to actively explore an unseen environment, construct a global representation from partial, egocentric RGB-D observations, and then localize targets in building-scale scenes given complex natural language instructions. We further propose 3D-LangNav, a strong baseline that combines a dual-sighted exploration strategy for high-quality 3D mapping with an LLM-based spatial reasoning module for language grounding. Experiments show that ERNav is highly challenging for existing approaches, while 3D-LangNav achieves strong performance, more than doubling the success rate of state-of-the-art 3D-MLLMs. By decomposing the pipeline into sequential subtasks, ERNav also enables fine-grained diagnosis of how exploration and representation quality affect downstream navigation.

Abstract:
In surgery, patient safety is threatened not only by technical issues but also by poor teamwork. However, existing surgical AI-based solutions focus mainly on visual workflow and technical execution, neglecting the modeling of team interactions and missing opportunities to actively support clinicians in improving their teamwork skills. To address this gap, we propose a tempo-relational framework for modeling surgical team dynamics from multimodal observations. By leveraging Time-Expanded graphs, the approach captures both relational structure and temporal evolution, achieving strong expressivity while remaining robust in the low-data regime typical of surgical settings. Beyond prediction, such modeling enables the generation of efficient, interpretable, and actionable suggestions for clinicians. More specifically, we generate suggestions via a counterfactual procedure that identifies minimal yet structured changes in individual behaviors and interaction patterns associated with improvements in team performance. Experiments with simulated surgical procedures show that our approach improves predictive performance in diverse behavioral and interaction goals while offering meaningful insights into team dynamics. This work advances surgical AI beyond outcome-driven prediction towards a socially grounded, team-centric, and actionable paradigm to better understand and support the development of team skills in surgical settings.

Abstract:
Vision Graph Neural Networks (Vision GNN, ViG) have demonstrated strong representational capabilities in visual tasks by modeling images as graph structures. However, effectively transferring pre-trained ViG models to downstream tasks via parameter-efficient visual prompt learning remains a significant challenge. Existing visual graph prompt methods share a common critical limitation: the generation of prompt signals is decoupled from the visual semantic content of input samples, which fundamentally restricts the ability of prompts to perceive discriminative information embedded in the visual graph structure. Specifically, this limitation manifests in two aspects. First, prompt parameters in existing methods are typically input-agnostic learnable vectors that cannot be dynamically adjusted according to the visual discriminative content of different samples during inference, thereby constraining the capacity of prompts to model discriminative information within the graph structure. Second, low-rank feature extraction in prompt designs predominantly relies on simple convolutional linear projections, lacking targeted modeling of semantically critical dimensions in the feature space, which hinders the effective capture of core semantic information in the visual graph. To address the above issues, we propose the Prompts That Understand Graphs (PTUG) framework. Specifically, we design the Dynamic Discriminative-Aware Prompt Injector (D2API) module, which introduces graph-level cross-patch attention and graph convolutional modeling, combined with local and global discriminative branches, to dynamically generate discriminative-aware prompts conditioned on each input sample, thereby alleviating the decoupling between static prompts and input semantics. Furthermore, we propose the SVD-Inspired Structured Semantic Prompter (S3P) module, which replaces conventional convolutional projections with SVD-inspired structured low-rank semantic prompting based on subspace projection, dimension-wise scaling, and basis transformation, thereby enhancing the expressive capacity of prompt features over core semantic dimensions. Extensive experiments on multiple visual classification benchmarks demonstrate that the proposed method significantly improves the transfer performance of pre-trained ViG models while maintaining high parameter efficiency, achieving performance on par with or superior to full fine-tuning across multiple datasets.

Abstract:
Anomaly edge detection in dynamic graphs is critical to applications like social networks, e-commerce, and financial systems. However, existing methods often decouple spatial and temporal features, losing joint correlations, and struggle to define robust decision boundaries due to noisy pseudo-labels and class imbalance. To address these issues, we propose a novel dynamic graph anomaly detection method, ST-DiffDyG. The method employs a Spatio-Temporal Unified Graph Representation (STUGR) module, which utilizes a unified spatio-temporal encoding strategy and position-aware attention to capture intricate dependencies and generate highly expressive edge representations. Meanwhile, a Calibrated Diffusion Anomaly Scoring (CDAS) module is introduced, which implements a calibrated dual diffusion framework to model normal and anomalous distributions separately. It employs a novel calibration loss to rectify the inherent noise in pseudo-anomalous labels and utilizes Bayesian inference for accurate anomaly scoring. Experimental results on multiple real-world datasets show that our ST-DiffDyG method outperforms state-of-the-art baselines. Our code is available in the supplementary materials.

Abstract:
Breast cancer is one of the leading causes of death among women worldwide, and accurate segmentation of breast ultrasound images is of crucial importance. Compared with static images, videos have advantages in capturing lesions' dynamic and spatial information. However, research on breast tumor ultrasound video segmentation remains extremely limited due to the lack of high-quality publicly available datasets and the high cost of frame-by-frame annotation. In this situation, sparse-frame annotation of videos serves as a more affordable strategy, and diagnostic reports can be used to provide global semantic guidance for videos. However, diagnostic reports are unavailable in advance during clinical diagnosis, which limits the reasoning ability of current multi-modal models. To solve these issues, we construct the first large-scale multi-modal benchmark dataset, containing 1,085 videos with the corresponding clinical diagnostic text and tumor grading information. This dataset establishes a foundation for algorithm research and performance evaluation. Furthermore, we propose a multi-modal training framework with Text-Free inference for Breast tumor ultrasound Video Segmentation (TF-BVS). TF-BVS incorporates a Label-aware Visual-To-Text Optimal Transport module (LV2TOT) to generate pseudo-text token embeddings highly consistent with clinical text distributions, thus enabling inference under text-free conditions in line with clinical workflows. TF-BVS employs a dual-branch joint training strategy, achieving robust segmentation performance even with less annotated training data. Extensive experiments demonstrate that TF-BVS achieves state-of-the-art segmentation results on both the proposed multi-modal breast ultrasound dataset and the public VTUS thyroid ultrasound dataset.

Abstract:
To address the threat of fake news, the task of Detecting and Grounding Multi-Modal Media Manipulation (DGM^4) has gained increasing attention. However, existing methods predominantly treat these sub-tasks in parallel, resulting in logical inconsistencies and neglecting the causal dependencies between tasks. Moreover, current models lack explanations for the forgery process. In this paper, we propose SMART, a framework Synergizing Multi-Turn ChAin-of-Thought Reasoning and Reinforcement Fine-Tuning. Implemented upon a Multimodal Large Language Model, SMART introduces three key innovations. First, we reformulate the DGM^4 task into a four stage causal reasoning chain to ensure coarse to fine logical coherence. Second, we augment the DGM^4 dataset into a structured multi-turn Chain-of-Thought dataset comprising 230K samples to facilitate explicit interpretable reasoning. Finally, a rule based reinforcement learning strategy is employed to directly optimize forensic metrics such as IoU and token-F1. SMART eliminates logical paradoxes and improves interpretability in multimodal forgery detection. Experimental results demonstrate that SMART achieves state-of-the-art performance across all sub-tasks of the DGM^4 dataset. Our dataset and code will be made publicly available.

Abstract:
The escalating complexity of deep neural networks hinders the deployment of federated learning (FL) on resource-constrained devices, especially for multimedia applications such as on-device image classification and speech recognition. Parallel split learning (PSL) has emerged as a promising solution by offloading a substantial portion of the computational workload to the server through layer-wise model partitioning, making it particularly suitable for resource-limited visual and speech intelligence scenarios. However, the absence of client-side aggregation renders PSL susceptible to severe training divergence and potential convergence failure under heterogeneous data distributions. To address this challenge, we propose PSL-RGHA, a gradient-aligned \underlinePSL framework that employs \underlinereference \underlinegradient \underlinehalf-space \underlinealignment mechanism to mitigate training divergence among client devices. PSL-RGHA comprises two novel components: reference gradient progressive construction (RGPC) and gradient half-space alignment (GHA). Specifically, RGPC evaluates the directional consistency of client gradients and progressively synthesizes a robust reference gradient via Weiszfeld updates, effectively approximating the global descent direction. Then, GHA constrains client gradients within the reference half-space defined by RGPC to mitigate gradient conflicts and rectify local updates. Extensive experiments demonstrate that our PSL-RGHA framework outperforms existing baselines in various settings.

Abstract:
Granular-ball clustering (GBC) is an emerging clustering paradigm that provides a multi-granularity representation of samples via a scale-variable granular-ball (GB) topology. By capturing multi-scale features and enhancing robustness, GBC is particularly effective for processing high-dimensional data such as images. However, existing shallow GBC remain constrained by two primary bottlenecks. First, the assumption of isotropic hyperspheres impedes the adaptability of topology evolution to real-world data distributions. Second, these methods fail to explicitly characterize and model data uncertainty. To address them, the first Deep Adaptive Topology Evolution GBC with intuitionistic fuzzy representation (D\textscateGBC) is proposed for unsupervised image segmentation. D\textscateGBC innovatively synergizes deep architectures with GBC to establish a deepGBC paradigm. By leveraging the self-representation of deep models, it makes split-and-merge topology evolution adaptive to anisotropic data distributions, improving segmentation performance. Furthermore, D\textscateGBC introduces intuitionistic fuzzy granular balls (IFSGBs) to explicitly model data uncertainty by measuring the non-membership and hesitation of GBs, resulting in improved robustness. Extensive experiments on natural and remote-sensing image datasets demonstrate that D\textscateGBC outperforms other conventional clustering and deep clustering algorithms.

Abstract:
Underwater object detection faces unique challenges from haze, color distortion, and low contrast, which arise from light absorption and scattering characterized by the JaffeMcGlamery physical imaging model, significantly degrading image quality and detection performance. To address these degradations, we propose HE-MoESNN, a spiking neural network that integrates a Mixture of Heterogeneous Enhancement Experts (HE-MoE) with a lightweight Forward Spiking Neural Network (FSNN) backbone. Unlike conventional MoE frameworks, HE-MoE incorporates physics-guided modality decomposition, where degradation cues derived from underwater imaging principles are mapped to three modality-specific inputs: dehazing, color correction, and contrast enhancement. These inputs are processed by dedicated experts and fused through a shared routing mechanism. By embedding physically grounded cues into the MoE design, the model learns complementary, degradation-aware feature representations while explicitly enforcing expert diversity and reducing redundancy. Furthermore, to address the efficiency challenges of MoE, FSNN replaces conventional ANN activations and dense convolutions with signed spiking neurons and efficient ternary convolutions, enabling sparse computation while maintaining competitive detection accuracy. Extensive experiments on the RUOD and DUO benchmarks demonstrate that HE-MoESNN achieves state-of-the-art performance while consuming orders of magnitude less energy.

Abstract:
Recent advances in 3D Gaussian Splatting (3DGS) have enabled significant progress in dense dynamic Simultaneous Localization And Mapping (SLAM). Prevailing methods typically discard predefined dynamic objects, ignoring that transiently static objects offer valuable geometric constraints for pose estimation. A recent work attempts to leverage this potential by employing per-pixel uncertainty maps to quantify the magnitude of motion. While this approach enables transiently static objects to enhance pose estimation, it erroneously integrates these objects into the static map, resulting in persistent artifacts. Moreover, its reliance on purely geometric information leads to ambiguous object boundaries in the uncertainty maps. To overcome these limitations, we present DL-SLAM, a monocular Gaussian Splatting SLAM system built upon a novel dual-level probabilistic framework. Our method computes dynamic probability maps by combining semantic and geometric information. These pixel-level probabilities are lifted to 3D and aggregated to derive an object-level dynamic probability for each instance. Object-level probability enables the categorical pruning of dynamic Gaussians, resulting in an artifact-free static map. The static map, in turn, provides a geometrically consistent guidance to refine the pixel-wise probabilities, enhancing their reliability. Experimental results demonstrate that DL-SLAM outperforms existing approaches, improving tracking accuracy by up to 13% while generating high-fidelity semantic maps.

Abstract:
Multimodal Sentiment Analysis (MSA) integrates text, audio, and visual signals to understand human emotions. Most methods for handling missing modalities assume that all samples share the same missing pattern, but in practice each sample has its own missing combination due to sensor failures, noise, or transmission errors. Such sample-wise variation leads to unpredictable performance drops. We propose EPA-MoE, a pattern-aware mixture-of-experts model with entropy-adaptive gating for uncertain missing modalities. EPA-MoE clusters samples by their missing configurations to identify recurring degradation patterns, and incorporates the pattern information as a structural prior into the expert routing mechanism. The gating distribution adapts its entropy to each sample's missing severity sharpening when modalities are mostly available to focus on specialized experts, and flattening under severe degradation to spread predictions across multiple experts. The resulting unified architecture handles both common and rare missing patterns without separate processing branches. Experiments on CMU-MOSI and CMU-MOSEI show that EPA-MoE outperforms existing methods under uncertain missing conditions, and analyses of routing patterns confirm that the gating mechanism adapts its behavior to different missing scenarios.

Abstract:
With the rapid evolution of generative artificial intelligence, deepfake detection faces severe challenges in cross-domain generalization. Existing methods essentially follow a passive observation paradigm, heavily relying on mining forgery artifacts from a single static image. However, such static features are often strongly coupled with specific generators, leading to overfitting and significant performance degradation when encountering unseen manipulations. To break this fundamental limitation, we rethink the paradigm of deepfake detection and advocate a shift from passive observation to active creation. Our key finding is that real and fake images exhibit asymmetric reconstruction responses under the same generative transformation: real images originating from the physical world experience "manifold collapse" when passing through the generative bottleneck, resulting in severe representational drift; whereas fake images residing within the generative manifold exhibit strong "manifold inertia" with relatively minor feature perturbations. Based on this observation, we propose a novel framework termed Probing Reconstruction Discrepancy (PRD). Rather than relying solely on static representations, PRD actively probes the input by modeling the representational variations induced during the reconstruction process, thereby capturing more stable and transferable forgery cues. Extensive experiments demonstrate that PRD achieves state-of-the-art cross-dataset frame-level performance, reaching an average AUC of 90.48%, while requiring only 3.7M trainable parameters.

Abstract:
Federated learning (FL) enables collaborative training without sharing raw data, but optimization becomes unstable under data heterogeneity and sparse participation. Federated primal--dual (FedPD) methods mitigate client drift by introducing dual variables to enforce consensus; however, under partial participation, inactive clients accumulate stale dual states, and their re-entry can induce severe primal--dual misalignment, degrading both convergence and generalization. To address this issue, we propose FedFGS, a FedPD framework that improves robustness under sparse participation through two complementary mechanisms. First, Friend-Guided Synchronization continuously evolves the reference dual states of inactive clients using cues from geometrically similar active clients, enabling a stable hot start upon re-entry. Second, Geometric Stabilization (GS) improves the reliability of friend estimation by stabilizing update geometry on the server side, preventing noisy similarity signals from corrupting synchronization. We further provide theoretical results on bounded virtual-dual deviation, GS perturbation stability, nonconvex stationarity, and generalization, where the effect of sparse-participation synchronization is explicitly characterized through a friend-mismatch term. Extensive experiments on standard federated benchmarks show that FedFGS consistently improves training stability, accelerates convergence, and enhances generalization, with especially pronounced gains under highly sparse participation.

Abstract:
Multimodal 3D object detection is a cornerstone of autonomous driving perception, yet its adversarial robustness remains poorly understood. Due to the lack of a unified cross-modal representation, existing adversarial attacks manipulate image and point cloud inputs separately, leading to limited physical realizability and suboptimal attack effectiveness. To address this problem, we propose MiuGS, a modality-integrated unified Gaussian representation framework for generating 3D adversarial objects with strong attack effectiveness and physical-world potential. For physical realizability, we design a multimodal 3D Gaussian coordination strategy to address the imprecise cross-modal fusion issue. Specifically, MiuGS models the adversarial object as a shared 3D Gaussian representation and instantiates it into consistent image and point cloud observations, enabling the two sensing streams to be coupled at the representation level rather than optimized independently. To further strengthen attack effectiveness, we introduce a cross-scenario guidance rebalancing strategy to mitigate cross-modal guidance drift across scenarios and the digital-to-physical domain gap. Extensive experiments show that MiuGS achieves strong attack effectiveness, reaching an average ASR of 78.8% on nuScenes and 44.3% on CARLA, supporting the effectiveness and physical-world potential of the proposed unified representation.

Abstract:
Modern Extended Reality (XR) systems are becoming increasingly heterogeneous, as latency-critical perception and rendering pipelines are now coupled with compute-intensive and bursty foundation-model-based agents. This trend makes edge-cloud offloading essential for sustaining interactive performance under dynamic workloads. However, existing offloading architectures are limited by a fundamental bottleneck: task migration in XR is often dominated by the transfer and restoration of heavyweight execution context, such as scene assets and model parameters, which incurs severe cold-start delay and undermines real-time responsiveness. We present \textttDEXA, a task offloading framework that enables low-overhead migration through static-dynamic context decoupling. \textttDEXA separates each task into a heavyweight static context and a lightweight dynamic context, pre-stages static assets on edge and cloud nodes, and reduces runtime migration to dynamic-state synchronization. On top of this design, \textttDEXA employs a penalty-based online scheduler to jointly optimize service latency and migration cost. We implement \textttDEXA on a prototype that integrates high-fidelity rendering and LLM-based agent serving. Results show that \textttDEXA effectively mitigates cold-start overhead during migration, maintains rendering SLA compliance, and improves system throughput and responsiveness under dynamic load.

Abstract:
Vision-language process reward models (VL-PRMs) score intermediate reasoning steps and rerank candidates under test-time scaling, yet their step scores are often hard to interpret: a low reward may indicate a genuine reasoning mistake, or simply unreliable visual grounding by the policy or the verifier. This entanglement yields systematic false positives (rewarding hallucinated visual premises) and false negatives (penalizing correct grounded statements), degrading both reranking and error localization. We propose Explicit Visual Premise Verification (EVPV), a lightweight, test-time framework that decouples visual premise reliability from step correctness. EVPV prompts the policy to emit a step-wise visual checklist and independently extracts structured visual constraints from the image. By matching checklist claims against constraints, EVPV computes a scalar visual reliability signal and uses it to gate rewards for visually dependent steps, avoiding per-step tool calls. To fully realize this framework, we train EVPV-PRM, a Qwen2.5-VL-Instruct-7B-based step verifier that provides probabilistic base rewards and can be calibrated by EVPV at inference time. Across VisualProcessBench and six downstream benchmarks, EVPV improves step-level verification and yields overall Best-of-N reranking gains, while our EVPV-PRM achieves strong performance as a deployable reranker. Under controlled constraint corruption, performance degrades monotonically, providing interventional evidence that the gains are driven by constraint fidelity.

Abstract:
Whole slide images (WSIs) offer rich multi-scale morphological information critical for survival prediction, yet the sparsity and spatial dispersion of prognostic evidence make it difficult for existing methods to identify critical regions and capture complex dependencies across regions and scales. To address these issues, a novel hierarchical semantics-guided hypergraph network for WSIs-based survival analysis is proposed, named SemHyperSurv. A large language model first generates pathology-aware hierarchical prompts to derive a slide-specific hierarchical semantic prior. Guided by this prior, SemHyperSurv adopts a coarse-to-fine cross-scale focusing strategy to localize prognostically relevant regions. It then constructs a structure-semantic hypergraph over the retained regions, explicitly injecting the semantic prior into hyperedge construction to encode cross-scale containment, intra-scale contextual dependencies, and concept-conditioned higher-order semantic relations. Finally, it performs a semantics-guided prognostic readout via concept queries and gated fusion for robust and interpretable risk prediction. SemHyperSurv is evaluated on six public cancer datasets and demonstrates superior performance over existing methods.

Abstract:
Multimodal assistants are increasingly deployed on resource-constrained devices, yet the core trade-off remains unresolved: Large Language Models (LLMs) are too heavy for on-device execution, while full cloud offloading incurs high latency, bandwidth costs, and privacy risks. Existing edge-cloud collaborative systems rely on \emphself-evaluative routing (SER), where the edge model decides whether to respond locally or offload to the cloud. On one hand, we find this strategy unreliable due to the inherent capability limitations of small models on edge devices. On the other hand, real-world model usage exhibits clustering, temporal stability, and partially shared topic characteristics across users. Based on these observations, we argue that the edge should act not as a general assistant, but as a cache of frequently used capabilities. In this paper, we propose SkillCache, which caches model capabilities via \emphmulti-view coupled skill operators for low-latency multimodal edge intelligence. In SkillCache, capabilities are represented in a unified space and can be combined through sparse composition over shared multimodal bases. As a result, SkillCache enables the discovery, routing, maintenance, and personalization of capabilities through simple coefficient operations. We implement a prototype of SkillCache using typical edge and cloud models, and extensive experimental results show that SkillCache closes 62.0% of the edge--cloud quality gap (Core GC) on MMBench, OCRBench~v2, and CharXiv at the matched 34% cloud budget, with smaller gains on the harder BLINK and MMMU stress tests. It reduces uplink traffic by 88% (85.34% net bidirectional reduction after amortizing skill-pack delivery) with only 30\,ms routing overhead.

Abstract:
Sparse trajectory recovery has achieved substantial progress in single-city settings, yet zero-shot generalization to unseen cities remains limited. A key challenge is that existing methods often entangle trajectory modeling with city-specific spatial layouts, road-network topology, and contextual distributions, causing them to rely on source-city priors rather than learn transferable motion patterns. To address this issue, we propose a motion-centric multimodal model for cross-city zero-shot trajectory recovery. The proposed model constructs a motion-aligned local frame from observed trajectories and local road geometry, models trajectory evolution as relative motion relations in this local coordinate system, and incorporates multimodal urban context through a controlled late-fusion design. In this way, the model preserves transferable motion dynamics while introducing structural and semantic environmental cues for recovery calibration. Extensive experiments on cross-city transfer tasks over three real-world trajectory datasets show that our method consistently outperforms existing approaches under strict zero-shot settings across different city pairs and sampling sparsities, while also maintaining strong in-domain recovery performance.

Abstract:
The increasing scale of spiking neural networks (SNNs) poses significant challenges for deployment on resource-constrained neuromorphic hardware, necessitating lightweight and learnable structural solutions. Interestingly, biological neural systems employ an efficient organizational strategyhierarchical structural reorganization around functional clusters, where new connections grow orthogonally to existing ones to expand representational capacity. Inspired by this mechanism, we propose a dynamic pruning and regrowth framework with channel-level orthogonality for SNNs (DPRC-SNNs) to enable scalable and efficient structural learning for SNNs. DPRC-SNNs introduce the spiking column subset selection mechanism for SNNs, which integrates channel-level pruning with orthogonality-driven regrowth, selectively restoring diverse and complementary channels to minimize information loss from aggressive pruning. Through iteratively pruning redundant channels and regrowing orthogonal ones, DPRC-SNNs preserve functional diversity while enhancing sparsity at the channel level. Extensive evaluations on CIFAR10, CIFAR100, ImageNet,DVS-Gesture, and DVS-CIFAR10 demonstrate that DPRC-SNNs achieve high compression rates and computational efficiency without compromising accuracy, showing strong potential for neuromorphic deployment.

Abstract:
Camera trajectory design plays a crucial role in video production, as it determines how a target object is presented throughout a shot. In object-aware camera trajectory generation, the key challenge is not only to produce plausible camera motion, but also to maintain stable cameraobject interaction as the target moves. Specifically, the trajectory should keep the object within a proper image region, preserve a stable and reasonable scale, and ensure the camera consistently faces the object. Existing methods improve controllability from text or point-based object cues, but they often fail to preserve this interaction over time. Language provides high-level camera-motion intent, while a time-varying 3D bounding box supplies frame-wise target geometry for preserving framing, visibility, and viewing direction. However, object location alone does not capture the targets spatial extent and orientation, and existing datasets lack aligned supervision for generating camera trajectories from both text and structured 3D target state. To address this, we establish target-object camera trajectory generation from text and 3D target state as a new benchmark task, and introduce BlockCam, a 41K-sequence benchmark with aligned textual motion descriptions, target-object 3D bounding box sequences, and camera trajectories collected from real and synthetic videos under continuous target visibility. We further propose Camera Operator, a geometry-conditioned flow-matching model whose intermediate clean trajectory estimate enables projection-space supervision of cameraobject relations. We evaluate this benchmark using trajectory-alignment and distributional metrics as well as two geometric measures, In-Frame Ratio and Look-at Error, to assess target visibility and cameratarget alignment. Experiments show that the full configuration strengthens relation preservation while maintaining strong texttrajectory alignment. Across progressively richer geometry-conditioned configurations, the full-box model outperforms its location-only counterpart, and a user study favors our results in terms of alignment and framing.

Abstract:
The exceptional generation capabilities of text-to-image diffusion models have raised copyright concerns, particularly the unauthorized reproduction of animation characters. Existing concept erasure methods fall short for animation character erasure: model modification methods struggle to identify suitable anchors for diverse, highly distinctive characters; prompt-based steering methods lack fine-grained control for precise intervention. These approaches often yield incomplete erasure and degraded image fidelity, hindering real-world deployment. In this paper, we propose a controllable method operating on the models continuous textual representation to erase target characters during generation. We optimizes an anchor embedding via structural and detailed constraints to serve as a character surrogate, then replaces target-related embeddings with the anchor via a structure-aware adaptive strategy. Experiments show that our method achieves state-of-the-art erasure effectiveness and image fidelity preservation, while supporting controllable erasure degree, multi-target removal, and model transferability. Moreover, our optimized anchors are plug-and-play with current model modification baselines to improve their erasure performance.

Abstract:
Diffusion Multimodal Large Language Models (DMLLMs) are highly effective for multimodal reasoning, yet their inference efficiency is significantly hindered by fixed-length generation constraints. Since the actual output length is unknown, output sequences are padded to a predefined maximum length, resulting in substantial redundant computation over unnecessary [EOS] tokens. In this work, we discover that DMLLMs implicitly reveal their valid semantic boundary at the very first denoising step through a distinct shift in MLP activation sparsity. Leveraging this observation, we propose Seer, a training-free framework that detects this boundary using a Signal-to-Noise Ratio (SNR)-based criterion and performs one-shot truncation of the redundant suffix for all subsequent computations. To preserve these theoretical gains during batched serving, Seer incorporates a hybrid execution strategy that maximizes throughput while seamlessly accommodating dynamic sequence lengths. Experimental results demonstrate that Seer effectively eliminates padding waste, accelerating throughput by up to approximately 31. Across nine benchmarks, Seer robustly maintains overall performance and even improves accuracy on complex visual tasks by mitigating noise leakage (e.g., the DocVQA score increases from 63.52 to 63.66), offering a highly efficient, plug-and-play solution for DMLLM acceleration.

Abstract:
The rapid evolution of generative models has precipitated a proliferation of fabricated content, posing significant challenges to existing Synthetic Image Detection (SID) methods. Capitalizing on advancements in vision-language models (\eg, CLIP), recent attempts have leveraged learnable textual prompts to identify synthetic images. However, they still leverage static prompt as a fixed boundary for real and fake images, failing to adapt to the varying types of forgery that emerge during inference. To overcome this issue, we propose HydraPrompt, an asymmetric prompting framework that dynamically adjusts the category centers by aligning with fine-grained visual cues. Specifically, we propose an Asymmetric Prompt Adapter (APA): (1) for authentic category, we introduce a single set of prompts to capture the consistent representative patterns, which serves as a unified anchor for real content. While (2) for fake category, we construct sample-adaptive prompts that specialize in capturing diverse cues from different samples, enabling adaptive modeling of forgery image variations. To increase pronounced discriminability within different synthetic images, we further introduce a Conditional Supervised Contrastive (CSC) objective. Extensive experiments on popular SID benchmarks demonstrate the state-of-the-art performance of our framework.

Abstract:
Large Audio-Language Models (LALMs) excel at general speech understanding. However, adapting them to fine-grained tasks like Speech Emotion Recognition (SER) remains a significant bottleneck. Current Parameter-Efficient Fine-Tuning (PEFT) methods typically operate in flat Euclidean space. This geometry fails to capture the multi-granularity nature of emotion cues, which range from low-level prosody to high-level semantics. To address this, we propose \method, a hyperbolic PEFT framework for LALM-based SER. \method leverages the Poincar\'e ball model, using the hyperbolic radius as an explicit proxy for representational granularity. The framework consists of two core components: a Hyperbolic Geometric Adapter (HGA) for layer-adaptive weight modulation, and an Emotion-aware Multi-capacity Cross-modal Aggregator (EMCA) that compresses multi-scale features into compact audio prefixes. Empirical results on standard benchmarks show that \method outperforms Euclidean PEFT baselines across all metrics on MELD and achieves a notable Unweighted Accuracy gain on IEMOCAP, particularly in class-imbalanced emotion recognition, with the accompanying slight Weighted Accuracy trade-off reflecting hyperbolic space's geometric prioritization of minority-class representations. Furthermore, \method achieves robust zero-shot cross-dataset generalization within a constrained parameter budget. By grounding the adaptation process in hyperbolic geometry, \method offers a highly efficient path for LALM fine-tuning.

Abstract:
Robotic-View Human Action Recognition (Robot-HAR) aims to recognize human actions from sequences captured by a moving robotic platform, where observer motion introduces substantial challenges beyond conventional fixed-camera settings. In addition to intrinsic human-motion variability, observer motion induces broadly shared spatiotemporal variations, while relative robot--human motion further amplifies the diversity of effective temporal and spatial scales. To address these issues, we propose RoboDAS, an efficient scan-based framework tailored to Robot-HAR. Its core DAS Block combines two complementary components: a Global Differential Scan (GDS), which performs differential global propagation along structured scan orders to attenuate observer-induced shared variations, and a Local Adaptive Scan (LAS), which applies scale-specialized local experts with lightweight top-k routing to capture heterogeneous action cues at appropriate receptive fields. This design yields a unified spatiotemporal representation that is robust to moving-observer disturbances and adaptive to diverse action dynamics across modalities. Extensive experiments on robotic-view point-cloud and UAV-view RGB benchmarks demonstrate that RoboDAS achieves state-of-the-art recognition accuracy with favorable computational efficiency.

Abstract:
Recognizing fine-grained bodily behaviors, such as self-grooming, postural shifts, and backchanneling, is a useful intermediate step for computational social signal analysis. But the behaviors that matter vary across social contexts, and retraining a model for each new taxonomy is expensive. Vision-language models (VLMs) can adapt to new categories through text prompts, yet they lack sensitivity to body structure and often confuse subtle postural distinctions. We propose PRIMA (Pose Refined Inductive Motion Adapter), a lightweight module that injects a body-structure prior into frozen VLMs using human pose as a training-time supervisory signal, requiring no pose input at inference. PRIMA separates a model-specific projection head from a shared core that encodes transferable body-configuration knowledge; adapting to a new VLM requires retraining only the head. We train PRIMA on a micro-action dataset and evaluate entirely through zero-shot prompting. The adapter improves recognition on its training domain, transfers to additional VLM backbones at different scales with the core frozen, and generalizes to two unseen target datasets from social psychology and daily activity recognition. In a supporting downstream study, the improved behavior predictions also yield higher correlations with human ratings of rapport and dominance in a collaboration context. Within the scope of the tested models and indoor datasets, the results indicate that pose-guided training can produce body-structure priors that generalize across unseen behavior vocabularies and can be partially reused across VLM backbones while preserving RGB-only inference.

Abstract:
In the rapidly developing zero-shot skeleton-based action recognition (ZSAR), the mainstream approaches typically construct a shared semantic space to align global skeleton representations with label semantics, using the Transformer framework. However, existing methods still suffer from the following issues: 1) Global skeleton features are of limited use for modeling fine-grained motions. It is hard to model complex interactions between body parts; crucial limb movement information is also partially lost. 2) Local features between text and skeleton are difficult to locate. The locating of local features relies solely on global information in cross-modal alignment, which may easily lead to difficulties in fine-grained semantic alignment. To tackle the above challenges, we propose BiPartFormer, a novel transformer-empowered ZSAR framework based on bipartite graph interaction and dynamic part-alignment. First, a part-aware bipartite graph transformer is designed for fine-grained motion and interaction modeling, with L2S Calibration using global cues to refine local representations. Second, to accurately locate local features for fine-grained semantic alignment, a dynamic part-alignment module is proposed for BiPartFormer. Finally, extensive experiments demonstrate that the proposed BiPartFormer achieves superior performance compared with recent state-of-the-art methods on NTU RGB+D, NTU RGB+D 120, and PKU-MMD datasets for ZSAR.

Abstract:
High-fidelity 3D medical visualization supports applications such as clinical assessment and surgical planning. Sparse-view computed tomography (CT) can reduce projection requirements and associated radiation exposure, but limited observations may introduce structural artifacts and reconstruction uncertainty. Although 3D Gaussian Splatting (3DGS) provides an efficient explicit representation for volumetric rendering, existing CT methods based on standard Gaussian primitives may be sensitive to unreliable observations under sparse-view acquisition. We present TR-GS, a Gaussian-splatting framework for sparse-view CT volumetric rendering. TR-GS replaces standard Gaussian primitives with projectable Student's \(t\)-distribution primitives and introduces a ray-confidence model that regulates their degrees of freedom according to local ray observability. Confidence-guided 3D wavelet regularization is further used to balance high-frequency detail preservation and noise suppression. Experiments on synthetic and real-world datasets show that TR-GS improves over representative baselines in most evaluated settings and remains competitive in the remaining cases. The resulting volumetric representations may support downstream medical multimedia applications, including XR-based visualization and interactive clinical rendering.

Authors: Mengjie Zhang (University of Science and Technology of China), Qihui Zhu (University of Science and Technology of China), Tao Zhang (University of Science and Technology of China), Shuangwu Chen (University of Science and Technology of China), Huihuang Qin (University of Science and Technology of China), Yu Guo (University of Science and Technology of China), Shenghao Ye (University of Science and Technology of China), ZijianWen (University of Science and Technology of China), Yunpeng Hou (Institute of Artificial Intelligence, Hefei Comprehensive National science Center), Dong Jin (Institute of Artificial Intelligence, Hefei Comprehensive National science Center), Xiaobin Tan (University of Science and Technology of China), Huasen He (University of Science and Technology of China), Jian Yang (University of Science and Technology of China)

Abstract:
Video large language models (VideoLLMs) exhibit exceptional video understanding performance, yet their inference costs remain prohibitively high due to the massive volume of visual tokens required, especially for long video understanding. Visual token pruning is a promising paradigm for mitigating such inference costs by eliminating redundant tokens across video frames. However, most existing token pruning methods adopt a segment-level pruning strategy, where videos are partitioned into isolated segments and tokens are selected independently within each segment. Such designs may cause information-dense segments to discard tokens that are seemingly non-salient from a local perspective but remain critical from a global one. To address this issue, we propose GSTEP (Global Spatio-Temporal Density Pruning), a plug-and-play pruning framework that models video as a continuous spatio-temporal information flow. GSTEP constructs a token-level spatio-temporal density by combining a continuous temporal density, obtained from a smoothed centered frame-level change signal, with intra-frame spatial density, and then performs global token sampling by jointly balancing information density and coverage. Extensive experiments on multiple VideoLLMs and public benchmarks demonstrate that GSTEP consistently achieves strong accuracy-efficiency trade-offs and generalizes well across model architectures and evaluation settings. On LLaVA-OneVision-7B, GSTEP prunes 75% of visual tokens, preserves up to 100.2% of the original average performance across benchmarks, and achieves a 1.17× end-to-end speedup.

Abstract:
LiDAR is widely used in autonomous driving perception for its accurate 3D geometry, but its inability to directly measure point-wise velocity limits dynamic object perception. In contrast, 4D radar provides informative Doppler velocity cues for moving object detection, yet its point clouds are extremely sparse and lack detailed spatial structure. Effectively combining the complementary strengths of LiDAR and 4D radar remains a key challenge for robust 3D object detection. Existing fusion methods typically initiate intermodal interaction only after the encoder stage, which limits the joint exploitation of LiDAR geometry and radar velocity. Meanwhile, they still struggle to effectively address intermodal misalignment. To overcome these limitations, we propose VLRFusion, a velocity-aware LiDAR-4D radar fusion framework that operates at the data, encoder, and feature levels. Specifically, VLRFusion introduces a 4D LiDAR generator that scatters radar velocity onto LiDARs precise geometry to construct a new 4D LiDAR modality. Based on this modality, we further design a velocity-decoupled backbone and a residual weighting mechanism to enhance motion object perception. In addition, a radar-guided alignment encoder and an attention-based fusion module are developed to mitigate intermodal misalignment and improve feature interaction. Experiments on the View of Delft and Dual-Radar datasets show that VLRFusion achieves state-of-the-art performance in LiDAR-4D radar fusion for 3D object detection, outperforming both single-modality and camera-4D radar fusion methods.

Abstract:
Active non-line-of-sight (NLOS) imaging from a single relay-wall observation is severely information-poor. In this work, we study the problem of segmenting a hidden person's 2D silhouette from one wall reflection captured by a conventional camera under flashlight illumination, where diffuse low-pass transport suppresses morphology-relevant structure and renders the inverse mapping highly ill-posed. We propose RecDiffusion, a two-stage framework for active NLOS human silhouette segmentation. Its principal component is a decoupled Spatial--Spectral Recalibration (dSSR) module that rebalances dominant wall content, enhances recoverable frequency- and phase-related cues, and aligns them with spatial energy anomalies to produce a more stable representation for mask prediction. On top of this stabilized representation, Annealed Radial Diffusion (ARD) serves as a weak-physics-guided refinement stage that improves mask coherence, boundary continuity, and geometric consistency, rather than generating the segmentation from scratch. To benchmark this setting, we construct ReflectHuman, a dataset of real and synthetic relay-wall reflection scenes for hidden human silhouette segmentation. Across four ReflectHuman subsets, RecDiffusion outperforms the evaluated NLOS, segmentation, GAN, and diffusion baselines on region-based and boundary-sensitive metrics; on the two real subsets, it achieves Dice scores of 0.711/0.728 and Boundary-F1 scores of 0.147/0.153 while reducing 95HD. Ablation studies further indicate that dSSR is the dominant contributor to stable silhouette recovery, whereas ARD provides smaller but consistent gains in global mask coherence and extreme-error suppression. These results demonstrate that coherent hidden-human silhouettes can be recovered from highly degraded active NLOS wall reflections using low-cost, off-the-shelf hardware.

Abstract:
Light fields exhibit entangled spatial-angular frequency responses that introduce substantial redundancy across views and frequency bands. Such redundancy obscures structural details and makes reliable saliency discrimination particularly challenging in unsupervised settings. To address this issue, we propose SF, a structured-sparsity driven frequency reasoning framework for Unsupervised Light Field Salient Object Detection. The framework consists of two key components: Light Field Sparse Feature Generation (SFG) and Multi-domain Light Field Frequency Fusion (MLF). Specifically, SFG learns compact spatial-angular representations through collaborative sparsity and adaptive local reweighting, thereby suppressing redundant responses while preserving geometric consistency across views. Built upon these sparse representations, MLF jointly exploits spatial- and angular-domain frequency cues to enhance saliency contrast, recover object structures, and refine boundary details. By coupling sparse structural representation with cross-domain frequency modeling, SF enables intrinsic saliency inference directly from the structural and frequency regularities of light field data. Extensive experiments on three public benchmarks demonstrate that SF consistently outperforms all state-of-the-art unsupervised methods and most weakly supervised counterparts.

Abstract:
We present RelayMask-Bench, a physically captured relay-wall hidden target mask benchmark for recovering hidden target masks from a single relay-wall RGB observation in a corner-occluded steady-state setting. Under the RelayMask protocol, RGB images from public segmentation datasets are displayed on a hidden emissive plane, while their paired binary masks are reused as supervision; the model input always remains a real relay-wall photograph formed by actual multi-bounce indirect transport. This design enables scalable and reproducible evaluation while preserving the real degradations of indirect observations, including path mixing, boundary attenuation, and sensor noise.To address this problem, we propose MOSTE-Net, a transport-aware encoder-decoder with three coordinated components: a Neumann-style Multi-order Transport Prior (NMTP) to separate and reweight effective transport responses, a Spectral-Discontinuity Gradient-Coherent Edge Attention (SD-GCEA) module to enhance weak boundary cues, and a forward structural consistency constraint to suppress transport-inconsistent predictions. Experiments on physically captured relay-wall measurements show that MOSTE-Net consistently outperforms representative visible-domain, physics-inspired, and NLOS-related baselines under the unified RelayMask protocol. For example, on RelayMask-ECSSD it improves Max-F from 0.625 to 0.716 and reduces MAE from 0.175 to 0.123, while also showing stronger robustness under realistic observation perturbations. Overall, our work establishes RelayMask-Bench as a reproducible benchmark and provides a strong transport-aware baseline for hidden target inference from real indirect visual measurements.

Abstract:
In semi-supervised video segmentation, recent methods built on SAM 2 have achieved strong performance. However, SAM 2 lacks explicit motion modeling during mask selection and relies on a short-term memory mechanism that provides limited context for re-identifying targets after prolonged occlusions. To address these limitations, this paper proposes ProtoSAM2, a training-free framework that enhances SAM 2 from both motion modeling and memory storage. Firstly, we combine a lightweight optical flow estimation method with Kalman filtering in an adaptive manner, enabling the model to handle complex non-linear motion while maintaining computational efficiency. Secondly, we introduce a persistent semantic anchor into SAM 2's memory bank through multi-scale object prototypes, which are continuously updated during tracking to capture the evolving semantic features of targets and facilitate reliable re-identification after long occlusions. Experiments on challenging long-term benchmarks demonstrate that ProtoSAM2 consistently outperforms the SAM 2 baseline and recent extensions across different model scales. Notably, ProtoSAM2-L achieves 81.5% \mathcalJ\&\mathcalF on the SA-V test set and improves over SAM 2 by 2.9% on LVOS-v1.

Abstract:
Multimodal Named Entity Recognition (MNER) aims to identify entities and their types from textimage pairs, using images to resolve ambiguity in short social media posts. Existing approaches rely heavily on large-scale annotated datasets and often struggle to generalize to unseen entities. To address these limitations, we formalize the zero-shot MNER (ZeroMNER) task, where models must perform entity recognition without any labeled data. We adapt large language models (LLMs) to this setting by leveraging their strong generalization capabilities. However, our experiments show that LLMs remain inadequate for ZeroMNER, despite performing substantially better in few-shot scenarios. Motivated by this gap, we propose GTA-MA, a Generative Test-time Augmentation Multi-Agent framework that reformulates ZeroMNER as a pseudo in-context learning (ICL) problem. GTA-MA synthesizes test-aligned multimodal demonstrations at inference time, conditioned on each input instance, ensuring strong semantic relevance to the target sample and improving ICL effectiveness. GTA-MA further employs coordinated agents to generate, annotate, select, and exploit these demonstrations, forming reliable pseudo in-context supervision that enables LLMs to self-hint without requiring labeled data. Experiments on Twitter-2015 and Twitter-2017 demonstrate substantial improvements, with F1 gains of 15.13 and 14.43 points over vanilla zero-shot baselines, respectively. Moreover, GTA-MA achieves performance comparable to few-shot approaches while consistently outperforming existing zero-shot methods.

Abstract:
Video reconstruction from brain signals is a complex interdisciplinary task involving text, images, and the brain. Existing reconstruction frameworks are based on individual training paradigm, requiring separate model training for each subject or large-scale fine-tuning for new subjects, which severely limits the development of brain-computer interface systems. Although some cross-subject approaches have been proposed, they mostly focus on invariant features of subjects while ignoring subject-specific information and temporal dynamic, resulting in limited reconstruction quality. Therefore, we propose NeuroCross, a novel cross-subject visual decoding framework. First, it decomposes fMRI signals into semantic streams, temporal streams, and subject-specific residuals through Semantic Temporal Routing (STR). Then, it separates invariant features and specific features by implementing adversarial constraints in the Shared Residuals Decoupling (SRD), and finally completes personalized feature fusion through Personalized Dynamic Recombination (PDR) under motion guidance. Extensive experiments have shown that NeuroCross has achieved excellent performance in cross-subject video decoding tasks, improving CLIP-PCC and SSIM by 23% and 114%, respectively, over the previous state of the art.

Abstract:
\beginabstract All-day UAV object tracking in day-and-night scenarios confronts multiple challenges, including drastic illumination variations, low-light noise, occlusion, and fast motion. Consequently, these factors induce representation degradation and significant appearance variation, rendering traditional template matching ineffective. Existing methods, which primarily rely on rule-based template updates or short-term feature adaptation strategies, struggle to maintain robust object representations under extreme lighting conditions and long-term appearance changes. To address these issues, we propose a diffusion-enhanced Mamba tracking framework with temporal memory for day-and-night UAV tracking. First, we design a conditional diffusion-guided enhanced representation learning mechanism (DERL). By introducing learnable noise priors and a single-step lightweight diffusion strategy within the feature space, we guide the reconstruction process conditioned on the original template. This mechanism adaptively enhances template representation while preserving semantic consistency, effectively mitigating scene noise such as low illumination and texture degradation. Second, we construct an object modeling module by designing a Mamba-fusion-based temporal memory (MFTM-OM). Utilizing a quality-aware memory storage mechanism and a State Space Model (SSM), we fuse forward, backward, and inter-layer token shift information to model the target's appearance evolution patterns over long temporal spans, thereby bolstering robustness against illumination and pose variations. Finally, we propose a Diffusion-anchored Memory-Template Adapter (DiMT-Adapter) to bridge the diffusion-enhanced template and the memory template. Anchored by the diffusion prior, it employs dynamic separable convolution kernels and gated residual connections to efficiently fuse the diffusion-enhanced representations with temporal memory representations. Extensive experiments on multiple day-and-night UAV tracking benchmarks demonstrate that our method achieves significant performance improvements across both day and night scenarios, providing an effective solution for all-day UAV object tracking. \endabstract

Abstract:
3D human mesh recovery from monocular sketches remains challenging due to inherent ambiguity and information sparsity in sketch inputs. While deep learning has enabled remarkable progress in 3D reconstruction from natural images, sketches pose unique challenges due to their lack of texture, depth cues, and precise geometric information. To address these challenges, we present SMPLiketch, a novel multi-scale framework that bridges this representational gap through three novel components: 1) a dual-branch encoder leveraging both local sketch features and semantic part segmentation priors, 2) a Multi-scale Context Fusion Module preserving and integrating critical information across different scales through adaptive feature enhancement, and 3) a multi-level decoder progressively recovering detailed 3D human mesh parameters. Our framework employs a two-stage training strategy, first learning from synthetic sketches before adapting to real hand-drawn inputs. Extensive experiments demonstrate that SMPLiketch achieves state-of-the-art performance, reducing Mean Per Joint Position Error by 11.14 mm (9.0%) on synthetic sketches and 14.53 mm (9.4%) on freehand sketches compared to existing methods. Qualitative results showcase our method's robustness across varying sketch styles while maintaining reconstruction fidelity.

Abstract:
We propose ECHO for dyadic 3D facial motion generation under a strict dual-stream audio-only setting, formulating the problem as an asymmetric task involving speech-constrained articulation and one-to-many listener reactions. To address this asymmetry, ECHO decomposes motion into a deterministic anchor that captures stable speech-correlated structure and a stochastic residual that models the remaining one-to-many interaction dynamics. On top of this backbone, Motion Memory acts as a training-only regularizer during brief late-stage fine-tuning to provide local priors for weakly conditioned listening windows, while semantic-group scaling controls residual injection across expression, jaw, and neck. This design balances speaking-side articulatory fidelity with listening-side realism and diversity in a single generation process. Results from unified, state-wise, and ablation evaluations show that conversational 3D motion benefits from decomposing stable and uncertain components rather than applying stochasticity uniformly. ECHO provides a practical formulation and technical basis for deployable conversational digital humans under strict audio-only conditions.

Abstract:
The rapid proliferation of social media has accelerated multimodal misinformation, where false claims paired with manipulated images make detection substantially more challenging. Existing multimodal fake news detection methods typically predict each sample independently, ignoring inter-sample relational evidence. While label propagation (LP) over inter-sample graphs captures propagation patterns, label scarcity limits its effectiveness. Recent methods inject LLM-generated pseudo-labels into LP to compensate, yet this naive integration introduces two issues: connecting all samples without selective distinction causes erroneous cross-class label propagation, and noisy pseudo-labels corrupt model parameters through a noise-amplification effect. To address these issues, we propose GCLP, a gradient-consistency-driven LP framework for multimodal fake news detection. We design Anchor-Driven Gradient Walk Graph Construction (ADGW), which leverages gradient consistency between high-confidence anchors and training nodes to build a class-aware propagation subgraph, eliminating erroneous label mismatch propagation. We further introduce a Gradient-Guided Denoising Masking mechanism (GGDM), which fuses per-parameter gradient signal-to-noise ratio with anchor-based consistency to selectively mask noisy updates, thereby protecting model parameters from label noise and mitigating the noise-amplification effect. Finally, Gradient Graph-based Label Propagation (GGLP) performs attention-weighted propagation over the augmented graph for the final prediction. Experiments on three benchmarks spanning two languages show that GCLP achieves state-of-the-art performance.

Abstract:
Text-guided asset placement in 3D Gaussian Splatting (3DGS) scenes presents a compelling frontier for immersive content creation. However, current progress is hindered by two primary limitations: (i) the lack of frameworks for purely text-guided layout within entangled representations, and (ii) the reliance of existing methods on cumbersome manual supervision or explicit geometric priors. In this paper, we introduce \scshape Loci-GS, a novel framework designed to automatically place assets into 3DGS scenes guided solely by textual instructions. Specifically, we employ a hierarchical Multimodal Large Language Model (MLLM) strategy for precise semantic grounding, coupled with a multi-view consensus mechanism to accurately localize interaction anchors. Subsequently, we derive optimal placement via geometry-aware alignment, utilizing physical regularization and appearance harmonization to ensure realistic integration. Extensive experiments demonstrate that our method generates physically plausible and visually coherent inserts, achieving state-of-the-art performance compared to editing baselines on a single RTX 3090 GPU.

Abstract:
In-context learning (ICL) is widely used in multimodal large language models (MLLMs) and achieves strong performance across a wide range of multimodal tasks. However, existing multimodal ICL methods often rely on surface level imitation of in-context demonstrations, making it difficult for MLLMs to align their responses with the reasoning path required by the given multimodal input. This limitation becomes more pronounced in complex multimodal tasks, thereby restricting further improvements in MLLM performance. To address this issue, we propose a new multimodal ICL framework that combines contrastive demonstration modeling with the self-refinement capability of MLLMs. Specifically, our framework reformulates each demonstration by explicitly contrasting a suboptimal response with a better response under the same input, together with a reasoning path that reveals how the response should be refined. This contrastive formulation makes the reasoning path toward the desired response more explicit and guides the MLLM beyond superficial imitation. Furthermore, because effective refinement depends on the current response, we introduce a response-conditioned retrieval mechanism to select demonstrations whose reasoning paths are more relevant to the current response. In addition, we use a lightweight alignment controller to predict response quality and determine whether further refinement is needed. Experiments on three types of multimodal tasks show that the proposed framework consistently improves MLLM performance, with particularly notable gains on visual question answering (VQA).

Abstract:
Medical image segmentation needs sharp boundaries and strong global context. Transformers provide global modeling, but self-attention scales quadratically with resolution. CNNs are efficient, but long-range dependencies remain weak. We propose GISNet, an efficient U-shaped segmentation network built on Grouped Incremental Shift (GIS) and checkerboard-guided selective shifting (GIS_\textCB). GIS assigns progressively larger cyclic shifts within channel groups. Channels therefore capture context at different ranges. This yields an implicit multi-scale receptive field with near-zero extra overhead. GIS_\textCB shifts only half of spatial positions and keeps the rest as anchors. This reduces misalignment artifacts and preserves boundaries. We build a Dual-Path GIS Module (DPGM) and fuse global and local cues in parallel via a Spatial-to-Token Adapter Block (STAB) and a lightweight local convolution branch. Experiments on eight public datasets show strong accuracy--efficiency trade-offs at 2.1M--13.5M parameters. GISNet also surpasses large Transformer baselines on multiple benchmarks. Ablations confirm that incremental shifts and checkerboard anchoring reduce boundary errors (HD95). GISNet is a practical linear-complexity alternative for high-resolution medical image segmentation.

Abstract:
Existing camouflaged image generation methods typically violate natural principles by inducing subject attribute degradation or enforcing inauthentic background adaptation to foreground. They typically lack a robust and controllable generation paradigm, significantly limiting their practical applicability. To address these challenges, this paper introduces a new task, ``controllable camouflaged image generation'', which accommodates versatile user inputs, including subject, background, regional mask, and textual prompt, to enable controllable generation of camouflaged images. In response, we propose a novel, training-free ``\underlineCamouflage \underlineDiffusion Model'' (\underlineCamDiff) for this task, which seamlessly embeds the foreground object into the designated region of the background while effectively preserving the subject's features. Technically, we formalize camouflaged image generation through stochastic optimal control, integrating camouflage and anti-detection descriptors to model terminal costs. For this, we derive an optimal controller that optimizes the drift field of the reverse process, enhancing both the concealment and anti-detection performance of the synthesized images. Additionally, we design a controllable attention manipulation strategy to ensure high controllability throughout the camouflage process. Specifically, multi-head aggregation incorporates both subject and background information, while mask-restricted modulation confines the correlation between object-specific tokens and concealment regions. Extensive experiments demonstrate the effectiveness of CamDiff in camouflage generation across various domains.

Abstract:
In recent years, multi-view clustering has attracted widespread research interest. However, due to limitations in data collection devices, data across different views often suffer from misalignment, leading to the partial view alignment problem (PVAP). To mitigate the impact of view asymmetry and irrelevant samples, this paper proposes a framework for partial multi-view clustering via dual alignment and structure enhancement (DAS-PMVC), which leverages view structure consistency and semantic relevance. Specifically, DAS-PMVC includes three parts: anchor graph structure alignment, where sample joint embedding representations with consistent latent space are derived from anchor point relationships for initial view alignment; structure-enhanced feature learning, where the model learns view structure information through pretraining and combines multi-view graph convolutional networks to further extract deep latent features from the aligned graph structure to improve the discriminative power of representations; and a dual alignment strategy, where initial alignment is performed through the anchor graph in the pretraining phase, and contrastive learning loss and the Hungarian algorithm are introduced in the training phase to further optimize the alignment of latent features. Experimental results on various datasets demonstrate that the DAS-PMVC framework outperforms existing state-of-the-art methods in clustering performance, showcasing its effectiveness and superiority.

Abstract:
Video object detection aims to accurately localize and classify objects in video frames by effectively aggregating spatiotemporal features. Recently, DETR-based methods have achieved remarkable success in video object detection due to their strong global feature aggregation capability. Despite this, these methods generally rely solely on the visual modality for detection by using different transformers and attention mechanisms with quadratic computational complexity, leading to insufficient and inefficient feature aggregation as well as incomplete object query representations. Notably, video object detection can benefit not only from the visual modality providing rich appearance information, but also from the textual modality offering complementary semantic cues. However, effectively leveraging features from the textual modality for video object detection remains largely underexplored. In this paper, we propose TM2Q-Det, a novel framework based on Text-Guided Spatiotemporal Mamba and Multimodal Query Learning for Video Object Detection, which fully leverages textual features to facilitate feature aggregation and enhance object query representations. Specifically, we first propose a text-guided spatiotemporal Mamba module that leverages a significance-aware rearrangement Mamba to aggregate multi-level spatiotemporal features under the guidance of textual features, facilitating more sufficient and efficient feature aggregation. Second, we introduce a text-guided multimodal query module, which jointly leverages textual and visual features to enhance object queries through hierarchical multimodal feature interaction, providing more comprehensive object query representations. Experimental results on the ImageNet VID and EPIC-KITCHENS datasets demonstrate that TM2Q-Det outperforms state-of-the-art methods.

Abstract:
With the continuous convergence of edge intelligence and multimedia applications, edge-assisted video object detection has attracted increasing attention. To enhance system quality of service (QoS), existing studies primarily focus on full-frame scheduling and resource allocation, often overlooking the dominant role of the region of interest (RoI) in detection accuracy and lacking robust detection models for regionally blurred scenarios. To address these challenges, this paper proposes EVOD-RoI, an edge-assisted video object detection system based on RoI transmission and processing. It adaptively adjusts detection location, non-RoI resolution, and model selection to reduce non-RoI resolution during edge upload, minimizing overall latency. Meanwhile, it improves accuracy by fine-tuning a model specifically for regionally blurred frames. Specifically, EVOD-RoI introduces three key innovations: (1) a spatiotemporal joint sliding window algorithm to determine the optimal RoI, which saves bandwidth during transmission stage; (2) regionally blurred feature-aware model fine-tuning to improve detection accuracy in processing stage; and (3) a deep reinforcement learningbased approach to dynamically and adaptively optimize transmission and resource scheduling under heterogeneous environments. Experiments on a real-world testbed demonstrate that, compared with SOTA methods, EVOD-RoI improves detection accuracy by 15.9%30.1% while reducing overall system latency by 16.8%41.8%, significantly enhancing QoS in resource-constrained edge scenarios.

Abstract:
Multimodal automated fact-checking (MAFC) verifies claims by retrieving and reasoning over external evidence. However, most existing static benchmarks risk contamination: they primarily consist of outdated claims verifiable using an LLM's internal knowledge without external evidence. This can inflate performance estimates and fail to reflect true capability on novel claims that require up-to-date information. To address this, emerging dynamic benchmarks collect claims published after LLMs' knowledge cut-off dates, assuming they are uncontaminated. This work revisits this assumption by empirically studying contamination risks in both the state-of-the-art (SOTA) static AVeriTeC benchmark and our newly constructed dynamic ClaimReview2025Q4 benchmark, as well as their impact on MAFC evaluation. Our experiments yield 16 findings, highlighting three key results: (1) Dynamic evaluation reduces but does not eliminate contamination risks, as 17.09%--29.30% of post-cut-off claims remain potentially contaminated; (2) Many newly published claims can be verified either directly or by synthesizing multiple pieces of public knowledge available before the cut-off; and (3) Contamination can induce statistically significant inflation in MAFC performance, increasing Macro-F1 by up to 11.34 points and distorting system rankings. In light of these findings, we re-evaluate SOTA LLMs under a strictly contamination-controlled setting. Our study provides practical guidelines for trustworthy MAFC evaluation.

Abstract:
We present TT4D, a large-scale, high-fidelity table tennis dataset. It provides 140+ hours of reconstructed singles and doubles gameplay from monocular broadcast videos, featuring annotations like high-quality camera calibrations, precise 3D ball positions, ball spin, time segmentation, and 3D human meshes over time. This rich data provides a new foundation for virtual replay, in-depth player analysis, and robot learning. The dataset's combination of scale and precision is achieved through a novel reconstruction pipeline. Prior methods first partition a game sequence into individual shot segments using the 2D ball track, and only then attempt reconstruction. However, 2D-based time segmentation collapses under occlusion and varied camera viewpoints, preventing reliable reconstruction. We invert this paradigm by first lifting the entire unsegmented 2D ball track to 3D through a learned lifting network. This 3D trajectory then allows us to reliably perform time segmentation. The learned lifting network also infers the balls spin, handles unreliable ball detections, and successfully reconstructs the ball trajectory in cases of high occlusion. This lift-first design is necessary, as our pipeline is the only method capable of reconstructing table tennis gameplay from general-view broadcast monocular videos. We demonstrate the datasets fidelity through two downstream tasks: estimating the rackets pose \& velocity at impact, and training a generative model of competitive rallies.

Abstract:
Hairstyle transfer aims to synthesize a photorealistic portrait by transplanting the hairstyle from a reference image onto a source subject, while preserving the source's identity. While recent large-scale foundation models exhibit remarkable generative capabilities, they struggle with the zero-shot disentanglement required for such precise local editing, inherently entangling the reference hairstyle with its original facial identity and pose. To address these limitations through structural decomposition, a standard pipeline for hairstyle transfer typically decouples the process by first generating a ``bald'' image from the source and extracting identity-agnostic hairstyle features from a reference, fusing them to produce the final result. However, this methodology is frequently prone to several types of artifacts, including identity inconsistency and mismatched hair geometry. In this paper, we demonstrate that these artifacts stem from a more fundamental issue, which we term the leakage problem. This leakage is twofold: First, Identity Leakage in Hairstyle occurs when hairstyle features remain entangled with the reference's identity and pose. Second, Flaw Leakage in Bald arises when subtle geometric flaws left in the ``bald'' image are propagated into the synthesized hairstyle. To address these issues, we propose the Dual-Purification Framework (DPF), which integrates two complementary purification strategies. The Adversarial Hairstyle Purification (AHP) module explicitly purges identity information from hairstyle features by adversarially minimizing hairstylebald mutual information. Concurrently, the Contrastive Geometric Purification (CGP) module introduces a contrastive objective that penalizes the model's reliance on these geometric flaws in the ``bald'' image, thereby suppressing the Flaw Leakage in Bald. By explicitly mitigating both components of leakage, DPF achieves state-of-the-art performance in high-fidelity, identity-preserving hairstyle transfer.

Abstract:
Passive non-line-of-sight (NLOS) imaging promises remote awareness in safety-critical settings including search and rescue, privacy-preserving monitoring, autonomous robotics, and optical channel measurement attacks, where only a diffusive surface is visible. Yet single-shot reconstruction remains brittle: transport through the diffusive surface is strongly space-varying and couples near and far geometry, while photon statistics are heteroscedastic and easily confused with scene content. We present Radiance-Physics Diffusion (RaP-Diffusion), which treats the diffusive surface as a geometry- and noise-coded aperture. A Physics-Aware Photon Noise Synthesis forward process (PAPS) calibrates Poisson shot noise and Gaussian readout variability with signal-to-noise ratio (SNR) equalization, so the model learns sensor randomness rather than hallucinating it. A Liquid Radiance Adaptive-Kernel Ordinary Differential Equation denoiser (LiRa-AKODE) expands or contracts its receptive field according to local transport, preserving edges and long-range structure while staying within a Gaussian-conjugate reverse update. Across four datasets, CtM, StM, SmM and WtM, our single-image reconstructor improves PSNR, SSIM, LPIPS, FID, and diversity and remains stable under moderate viewpoint changes. On four indoor screen-content datasets and a 256256 setting, RaP-Diffusion achieves the above gains with 48.41 ms latency on an RTX 5090, with 133.55M parameters and 39.90G FLOPs, indicating practical efficiency within these operating conditions.

Abstract:
This paper addresses the challenge of reconstructing photorealistic and animatable 3D human avatars from monocular videos. While existing methods rely on combining per-subject optimization with generic human priors, they often fail to capture fine-grained details when training frames are limited. To mitigate this data scarcity, we propose TrioMan, a systematic tri-module framework for augmented 3D avatar learning. Our approach comprises three synergistic components. The Generator creates diverse unseen samples by imposing Gaussian perturbations on pose and camera. The Refiner improves the quality of generated data through one-step diffusion guided by texture and geometry cues. The Examiner selects subject-consistent samples using a dual-branch attention-based similarity evaluation. Experiments on the X-Humans and NeuMan benchmarks show that TrioMan outperforms state-of-the-art methods.

Abstract:
Due to the scarcity of real-world 3D indoor data, training on synthetic datasets and transferring to real environments has become a promising approach for enhancing embodied 3D perception. However, for the challenging 3D object detection (3DOD) task, which requires predicting accurate 3D object bounding boxes and class labels, the substantial domain gap between synthetic and real indoor data leads to severe performance degradation. Moreover, the few unsupervised domain adaptation methods for indoor 3DOD require access to both source and target data during training, leading to high computational cost and limited practicality when source data are unavailable. To address these challenges, we present the first source-free synthetic-to-real domain adaptation study for cross-modal indoor 3D object detection, leveraging 3D point clouds and associated multi-view 2D images to bridge the domain gap. We propose a novel framework that establishes both exploiting transferable priors cross modalities and training model for robust target-domain adaptation. Specifically, we leverage a 2D foundation model and a source-pretrained 3D detector to produce multi-view 2D masks and 3D proposals, respectively. Then the iterative pseudo-label refinement and soft optimal transport are conducted for cross-modal 2D-to-3D association, transferring semantic information from associated 2D masks to refined 3D proposals, yielding reliable open-set pseudo-labels. To further suppress residual detection errors and semantic noise, we introduce a robust training strategy with Gaussian prior regularization, which improves detector adaptation and helps model recover previously missed objects. Extensive experiments show that our method significantly outperforms existing domain adaptation and cross-modal 3DOD methods.

Abstract:
Token compression is crucial for vision-language models (VLMs) inference due to its tremendous computational complexity. Although substantial works with various model-driven methods have been done to mine importance rankings among tokens for compression~(e.g., rank according to attention scores or matrix ranks), they are all constrained by one-sided handcrafted information, thus being trapped in local optimum. To utilize comprehensive information for global optimum, we present a Data-driven Token Ranking (DTR) framework, which trains a plug-and-play token-ranking model with self-gathered token-ranking data for VLM token compression at runtime. Specifically, first, we propose a dataset construction method to efficiently gather importance rankings of tokens based on original VLM datasets. Then we present a training method to build a token-ranking model for predicting a ranked-list of token importance based on input vision and text tokens. Finally, the ranking model can be plugged in the model, then filter tokens with an user-defined token number at runtime for acceleration. Extensive experimental results across 8 mainstream benchmarks show that DTR achieves the state-of-the-art token compression performance compared with 8 challenging comparatives. Moreover, a comprehensive analysis shows that DTR as well as data-driven methods possess tremendous potential, which can comprehensively outperform the vanilla VLM with much fewer tokens.

Abstract:
Scanpath prediction is fundamental to understanding human viewing behavior and has broad applications across numerous domains. However, most existing methods focus on population-level attention modeling, and how personality traits shape individualized visual viewing behavior remains largely underexplored. To address this gap, we introduce BF360, a comprehensive benchmark featuring diverse immersive 360 scenes and fine-grained personality annotations for studying trait-dependent visual patterns. Our analysis reveals compelling insights: conscientiousness drives gaze to systematically focus on geometric scene boundaries, while high openness prompts high-entropy, divergent exploration of unstructured surrounding areas. Based on these insights, we propose the Personality-driven Hierarchical Network (P-HNet), a novel model that incorporates personality traits as conditioning signals for individualized gaze inference. Specifically, P-HNet comprises three core modules: (1) a personality-guided multimodal interaction module that utilizes personality embeddings as a conditioning signal to guide the modulation of visual preferences, enabling semantically robust multimodal representation learning; (2) a hierarchical adaptive attention module that effectively derives personality-related visual exploration tendencies; and (3) a spatial-probabilistic generative module that captures the intricate temporal dependencies of fixation transitions, thereby synthesizing realistic and individualized viewing trajectories. Extensive experiments demonstrate that P-HNet effectively predicts individualized scanpaths on BF360 and generalizes well to other datasets without personality annotations. This work advances the understanding of personality-driven eye movements and provides a scalable framework for personalized vision modeling.

Abstract:
Multimodal Large Language Models (MLLMs) have demonstrated great performance, yet they often suffer from severe modality misalignment when confronted with densely constrained spaces for interior design. Due to the loss of high-frequency local topological details and fine-grained aesthetic shifts during visual encoding, existing MLLMs frequently hallucinate, yielding physical spatial collisions and visual aesthetic dissonance. To address this, we propose Dual-prior Activation Residual Task-vectors Injection mechanism (DART-I) for MLLMs. It shifts the paradigm from lossy text-prompting to direct latent intervention, utilizing weak-to-strong deterministic rules to anchor the causal reasoning of MLLMs for interior design. Specifically, DART-I operates in three steps: it first explicitly extracts continuous spatial distance and color typography features from images using extremely lightweight weak experts; subsequently, it transforms these deterministic priors into directional task vectors via a linear projection network; these vectors are dynamically injected as residual terms into the latent space of the frozen MLLMs, steering MLLMs towards precise reasoning for interior design. Stepping outside the conventional paradigms, our method achieves precise reasoning without fine-tuning the MLLMs, effectively bypassing expensive computational costs and catastrophic forgetting. Extensive experiments on various benchmarks demonstrate the effectiveness and advantages of DART-I.

Abstract:
Panoptic Scene Graph generation (PSG) requires holistic yet fine-grained reasoning over complex relations among objects. However, current Multimodal Large Language Models (MLLMs) applied to PSG suffer from a perception bottleneck: they function as static observers limited by fixed-resolution views. As a result, models often overlook critical relations involving small or peripheral objects and struggle to differentiate ambiguous predicates in complex scenes. To address these limitations, we propose PSGAgent, an agentic framework that enhances visual perception for PSG through a think--act--observe loop. PSGAgent allows MLLMs to dynamically refine their visual perception using two tools: a Region Zoom-In Tool for local exploration of small or boundary relations and a Visual Highlight Tool for relation-focused segmentation that filters out irrelevant contextual noise. To optimize decision-making, we construct a two-stage training pipeline. First, we create a high-quality, error-aware Multimodal Chain-of-Thought (M-CoT) dataset that provides robust supervision for cold-start learning. Then, we design a novel hierarchical reward mechanism integrated with Group Sequence Policy Optimization (GSPO) to enable strategic and adaptive tool usage. Extensive experiments on the PSG benchmark show that PSGAgent achieves SOTA performance, demonstrating strong capability in both holistic and fine-grained scene understanding.

Abstract:
The inherent irregularity and lack of structure in point cloud data present fundamental challenges for efficient processing, representation learning, and spatiotemporal modeling. Point Geometry Image (PGI) methods offer a promising direction by regularizing point clouds into structured 2D representations, yet existing approaches typically learn direct mappings from 3D point clouds to 2D grids, placing a heavy burden on the network to simultaneously discover global organization and local geometric refinement. This often results in unstable geometric correspondences and limited temporal consistency. We propose PGI-Net, a geometry-guided structured representation framework that converts 3D point clouds into compact and regular 2D Point Geometry Images through spherical and octahedral geometric initialization. A multi-stage progressive refinement strategy with adaptive resampling further improves local neighborhood preservation, global geometric fidelity, and temporal coherence across frames. The resulting representation is compact, structured, and reconstruction-consistent, while being readily compatible with standard 2D backbones for downstream analysis. Extensive experiments demonstrate that PGI-Net achieves strong geometric fidelity, structural consistency, and computational efficiency, while effectively supporting downstream tasks, including action recognition, temporal interpolation, and point cloud compression, establishing geometry-guided PGI construction as an effective paradigm for structured point cloud representation.

Abstract:
The rapid evolution of multimodal large models has revolutionized the simulation of diverse characters in speech dialogue systems, enabling a novel interactive paradigm. Character attributes are manifested not only in textual responses but also through vocal features, as speech conveys rich paralinguistic information that is challenging to quantify. This poses significant difficulties in evaluating the character alignment of role-playing agents. To address these challenges, we present RoleJudge, an evaluation framework that leverages audio large language models to systematically assess the alignment between speech and character across multiple modalities and dimensions. Furthermore, we introduce RoleChat, the first voice role-playing evaluation dataset enriched with chain-of-thought reasoning annotations, comprising a diverse set of authentic and LLM-generated speech samples. Utilizing this dataset, we implement a multi-stage training paradigm and incorporate Standard Alignment in reinforcement learning to mitigate reward misalignment during optimization. Experimental results in terms of accuracy and subjective assessment demonstrate that RoleJudge outperforms various baseline models, validating the effectiveness of our multidimensional evaluation framework.

Abstract:
Olfactory feedback enhances immersion in virtual reality (VR), yet its deployment is constrained by the limited number of odor sources. We investigate whether visual cues can influence users' interpretation and affective evaluation of the same physical odor in immersive environments without increasing hardware complexity. We conducted a formative study to identify odor characteristics susceptible to visual reinterpretation, a pre-study to calibrate odor intensity, and a within-subject VR experiment manipulating contour recognizability (concrete vs. abstract) and lighting conditions (achromatic lighting, cool-toned lighting, warm-toned lighting). Participants evaluated virtual objects paired with two base odorants in terms of odor matching, perceptual clarity, and affective responses. Results demonstrate the feasibility of visually guided odor reinterpretation, with an average matching success rate of 80.45% across the three phenylethanol-associated objects. Our findings show that visual design systematically influences users' interpretation of odor categories and their experience of pleasure through ease of recognition, authenticity, and the color atmosphere created by lighting. We further discuss how visual cues can influence odor interpretation and how cross-modal olfactory illusions affect the virtual experience.

Abstract:
Road anomaly segmentation is a safety-critical open-set perception task in autonomous driving. However, existing evaluation is largely scale-agnostic, making it difficult to reliably assess model sensitivity across anomaly scales and potentially obscuring severe failures in mixed-scale scenes. In real driving environments, anomalies often appear at highly different scales, and dense prediction models tend to favor large and visually dominant anomalous regions, making weaker responses from small anomalies easier to suppress. To better understand this issue, we first conduct a scale-aware diagnosis and introduce MS-Protocol, a scale-aware diagnostic evaluation protocol for road anomaly segmentation. It defines anomaly scale in a resolution-invariant manner, reorganizes existing benchmarks into a common scale-aware evaluation suite, and includes MS-ROAD, a diagnostic subset tailored to mixed-scale analysis. Under this protocol, we observe a consistent failure mode of current models. When anomalies of different scales coexist, sensitivity to \emphsmall-but-hazardous anomalies drops markedly. Motivated by this finding, we propose SHIELD, a coarse-to-fine framework built on \emphregion-to-pixel hazard reasoning. SHIELD first predicts a coarse anomaly map, then harvests small-sensitive candidate regions, reasons about their hazardness using scale-aware visual-semantic cues, and finally injects the resulting small-hazard evidence back into dense prediction for refinement. In this way, SHIELD explicitly complements pixel-level anomaly prediction with region-level hazard assessment. Experiments across scale-aware and standard benchmarks show that SHIELD effectively alleviates this mixed-scale failure mode, improving sensitivity to \emphsmall-but-hazardous anomalies in challenging mixed-scale scenes while remaining competitive under conventional evaluation.

Abstract:
It is well-known that audio classifiers often rely on non-musically relevant features and spurious correlations to classify audio. Hence, audio classifiers are easy to manipulate or confuse, resulting in wrong classifications. While inducing a misclassification is not hard, until now the features that the classifiers rely on have not been well understood. In this paper we introduce a new method that uses post hoc causal reasoning to reveal those features of the frequency space that are sufficient and necessary for a given classification. We describe an implementation of this algorithm in the tool \textscFreqReX and provide experimental results on a number of standard benchmark datasets. Our experiments show that sufficient and necessary frequency subsets are discoverable and allow us to manipulate model outputs by changing only sufficient parts of the signal. Namely, a change to 1 out of 240,000 frequencies results in a change in classification 33% of the time, and the change can often be localized to just a part of the signal. These results show that causal analysis can usefully extract the information relied upon by audio classifiers in their decision process and can also be used to successfully manipulate their outputs.

Abstract:
Most snapshot spectral imaging systems still adhere to a reconstruction-first paradigm: a compressed, coded measurement is first reconstructed into a hyperspectral image and only then passed to a downstream task model. While natural from the standpoint of inverse imaging, this design hardwires a stronger assumption than necessarythat semantic perception must be mediated by explicit reconstruction. In this work, we challenge that assumption for camouflaged object detection (COD) in filter-array-based snapshot spectral imaging. We show that COD can be performed directly from compressed measurements, without reconstructing a hyperspectral cube. Our central observation is that the coded measurement, despite severe compression, retains exploitable spatialspectral structure for downstream camouflage reasoning. Building on this insight, we formulate snapshot spectral COD as a reconstruction-free measurement-to-task problem and introduce the Measurement-domain Camouflaged Object Detection Network (MeasCODNet), a hardware-aware framework that unifies mosaic-aware measurement encoding, spectral-response-conditioned feature learning, and dense COD prediction in a single end-to-end model. Rather than expending computation on an intermediate image formation objective, MeasCODNet leverages physical priors from the sensing process itself to learn task-oriented representations natively in the measurement domain. Experiments show that this measurement-native formulation approaches the performance of conventional reconstruction-then-task pipelines while substantially simplifying the overall workflow and improving computational efficiency. More broadly, our results argue that explicit hyperspectral reconstruction is not an obligatory precursor to snapshot spectral perception, and position task-first inference in the measurement domain as a compelling new direction for snapshot spectral understanding.

Abstract:
Multimodal embodied systems interact within highly dynamic environments, necessitating pre-trained models to adapt to continuous distribution shifts during deployment. This critical demand gives rise to the task of Continual Test-Time Adaptation (CTTA), which aims to continuously update models using unlabeled data streams in target environments to prevent performance collapse. However, existing methods, predominantly originating from pure vision tasks, rely heavily on passive observational data for model updates, thereby overlooking the most crucial aspect of embodied interaction: action stability. Neglecting action stability allows the model's kinematic representations to be corrupted by transient visual alignment, inevitably leading to severe error accumulation during continuous execution and catastrophic forgetting of original physical manipulation skills. To address this challenge, we introduce See to Act, Adapt to Act, pioneering a new CTTA paradigm for embodied scenarios. Fundamentally, this mechanism relies on real-time perception to extract representations (See to Act) while continuously calibrating feature representations on the fly to ensure subsequent actions remain robust against perturbations (Adapt to Act). Specifically, in the See to Act phase, the model explicitly extracts and decouples transient and steady features. Subsequently, in the Adapt to Act phase, the framework leverages a teacher-student architecture to implement a boundary-aware feature updating mechanism alongside an adaptive action generation mechanism. This design utilizes either steady or transient features in a targeted manner, driving the model to continuously output highly successful and stable action sequences. Extensive evaluations on our newly proposed closed-loop ALOHA-C benchmark demonstrate that the proposed framework outperforms existing baselines across diverse, highly dynamic scenarios, achieving performance gains of over 40% under multiple settings. Our results demonstrate that robust embodied system requires a paradigm shift from passive perceptual alignment to active action stability.

Abstract:
Modeling 4D human-object interaction (HOI) is a compelling challenge in computer vision and an essential technology powering virtual and mixed-reality applications. While existing works have achieved promising results on specific HOI taskssuch as text-conditioned HOI generation and human motion generation from object motion, they typically rely on task-specific architectures and lack a unified framework capable of handling diverse conditional inputs. Building on this, we propose Uni-HOI, a unified framework that learns the joint distribution among text, human motion, and object motion. By leveraging large language models (LLMs) and two motion-specific vector quantized variational autoencoders (VQ-VAEs), we convert heterogeneous motion data into token sequences compatible with LLM inputs, enabling seamless integration and joint modeling of all three modalities. We introduce a two-stage training strategy: the first stage performs multi-task learning on a large-scale HOI dataset to capture the underlying correlations among the three modalities, while the second stage fine-tunes the model on specific tasks to further enhance performance. Extensive experiments demonstrate that Uni-HOI achieves remarkable performances on multiple HOI-related tasks including text-driven HOI generation, object motion-driven human motion generation (optionally with text) and human motion-driven object motion prediction within a unified framework.

Abstract:
In real-world federated systems, distributed clients frequently engage in multiple tasks with diverse modality combinations, highlighting the practical need for Federated Multimodal Multi-task Learning (FMML). However, task heterogeneity poses a fundamental challenge: substantial differences in task objectives, modality compositions, and model structures render a single shared model insufficient. While this necessitates personalization, excessive localization fragments the federation into isolated models, severely limiting global generalization. Therefore, the core objective in this setting is to synthesize a highly generalizable global model while strictly preserving the distinct personalized capabilities of heterogeneous tasks. To achieve this, we propose FedM^3, a Mixture-of-Experts framework explicitly designed for FMML. By treating each client as an independent expert, FedM^3 inherently preserves task-specific structures and multimodal decoding patterns. Simultaneously, it organizes these heterogeneous local experts into a unified global expert bank, empowering a single global model to seamlessly absorb transferable knowledge across all clients. To bridge local specialization and global learning, FedM^3 shifts from traditional parameter-level alignment to representation-level alignment via conflict-averse distillation and adaptive resampling, complemented by task-aware routing for dynamic expert selection across diverse multimodal queries. Extensive experiments across highly heterogeneous vision, text, and multimodal benchmarks validate our approach.

Abstract:
Medical images constitute a class of highly complex visual multimedia data, where accurate segmentation heavily relies on the synergistic modeling of global semantics and fine-grained local details. However, existing deep learning methods often fall short in depicting complex geometric morphologies and facilitating efficient feature interactions, while suffering from the inherent explainable bottleneck of black-box decision making. To address these issues, we propose SGDA-Net, a novel segmentation network that integrates second-order weighted graph convolution and dynamic agent attention within an asymmetric encoder-decoder architecture. In the decoder, we introduce a second-order weighted graph convolution that explicitly models second-order curvature, enabling precise reconstruction of boundary structures and complex geometric morphologies. Complementing this, a dynamic agent attention module is designed to employ a saliency-guided strategy to adaptively focus on key regions with high feature activation levels. This synergistic mechanism, based on geometric curvature perception and saliency-driven feature selection, not only achieves efficient fusion of global and local features but also provides feature-level transparency for network decisions. Consequently, our method effectively enhances computational explainability by offering intuitive visual cues for complex clinical reasoning. Extensive experiments on two public benchmarks demonstrate that SGDA-Net yields average Dice coefficients of 85.44% on the Synapse abdominal multi-organ CT dataset and 92.44% on the ACDC cardiac MRI dataset, outperforming the existing state-of-the-art methods with particularly significant performance gains on challenging, small-volume, and morphologically complex targets (\emphe.g., gallbladder and pancreas). Our work illustrates how combining geometry-aware and attention-driven mechanisms can advance both the performance and transparency of understanding medical multimedia content.

Abstract:
Short-answer VQA benchmarks often mix two distinct aspects of model performance: whether an answer is semantically correct and whether its surface form matches the expectation of the automatic evaluator. We study this mismatch across six vision--language models and six benchmarks, using a semantic judge validated by human annotations (97.6% precision) to audit over 37k rejected answers. A second text-only judge yields consistent benchmark-level mismatch patterns, showing that the observed effect is not an artifact of a single audit model. On text-rich benchmarks, up to 47.7% of rejected answers are judged semantically acceptable despite being marked incorrect by the benchmark evaluator. Evaluator sensitivity varies systematically across answer types, with readout and multi-span answers being particularly evaluator-sensitive compared with scalar answers. Alternative short-answer prompt formulations further destabilize benchmark outcomes, producing substantial item-level flips for the same image--question pairs. A deterministic CPU-only contract repair further shows that part of this undercount is recoverable.

Abstract:
Infrared (IR) and Visible (VIS) image segmentation integrates multi-modality complementary cues to assign accurate semantic labels to pixels. While successful on pre-aligned images, existing IR-VIS segmentation methods are vulnerable to pixel-level misalignments that disrupt multi-modality complementarity, resulting in label mismatches. Existing cascaded registration-fusion-segmentation pipelines, while improving segmentation accuracy, compromise the adaptability of visual and semantic features, which limits potential gains. Therefore, this paper proposes a Tri-optimized \textbf Feature-level \textbf Registration and \textbf Segmentation method, termed Tri-FRS. Specifically, to mitigate label mismatches, an end-to-end segmentation network for misaligned IR-VIS images is introduced to perform deformable registration directly in the feature space, ensuring semantically consistent feature learning. To ensure semantic consistency between feature-level registration and segmentation, we establish a tri-level optimization learning strategy. It leverages self-learning gradient responses in the middle-lower layers to reinforce the coupling between registration and segmentation, while employing an automated search to determine the optimal loss balance in the upper-middle layers, ultimately improving segmentation accuracy. In addition, to bridge the modality gap, we introduce a Cross-modality Image Translation Engine (CITE) that leverages LoRA fine-tuning on a diffusion-based style transfer model to generate high-quality pseudo-IR images from VIS images. Extensive experiments show Tri-FRS is highly robust to segmentation even under severe misalignments.

Abstract:
Multimodal empathetic dialogue generation aims to produce emotionally nuanced and compassionate responses. However, existing multimodal empathetic response generation (MERG) methods primarily focus on the current query, overlooking the emotional dynamics across dialogue history. Moreover, they lack effective mechanisms to handle cross-modal emotional inconsistencies that frequently arise in real-world interactions. To address these limitations, we propose MEC^2-TT, a novel framework that integrates Multimodal Emotion Consistency Correction and Trajectory Tracking, enabling more coherent and context-aware empathetic responses. Specifically, an Emotion Consistency Correction module is first introduced to mitigate cross-modal inconsistencies via cross-modal alignment. Afterwards, we design an Emotion Trajectory Tracking module to capture the temporal evolution of user emotions throughout the dialogue. Finally, we leverage large language models with structured reasoning to improve empathetic understanding and guide response generation. Comprehensive experiments on the AvaMERG and MELD benchmarks, including automatic and human evaluations, demonstrate that our method consistently outperforms existing state-of-the-art approaches, particularly in emotional alignment, contextual coherence, and empathetic response quality.

Abstract:
Unmanned Aerial Vehicle (UAV) detection is important for low-altitude security, yet collecting and annotating large-scale real-world UAV data is difficult. Simulation is an attractive substitute, offering scalable annotations and same-scene multimodal observations, such as infrared (IR) imagery, whose thermal contrast keeps small, low-contrast UAVs salient in poor illumination and adverse weather. In real deployment, however, most UAV systems carry only a visible-light camera, leaving this multimodal advantage seemingly unusable. Existing simulation-to-reality (sim-to-real) methods and multimodal detectors typically either focus on unimodal transfer or require all modalities at inference; neither fits RGB-only deployment. To address this, we propose a deployment-constrained sim-to-real UAV detection framework with labeled simulated RGB and IR data available in the source domain, while the target domain and final deployment remain strictly RGB-only. It features two components: 1) a cross-modal knowledge distillation scheme that embeds thermally informed priors from a frozen IR diffusion detector into an RGB branch; 2) a diffusion-guided Cross-Step Proposal Consistency (CSPC) module for unsupervised domain adaptation (UDA). By exploiting intrinsic cross-step denoising stability beyond confidence, CSPC calibrates proposal reliability and carries distilled IR priors across the sim-to-real gap. Experiments on Anti-UAV300 and the Real-World dataset show that our method achieves state-of-the-art performance with 36.5 and 23.1 AP, respectively. The results demonstrate that source-side simulated IR priors can be effectively transformed into real-world gains for RGB-only UAV detection.

Abstract:
\beginabstract High-resolution visual inputs can substantially enhance the fine-grained understanding of documents, charts, and complex layouts in multimodal large language models (MLLMs), but they also incur higher inference latency and greater memory overhead. Existing visual token compression methods primarily rely on pruning or merging strategies, yet they remain inadequate for document-centric scenarios: the former may erroneously discard low-saliency but semantically critical regions, such as small-font text and symbols, while the latter may attenuate the high-frequency local details essential for OCR through feature averaging. To address these limitations, we propose MERIT, a training-free visual token compression framework for high-resolution document understanding. MERIT introduces a multi-view token importance estimation mechanism based on inter-layer dynamics, jointly modeling token importance from three complementary perspectives: global semantics, local structure, and signal saliency. Building on this design, we further develop Dynamics-Aware Proxy Compression (DAPC), which preserves high-value tokens via feature-faithful hard sampling and compensates for the influence of compressed regions in subsequent attention through quality redistribution. Experimental results demonstrate that MERIT achieves strong performance across document understanding, OCR, and general multimodal benchmarks. On LLaVA-OneVision-7B, MERIT retains 97.6% and 93.0% of the original performance at compression ratios of 66.7% and 77.8%, respectively, while reducing time-to-first-token (TTFT) by 62.9% and 60.7%, respectively, relative to the best competing baselines under the respective settings. These results suggest that dynamics-aware, feature-faithful compression provides an effective pathway toward efficient high-resolution multimodal document understanding. \endabstract

Abstract:
Vision-Language Agents (VLAs) that operate graphical user interfaces (GUIs) face a distinctive privacy-alignment problem: sensitive information is embedded in screen regions and interaction targets, so failures can arise through both textual disclosure and grounded actions. Existing text-centric safety guardrails are poorly matched to this setting because they treat privacy as a fixed property of an entity type, rather than as a context-dependent information-flow constraint. We formulate this problem as Multimodal Contextual Integrity (M-CI), where legality depends on the current contextual relation among actors, attributes, and transmission principles. To operationalize M-CI, we propose the Visual-Agent Guardian (VAG), which represents each interaction state using a contextual tuple together with a region-level authorization map over GUI elements. VAG combines a context-conditioned inverse rendering pipeline for counterfactual supervision, an M-CI-CoT module for contextual tuple and authorization inference, and a Visual-Semantic Reward Model (V-SRM) that scores both textual and spatial actions against the inferred CI state. We train the agent under a privacy budget using Constrained Group Relative Policy Optimization (C-GRPO), a critic-free constrained reinforcement learning method with PID-Lagrangian updates, and further improve robustness through automated adversarial self-play. Experiments on privacy-critical benchmarks demonstrate that VAG fundamentally shifts the privacy--utility Pareto frontier: it reduces visual privacy leakage to under 5% across all three benchmarks while averaging over 80% task success, and reduces the average multimodal social-engineering attack success rate to below 4%.

Abstract:
\beginabstract Action-driven stochastic human motion prediction is a fundamental task for achieving controllable motion generation. However, existing approaches often suffer from three major limitations: (1) Semantic inconsistency between generated motions and action labels occurs because certain action categories share highly similar semantic features; (2) Incoherent transitions, as generating smooth connectivity is hindered by the high variability in transition speeds and kinematic patterns across different actions; and (3) Compounded error accumulation inherently exists in multi-step autoregressive generation, leading to progressive quality degradation in long-horizon sequences. % To surmount these challenges, we propose Anchor-based Residual Diffusion Guided by Motion Library (ARDML). To address these challenges, we propose Anchor-based Residual Diffusion Guided by Motion Library (ARDML), a novel framework designed for human motion prediction. At its core, we formulate a Residual Diffusion Mechanism explicitly anchored to Action Prototypes. Unlike conventional diffusion models that denoise from uninformative Gaussian noise, our framework leverages Action Anchors as foundational kinematic references. % By re-defining the diffusion objective as learning the residual refinement relative to these semantic anchors, we effectively bridge the gap between abstract labels and high-fidelity motion details. This synergistic coupling not only mitigates semantic misalignment but also enables non-autoregressive, multi-step generation, ensuring long-term physical plausibility. To further enhance transition naturalness, we construct a Motion Library as a retrieval-augmented prior, providing fine-grained guidance on style and velocity. Extensive experiments demonstrate that ARDML significantly outperforms state-of-the-art baselines. We reformulate the diffusion objective as learning a residual refinement with respect to semantic anchors, thereby effectively bridging the gap between abstract labels and high-fidelity motion details. This synergistic coupling not only alleviates semantic misalignment but also enables non-autoregressive, multi-step generation, ensuring long-term physical plausibility. To further enhance the naturalness of transitions, we construct a Motion Library that serves as a retrieval-augmented prior, offering fine-grained guidance on style and velocity. Extensive experiments demonstrate that ARDML substantially outperforms state-of-the-art baselines. \endabstract

Abstract:
Despite significant advances in instruction-following and auditory comprehension, the evaluation of Emotional Intelligence (EI) in Spoken Language Models (SLMs) remains confined to rudimentary paralinguistic perception, lacking a systematic, theory-driven cognitive framework. We introduce EmoSBench, the first comprehensive EI evaluation benchmark for SLMs constructed upon the four-branch theoretical model, covering Perceiving, Understanding, Using, and Managing Emotion across ten sub-tasks. Preliminary assessments on EmoSBench reveal a substantial gap: even leading proprietary models like GPT-4o-Audio achieve only 52.6%, significantly trailing human baselines. To bridge this gap, we develop EmoS, a specialized evaluator model optimized via Supervised Fine-Tuning (SFT) and Group Relative Policy Optimization (GRPO). To facilitate its effective training, we curate EmoDialogue, a bilingual dataset providing necessary fine-grained supervision through response pairs with rigorously defined EI gradations. Concurrently, we introduce a reward mechanism integrating a Steep Exponential Accuracy Reward (SEAR) and a Rationale Fidelity Reward (RFR) to enforce precise ordinal scoring and valid reasoning. Experiments demonstrate that EmoS reaches 83.8% accuracy, approaching human-level performance. Furthermore, evaluations on authentic, unconstrained spoken interactions validate its robust real-world generalization, establishing a foundational framework for advancing emotionally intelligent dialogue systems.

Abstract:
Large vision-language models (LVLMs) have demonstrated remarkable capabilities in multimodal understanding. However, they remain prone to hallucinations, generating responses that are inconsistent with the visual evidence. Existing mitigation methods largely address language-prior bias or cross-modal imbalance, while progressive visual degradation across perception and memory remains underexplored. In this work, we propose Saliency-Driven Perceptual Realignment (SDPR), a training-free framework that mitigates the degradation of visual awareness throughout inference. Specifically, we first introduce saliency-driven attention redistribution to release attention hijacked by non-semantic sink tokens, thereby recovering critical visual evidence. Second, we identify spatial distortion in the KV cache and propose saliency-driven cache alignment to preserve query-relevant visual features during generation. Finally, we introduce prior-constrained contrastive decoding to penalize unfaithful predictions induced by dominant language priors. Through holistic perceptual realignment across the generation process, SDPR robustly mitigates hallucinations. Extensive experiments across diverse LVLM architectures show that SDPR outperforms state-of-the-art methods on both hallucination and general-purpose benchmarks, requiring no additional training and incurring minimal runtime overhead.

Abstract:
\beginabstract Vision-language models (VLMs) generate hundreds to thousands of visual tokens per image, leading to high computational cost and memory consumption during multimodal reasoning. A common efficiency strategy is to reduce input resolution; however, pixel-level downsampling often removes fine-grained semantic information and degrades performance on downstream tasks. We propose CRISPR, a context-refined information spatial pooling framework with region-awareness for visual token compression. Instead of reducing image resolution, CRISPR operates on encoded visual representations and compresses them into a smaller set of topology-preserving tokens. A \emphToken Refiner first enriches local tokens with global context through self-attention, followed by a \emphLocal Token Compressor that performs block-wise cross-attention aggregation while maintaining spatial coherence, and a \emphGlobal Token Fusion module that enables information exchange among compressed tokens. CRISPR supports controllable efficiency--accuracy trade-offs through the block-size parameter; in this paper, we evaluate 9× and 16× token reduction. Extensive experiments on multimodal benchmarks, including visual question answering and image captioning, demonstrate that CRISPR consistently outperforms low-resolution baselines at a comparable compressed sequence length. The proposed approach reduces both inference FLOPs and prefill latency. \endabstract

Abstract:
Generic video object detection has advanced considerably in recent years. Nevertheless, its application to drone-view videos, which are often characterized by low object distinctiveness within vast frames and imbalanced object density across frames, still remains largely underexplored. Existing generic video object detection methods do not work well on drone-view videos, as they typically rely on visual features solely and fail to exploit contextualized text prompts of objects. In this paper, we propose a Language-Informed Mamba Diffusion Network (LMDNet) for drone-view video object detection, with its key insight of leveraging the vision-language model (CLIP) to produce contextualized text prompts based on visual features to guide both spatiotemporal feature aggregation and object query generation. Specifically, our LMDNet includes the following two principal modules: (1) a prompt-guided dynamic Mamba module that adaptively adjusts the scanning path to prioritize task-relevant features and aggregates multi-scale spatiotemporal information that is likely to encompass target objects or contextually important cues, enabling more discriminative feature representations in intricate drone-view video scenes; and (2) a prompt-guided conditional diffusion module that adaptively determines the number of object queries for each frame and refines object queries in a progressive paradigm, enabling efficient and accurate video object detection under imbalanced object densities in drone-view video scenes. Extensive experiments on the VisDrone-VID and UAVDT datasets show that our LMDNet achieves up to 3.1%?11.3% higher mAP against state-of-the-art methods.

Abstract:
Weakly-Supervised Domain Adaptive (WSDA) object detection aims to transfer knowledge from a fully annotated source domain to a target domain with only image-level labels. Most existing methods attempt to extract Domain-Invariant Attributes (DIA) by globally aligning feature distributions across domains. However, certain implicit domain knowledge that carries commonalities, such as object contours and morphology, is often overlooked. We categorize the knowledge contained in such features as Domain-Similarity Attributes (DSA). We argue that this simplistic domain alignment strategy that ignores DSA may lead to inadequate model learning. If DSA can be properly decoupled from DIA, DSA can provide meaningful structural cues to the detector. To this end, we propose the dynamic feature disentanglement framework via Manifold-Aligned Representation Splitting (MARS) for WSDA object detection. Our method consists of three key modules. Firstly, we design the Semantic Similarity Matching (SSM), which constructs image pairs with similar semantics by dynamic matching, providing a semantic similarity basis for disentanglement. Secondly, we design the Manifold-based Representation Splitting (MRS), which projects domain features onto a shared manifold. MRS regards DSA as an orthogonal and complementary knowledge source to the common DIA, and decomposes them using a dynamic feature disentanglement strategy with learnable decomposition gates. Thirdly, we design the Parallel Knowledge Mining (PKM) module, which separately extracts knowledge from DIA and DSA. Extensive experiments demonstrate that our proposed framework achieves state-of-the-art (SOTA) performance across various transfer scenarios.

Abstract:
Recent multimodal large language models (MLLMs) have shown impressive performance on various vision-language tasks. However, existing paradigms typically produce a generic response conditioned only on the image and the user's direct prompt, without considering the user's underlying intent. In practical scenarios, instructions may be brief and ambiguous, yet users still expect models to understand their actual needs and provide responses tailored to their background, goals, and preferred knowledge depth. To better integrate user intent into typical MLLM interactions, we introduce intent-oriented visual dialogue, a unified formulation that treats diverse vision-language tasks as the interpretation and expansion of image content conditioned on different user needs. To address such scenarios, we propose the IntentPro framework. At the beginning of the dialogue, it proactively infers user interests from the image content, user profile, and initial prompt, and further refines the query through interaction. Once the intent is clarified, the model plans an appropriate response strategy, and then engages in tool-aware chain-of-thought reasoning, which invokes external tools, such as grounding, retrieval, and OCR, for more accurate and informative generation. To support model training and evaluation, we construct InfoVD, a multi-domain dataset for intent-sensitive multimodal interaction, including the form of captioning, VQA, image narrative generation, etc. Experiments across diverse specialized domains, including encyclopedic knowledge, food, and fashion, demonstrate that our approach effectively captures user intent and enhances response accuracy by leveraging external tools, highlighting the practical potential of intent-aware interaction.

Abstract:
RGB-infrared detectors typically discard fusion-time statistics, leaving downstream modules unable to assess interaction reliability. We preserve these statistics as a parameter-free, 7-dimensional spectral reliability descriptor spanning band energy, amplitude ratio, phase consistency, and cross-modal correlation. The descriptor drives Spectral Reliability Fusion (SRF), which gates a spectral residual against a conservative spatial base, and Reliability-Conditioned Expert Routing (RCER), which combines it with pooled content to route sparse post-fusion experts. Under matched ablations, descriptor-aware gating improves mAP50 over content-only adaptive gating; a 22 factorial analysis further shows that descriptor-conditioned routing provides the larger marginal gain over expert architecture alone with comparable parameter counts. Across six DroneVehicle degradations, average retention reaches 95.0%, compared with 92.0% for content-only MoE and 87.9% for concatenation, with the largest gain under modality drop; natural day/night mAP50 improves by +5.2/+5.3. These results show that an explicit fusion-reliability signal benefits both adaptive fusion and conditional computation.

Abstract:
Although infrared small target detection (IRSTD) have achieved notable progress, real-world infrared data often exhibit domain shifts, requiring models to continually adapt to newly arriving domains. Directly fine-tuning a model often causes severe forgetting of previously acquired knowledge. In this paper, we formulate Incremental InfraRed Small Target Detection (I^2RSTD) and show that catastrophic forgetting in this setting mainly arises from background distribution shift. To facilitate research, we first construct a benchmark based on five public datasets under three domain orders. Building upon this benchmark, we propose a prompt-based framework with a freeze-and-expand strategy, which maintains three domain-specific parameter pools, namely the Domain-Aware Adapter (DAA) Pool, the Domain Prompt Query (DPQ) Pool, and the Domain Prediction Head (DPH) Pool, to support domain-adaptive feature extraction and target detection. Additionally, we introduce a Domain Identification Module (DIM) to infer the domain ID of test image, enabling the model to retrieve the corresponding domain-specific parameters during inference. Extensive experiments demonstrate that our method consistently outperforms representative continual semantic segmentation methods and effectively alleviates catastrophic forgetting. The source code is available in the supplementary materials.

Abstract:
Mixed Reality (MR) is increasingly used in exhibition settings to situate digital artworks within physical environments. However, existing MR exhibition systems are often limited to prototypes or case-specific deployments, providing limited guidance for large-scale implementation. To address this gap, we present a practical pipeline for designing and deploying large-scale MR art exhibitions, framing spatial alignment as both a technical mechanism and an experiential design decision. We first conducted a pilot study comparing marker-based and Simultaneous Localization and Mapping (SLAM)-based alignment methods in an MR exhibition setting. Based on the results, we developed a SLAM-based pipeline that coordinates technical deployment with exhibition curation. We then evaluated the pipeline through system-level measures and visitor feedback in a real-world, large-scale exhibition. The findings suggest that spatial alignment is associated not only with technical stability but also with exhibition coherence, visitors' perceived continuity and immersion, and artwork interpretation. This work provides an empirically grounded reference for future large-scale MR art exhibition deployment.

Abstract:
\beginabstract Leveraging vision large language models (VLMs) for long frame medical imaging is hindered by context capacity constraints and the prohibitive cost of per task finetuning. In this paper, we propose MedTRACE that decouples lightweight process reward model training from a fully zero-shot inference pipeline, enabling a frozen VLM to generalize to unseen medical imaging tasks without any gradient updates at deployment. Evaluated on cross task generalization spanning volumetric CT and histopathology WSI benchmarks, MedTRACE achieves competitive performance compared to modality specific finetuning approaches and MIL baselines. \endabstract

Abstract:
Large multimodal models (LMMs) often suffer from severe inference inefficiency due to the large number of visual tokens introduced by image encoders. While recent token compression methods, such as pruning and merging, have shown promise in reducing redundancy, their evaluation remains fragmented and inconsistent. In this work, we present UniPruneBench, a unified and extensible benchmark for visual token pruning in multimodal LLMs. UniPruneBench provides standardized protocols across six ability dimensions and ten datasets, covering ten representative compression algorithms and three families of LMMs (LLaVA-v1.5, Intern-VL3, and Qwen2.5-VL). Beyond task accuracy, it incorporates system-level metrics such as runtime and prefilling latency to provide a holistic view. Our experiments uncover several key findings: (1) random pruning is a surprisingly strong baseline, (2) no single method consistently outperforms others across scenarios, (3) pruning sensitivity varies significantly across tasks, with OCR being most vulnerable, and (4) pruning ratio is the dominant factor governing performance degradation. We believe UniPruneBench will serve as a reliable foundation for future research on efficient multimodal modeling.

Authors: Xiaohe Li (Aerospace Information Research Institute, Chinese Academy of Sciences), Jiahao Li (Aerospace Information Research Institute, Chinese Academy of Sciences), Kaixin Zhang (Aerospace Information Research Institute, Chinese Academy of Sciences), Yuqiang Fang (Space Engineering University), Leilei Lin (Capital Normal University), Hong Wang (Aerospace Information Research Institute, Chinese Academy of Sciences), Haohua Wu (Aerospace Information Research Institute, Chinese Academy of Sciences), Zide Fan (Aerospace Information Research Institute, Chinese Academy of Sciences)

Abstract:
While Multimodal Large Language Models (MLLMs) excel in general vision-language tasks, their application to remote sensing change understanding is hindered by a fundamental "temporal blindness". Existing architectures lack intrinsic mechanisms for multi-temporal contrastive reasoning and struggle with precise spatial grounding. To address this, we first introduce Delta-QA, a comprehensive benchmark comprising 180k visual question-answering samples. Delta-QA unifies pixel-level segmentation and visual question answering across bi- and tri-temporal scenarios, structuring change interpretation into four progressive cognitive dimensions. Methodologically, we propose Delta-LLaVA, a novel MLLM framework explicitly tailored for multi-temporal remote sensing interpretation. It overcomes the limitations of naive feature concatenation through three core innovations: a Change-Enhanced Attention module that systematically isolates and amplifies visual differences, a Change-SEG module utilizing Change Prior Embedding to extract differentiable difference features as input for the LLM, and Local Causal Attention to prevent cross-temporal contextual leakage. Extensive experiments demonstrate that Delta-LLaVA decisively outperforms leading generalist MLLMs and specialized segmentation models in complex change deduction and high-precision boundary localization, establishing a unified framework for earth observation intelligence.

Abstract:
Online continual learning (CL) enables models to learn incrementally from non-stationary data streams while preserving previously acquired knowledge. However, current online CL approaches primarily focus on the stability-plasticity tradeoff, overlooking the temporal overhead of model adaptation, which is a critical constraint for time-sensitive online applications. In this work, we identify model initialization as a crucial yet under-explored factor in online CL and propose Terra, a two-stage model merging framework to decouple fast model adaptation from long-term knowledge consolidation. To facilitate fast model adaptation to newly arriving tasks, Terra features a novel model zoo design and employs plasticity-aware model merging to construct task-specific model initialization by aggregating the top-k relevant models within the model zoo. To maintain performance stability within the fixed-capacity model zoo, Terra further introduces a stability-oriented merging strategy that consolidates newly acquired knowledge into existing models via geodesic interpolation. Extensive evaluations on three different datasets across various network architectures demonstrate that Terra significantly outperforms state-of-the-art solutions, improving average accuracy by 10.13% and achieving up to 1.42× faster convergence speedup.

Abstract:
\beginabstract Visual question answering failures in multimodal large language models are commonly attributed to insufficient reasoning capacity and addressed through additional training. We ask a prior question: are these failures perceptual rather than inferential---and if so, can external perception tools recover them without any training? We introduce a \emphtool-assisted oracle ceiling that measures the maximum performance a frozen model can achieve through ideal per-sample tool selection, cleanly separating information-access failure from capacity failure. On Qwen2-VL-7B across five benchmarks, the oracle reaches +18.4\,pp above the frozen baseline---comparable to fine-tuning on 522k samples---while naive indiscriminate tool use captures only +9.2\,pp, establishing routing quality as the decisive bottleneck. A per-sample failure taxonomy reveals that perception tools fix 61.4% of baseline errors without any training, outpacing backbone scaling to Qwen3-VL-8B (50.9%); and that 77.1% of capacity failures are also oracle-fixable---suggesting that gains commonly attributed to improved model capacity are, in part, better explained as improved information access. Building on this analysis, we propose PercepAgent, a training-free agent that selects tools by diagnosing evidence gaps rather than assessing answer confidence. PercepAgent recovers 61% of the oracle headroom at zero training cost, surpasses fine-tuned VPT on out-of-distribution benchmarks, and transfers across backbone generations with zero reconfiguration. \endabstract

Abstract:
Open-vocabulary audio-visual event localization (OV-AVEL) grounds a text-queried event in time from video, audio, and language. The supervision sources available to this task can differ in temporal-boundary reliability: on OV-AVEBench, our configured visual teacher gives more reliable boundary cues than the configured audio teacher, although the latter is a strong pretrained audio model and remains semantically informative. This is a setting-specific diagnostic rather than a universal ranking of vision and audio. We formulate the resulting challenge as supervision placement: which teacher signals may shape the localization decision, and which should remain auxiliary. Based on this view, we propose OV-OrthKD, a reliability-aware asymmetric distillation framework. Visual feature transfer shapes a decision-aligned representation, audio feature transfer enriches a complementary auxiliary subspace, a text prototype anchors seen/unseen category semantics, and an orthogonality loss limits directional overlap between the two teacher-specific projections. The student continues to use both modalities through query-aware fusion at inference, while the default training recipe keeps audio-teacher supervision off the segment-logit path. On OV-AVEBench, OV-OrthKD achieves 0.816 segment AP and improves F1@0.5 over the official fine-tuning baseline by 2.7 points overall and 3.4 points on unseen categories. Path-assignment, role-swap, corruption, and transfer analyses consistently support supervision placement as a task-specific design axis for OV-AVEL.

Authors: Yingjie Zhou (Shanghai Jiao Tong University), Ronghui Li (Tsinghua University), Zhaohe Lv (China University of Mining and Technology), Chuyi Lin (China University of Mining and Technology), Ziying Wang (China University of Mining and Technology), Jiale Tang (China University of Mining and Technology), Farong Wen (Shanghai Jiao Tong University), Zicheng Zhang (Shanghai Jiao Tong University), Yu Zhou (China University of Mining and Technology), Jiezhang Cao (Shanghai Jiao Tong University), Yue Shi (Shanghai Jiao Tong University), Xiaohong Liu (Shanghai Jiao Tong University)

Abstract:
Although recent advances in AI-based Text-to-Motion (T2M) algorithms have enabled text-driven 3D Animated Human (3DAH), often referred to as ``Puppeteers for their ability to streamline 3D animation design, the quality of the generated motion remains a critical concern. More importantly, the T2M field continues to lack reliable and objective evaluation metrics, which substantially hinders the advancement of puppeteers and negatively impacts user experience. To address this gap, we introduce the first large-scale quality assessment dataset for generative 3DAHs, termed AHQA-3D, which comprises 1,200 3DAHs sequences generated by 6 representative puppeteers. Through carefully designed subjective experiments involving recruited participants, we obtain human quality ratings for each 3DAH. The results highlight the impact of different factors on animation quality. We further propose ViSCA, an objective quality assessment agent based on Qwen2.5-VL. ViSCA supports end-to-end quality evaluation across four dimensions: 3D motion fidelity, 2D human body representation, visual quality, and semantic consistency. Experimental results demonstrate that ViSCA outperforms existing evaluation methods and exhibits stronger alignment with human perception.

Abstract:
Time-series forecasting supports decisions in domains such as finance and healthcare, where forecasts may depend on events not fully represented in numerical histories. Multimodal forecasting can supply this context through text, but the additional information is useful only when it remains relevant to the target and forecast horizon. Existing methods commonly fuse text at every prediction step, even when it is stale, weakly related to the target, or informative only for certain horizons. We therefore frame text use as a horizon-wise decision and introduce CUTS, a Counterfactual Utility Routing framework for selective fusion. For each forecast step, CUTS retrieves horizon-specific evidence and produces matched text-conditioned and no-text forecasts. A utility estimator predicts the benefit of the retrieved evidence, while its uncertainty moderates the final interpolation between the two forecasts. On Time-MMD, CUTS achieves the best nine-domain normalized MSE and MAE among the evaluated methods. It records the lowest MSE in five of nine domains. On FNSPID, it achieves the lowest average MSE and MAE and obtains the lowest MSE at the 3-day and 5-day horizons. Ablations identify horizon-specific evidence selection and staged training as key contributors. Corruption and routing analyses further show that uncertainty-aware routing helps CUTS limit its reliance on unreliable text.

Abstract:
Large Multimodal Reasoning Models (LMRMs) extend large language models to heterogeneous modalities, demonstrating strong potential across a wide range of open-ended tasks. Reinforcement Learning with Verifiable Rewards (RLVR) has been widely adopted to improve reasoning performance in LMRMs; however, existing approaches overlook the critical role of visual perception uncertainty in guiding reasoning processes. We term this limitation Reasoning Allocation Misalignment, where the misalignment between visual perception uncertainty and reasoning depth leads to over-allocation of deep reasoning under uncertain visual grounding, resulting in unnecessarily long reasoning trajectories and degraded performance. To address this issue, we propose ThinkWise, a perception-aware RLVR framework that jointly models task difficulty and visual perception state to regulate reasoning depth. Specifically, we introduce an instance-adaptive visual perception modeling mechanism that enables context-aware calibration of perception uncertainty, leading to more accurate perception estimation. We further design a structured reward formulation that jointly considers perception, correctness, and task difficulty. In particular, an answer reward provides supervision for correctness, a perception- and difficulty-conditioned length reward adaptively regulates reasoning depth, and a perception reward encourages trajectories with stronger visual grounding. Extensive experiments on multiple multimodal reasoning benchmarks demonstrate that ThinkWise consistently improves the trade-off between accuracy and reasoning efficiency across both 3B and 7B models. These results highlight the importance of perception-aware reasoning and establish a principled framework for advancing LMRMs.

Abstract:
Recent text-to-video (T2V) generation models enable fake news videos to be synthesized from scratch, shifting the threat beyond cheap fakes assembled from existing footage. Such news videos can closely match fabricated narratives, creating a modality alignment trap for existing detectors. Existing datasets lack pure synthesis fake news videos. Although directly prompting T2V models with descriptions of fake news videos can yield perfectly aligned samples, it reduces the fake news video detection (FNVD) to unimodal shortcuts and causes semantic-visual degeneration. To counter this, we formulate T2V-FNVD as a novel ternary classification task with three labels (real, cheap fake, and pure synthesis fake) and construct the first pure synthesis fake news video dataset (PS-FNVD). PS-FNVD includes fabricated events with aligned deception (Type 1) and true events with false visual provenance (Type 2), preventing models from exploiting unimodal shortcuts. Furthermore, we propose the Reasoning-guided T2V-FNVD (R-T2V) framework. Trained through conditioned rationale generation and supervised fine-tuning, R-T2V integrates high-level semantic logic with low-level physical generative traces to predict the ternary veracity label. Extensive experiments across 10 prevailing baselines show that R-T2V achieves the state-of-the-art performance, outperforming the second-best baseline by 12.20 percentage points in accuracy and 8.46 percentage points in macro F_1.

Abstract:
Large visionlanguage models (VLMs) have become increasingly powerful, but their security risks have risen just as sharplyespecially backdoor attacks, in which a model behaves normally on clean inputs yet produces malicious outputs when a hidden trigger is present. To our knowledge, we are the first to expose how vulnerable visionlanguage models (VLMs) are to soft-prompt backdoors and introduce the Dual-stream Soft Prompt (DSP) Attack. DSP inserts tiny soft prompt channels into both the visual and textual branches, enabling output manipulation while keeping the entire backbone frozen. We further design two dual-stream prompt variants: Orthogonal Prompts (OP), where visual and textual prompts are independent, and Mutual-Guided Prompts (MGP), where the two prompts are directionally aligned. Experiments on several mainstream VLM architectures and public datasets show that DSP requires only a small poisoning budget and minimal additional parameters, yet markedly increases both single-trigger and dual-trigger attack success rates. These findings expose an overlooked threat surface and establish a stronger baseline for future research in defending against backdoor attacks on VLMs.

Abstract:
Current advancements in Multimodal Anomaly Detection (MAD) are largely driven by enhancing multimodal fusion, particularly through the integration of RGB and Depth data for richer anomaly representation. However, less attention was devoted to analyzing the role of cross-modal fusion biasa well-known challenge in multimodal learningin MAD. This gap motivates a key question: can we overcome this bias to break the performance bottleneck of current work? In this paper, we first analyze the impact of cross-modal fusion bias in MAD via the Fisher Information Matrix. Then, grounded in these findings, we propose UCFB, a simple yet effective plug-and-play framework designed to mitigate cross-modal fusion bias in MAD. It achieves this by jointly employing Fisher-information-guided dynamic calibration to adjust modality-specific regularization weights and canonical similarity analysis to improve inter-modal interactions. Extensive experiments on the MVTec 3D-AD and Eyecandies datasets demonstrate that UCFB achieves consistent improvements in single-class, multi-class, and few-shot settings.

Abstract:
In this work, we examine hateful memes from three complementary angles -- how to detect them, how to explain their content and how to intervene them before being posted -- by applying a range of strategies built on top of generative AI models. To the best of our knowledge, explanation and intervention have typically been studied separately from detection, which does not reflect real-world conditions. Further, since curating large annotated datasets for meme moderation is prohibitively expensive, we propose a novel framework -- PEST -- that leverages task-specific generative VLMs and the few-shot adaptability of large VLMs to cater to different types of memes. We believe this is the first work focused on generalizable hateful meme moderation under limited data conditions, and has strong potential for deployment in real-world production scenarios. \textcolorredWarning: Contains potentially toxic contents.

Abstract:
Single image reflection removal (SIRR) is a fundamentally ill-posed problem that requires separating superimposed transmission and reflection layers from a single observation. While recent Diffusion Transformers (DiTs) offer powerful synthesis priors, applying them directly to SIRR exposes two deficiencies: semantic bleeding, where entangled layer signals lead the model to reconstruct reflection artifacts as genuine scene content with photorealistic fidelity; and structural topology collapse, where the absence of explicit 3D awareness causes reflection boundaries to be misinterpreted as real object edges. In this paper, we propose OrthoGeo-DiT, a framework that reconceptualizes reflection removal as a multi-dimensional perceptual decoupling process rather than a 2D pixel-level denoising task. Inspired by how the human visual system exploits both semantic conflict detection and monocular spatial perception, our approach introduces three synergistic components: an Orthogonal Semantic Decoupling module that separates frozen DINOv2 features into transmission and reflection representations via dual-path MLPs under orthogonality regularization; an Implicit Geometric Anchoring mechanism that extracts latent features from an adapted vision foundation model as robust topological priors, circumventing the fragility of explicit depth estimation under reflection interference; and a Zero-Init Cross-Attention Fusion strategy that injects both cues into the DiT backbone while preserving its pre-trained generative capabilities. Extensive experiments across seven real-world benchmarks show that OrthoGeo-DiT achieves state-of-the-art performance in both quantitative metrics and perceptual quality, particularly on scenes with heavy reflections, complex geometry, and adverse lighting.

Abstract:
Monocular depth estimation (MDE) faces challenges with non-Lambertian surfaces and adverse weather conditions due to the visual ambiguities inherent in single-image limited information. Existing works address them in isolation via image inpainting or augmentation, yielding limited robustness gains. Language, as a powerful complementary modality to vision, is demonstrated to enhance the visual perception capabilities of vision-language models (VLMs) via detailed long captions. However, prior language-integrated MDE methods fail to fully harness this potential due to short text input with limited information, coarse global text feature learning, and limited language guidance during depth decoding. To address these limitations, we propose CapDepth, a novel framework for robust MDE that leverages guidance from detailed long captions to alleviate visual ambiguities in both challenging scenarios. First, we design a detailed long caption input template that explicitly conveys rich spatial relationships among multiple atom sentences. Second, a dynamic caption encoder is introduced to extract fine-grained depth-relevant text features via progressive masked attention. Finally, we propose a text-adaptive decoder that guides enhanced depth decoding with text features via stable adaptive layer normalization. Extensive experiments validate the efficacy of CapDepth, which outperforms state-of-the-art methods, achieving depth error reductions of 25.0% on non-Lambertian surfaces and 22.0% under adverse weather conditions.

Abstract:
As the complexity of 3D digital content grows exponentially, understanding human visual attention is critical for optimizing rendering and processing resources. Therefore, reliable 3D mesh saliency ground truth (GT) is essential for human-centric visual modeling in virtual reality (VR). However, existing VR eye-tracking frameworks are fundamentally bottlenecked by their underlying acquisition and generation mechanisms. The reliance on zero-area single ray sampling (SRS) fails to capture contextual features, leading to severe texture aliasing and discontinuous saliency signals. And the conventional application of Euclidean smoothing propagates saliency across disconnected physical gaps, resulting in semantic confusion on complex 3D manifolds. This paper proposes a robust framework to address these limitations. We first introduce a view cone sampling (VCS) strategy, which simulates the human foveal receptive field via Gaussian-distributed ray bundles to improve sampling robustness for complex topologies. Furthermore, a hybrid Manifold-Euclidean constrained diffusion (HCD) algorithm is developed, fusing manifold geodesic constraints with Euclidean scales to ensure topologically-consistent saliency propagation. We demonstrate the improvement in performance over baseline methods and the benefits for downstream tasks through subjective experiments and qualitative and quantitative methods. By mitigating topological short-circuits and aliasing, our framework provides a high-fidelity 3D attention acquisition paradigm that aligns with natural human perception, offering a more accurate and robust baseline for 3D mesh saliency research.

Abstract:
Multimodal emotion recognition has achieved promising progress by jointly modeling heterogeneous signals. However, most existing approaches are trained under a fixed set of modalities, which severely limits their ability to flexibly expand to new modalities without retraining the entire model. To address this challenge, we propose , we propose MEP-ARAS (Modality-Expanded Prompting with Attention Reallocation via Adaptive Shifting), a parameter-efficient framework for modality expansion in multimodal emotion recognition. MEP-ARAS decomposes modality expansion into two complementary components: (i) a Modality-Conditioned Prompted Expansion (MCPE) module that selectively introduces informative tokens from newly added modalities under the guidance of existing modality context, and (ii) an Attention Reallocation via Adaptive Shifting (ARAS) mechanism that explicitly corrects attention bias toward original modalities through lightweight LoRA-based attention shifting and attention-guided rank allocation. Experiments on CMU-MOSEI and CMU-MOSI demonstrate that MEP-ARAS achieves effective modality expansion with only a minimal parameter overhead, consistently outperforming existing methods across various settings. These results suggest that MEP-ARAS provides a practical and scalable solution for extending pre-trained Transformers to open-world multimodal emotion recognition.

Abstract:
Modern deepfakes have evolved into localized and intermittent manipulations that require fine-grained temporal localization to mitigate severe digital security risks. The prohibitive cost of frame-level annotation makes weakly supervised methods a practical necessity, which rely only on video-level labels. To this end, we propose Reconstruction-based Temporal Deepfake Localization (RT-DeepLoc), a weakly supervised temporal forgery localization framework that identifies forgeries via reconstruction errors. Our framework uses a Masked Autoencoder (MAE) trained exclusively on authentic data to learn its intrinsic spatiotemporal patterns; this allows the model to produce significant reconstruction discrepancies for forged segments, effectively providing the missing fine-grained cues for accurate localization without demanding dense human annotations. To robustly leverage these indicators, we introduce a novel Asymmetric Intra-video Contrastive Loss (AICL). By focusing on the compactness of authentic features guided by these reconstruction cues, AICL establishes a stable decision boundary that enhances local discrimination while preserving generalization to unseen forgeries by advanced generative models. Extensive experiments on large-scale datasets, including LAV-DF, demonstrate that RT-DeepLoc achieves state-of-the-art performance in weakly-supervised temporal forgery localization.

Abstract:
High-quality remote sensing (RS) image acquisition is fundamentally constrained by physical limitations. While Multi-Frame Super-Resolution (MFSR) and Pansharpening address this by exploiting complementary information, they are typically studied in isolation: MFSR lacks high-resolution (HR) structural priors for fine-grained texture recovery, whereas Pansharpening relies on upsampled low-resolution (LR) inputs and is sensitive to noise and misalignment. In this paper, we propose SatFusion, a novel and unified framework that seamlessly bridges multi-frame and multi-source RS image fusion. SatFusion extracts HR semantic features by aggregating complementary information from multiple LR multispectral frames via a Multi-Frame Image Fusion (MFIF) module, and integrates fine-grained structural details from an HR panchromatic image through a Multi-Source Image Fusion (MSIF) module with implicit pixel-level alignment. To further alleviate the lack of structural priors during multi-frame fusion, we introduce an advanced variant, SatFusion, which integrates a panchromatic-guided mechanism into the MFIF stage. Through structure-aware feature embedding and transformer-based adaptive aggregation, SatFusion enables spatially adaptive feature selection, strengthening the coupling between multi-frame and multi-source representations. Extensive experiments on four benchmark datasets validate our core insight: synergistically coupling multi-frame and multi-source priors effectively resolves the fragility of existing paradigms, delivering superior reconstruction fidelity, robustness, and generalizability.

Abstract:
The open-sourcing of powerful image generation models has created a vibrant ecosystem where creators curate and combine a vast array of community-contributed models. This practice stands in sharp contrast to using closed-source tools like Midjourney. Yet, little is known about these emerging creative workflows. To bridge this gap, this paper presents the first large-scale empirical study of creator model usage behavior within this open-source image generation ecosystem. We construct a novel dataset of 6 million images with their embedded generation metadata --- a detailed recipe of the creation process, including the models used and the prompts. By linking the usage of 22.4K base models and 154K LoRA models to the images, our findings underscore the ecosystem's unique strengths and its inherent obstacles. This provides valuable insights for making this ecosystem more sustainable and innovative. Moreover, we make our dataset publicly available, providing creators with practical references for producing better artworks and researchers to facilitate further studies.

Abstract:
We present DynamicDriver, a generalized Gaussian splatting framework tailored for reconstructing dynamic driving scenes and synthesizing novel views across multiple lanes. In real-world autonomous driving data collection, videos are typically captured from a single trajectory, limiting scene coverage. To enrich training data with diverse viewpoints and road positions, we virtually shift the vehicle's trajectory to generate novel lane-level views. However, rendering views from adjacent or unseen lanes remains highly challenging due to limited coverage and occlusions in the original observations. Generalized Gaussian splatting presents a promising solution when trained on large-scale driving datasets. Previous works in this field mainly focus on static scenes, while DynamicDriver enables the reconstruction of dynamic scenes by integrating LiDAR and optical flow, which also benefits static scene modeling. To address the limitations of depth estimation in both dynamic and static regions, we propose a Cross-Cue LiDAR Fusion module that injects accurate LiDAR depth into the image feature space. Additionally, we introduce a Flow-Based Gaussian Optimizer that improves the representation of dynamic objects by enhancing temporal consistency, allowing their Gaussians to be more clearly separated from the background. Experiments demonstrate that DynamicDriver achieves state-of-the-art performance in generating high-fidelity and temporally consistent novel views for complex lane-switching scenarios involving both static and dynamic elements.

Abstract:
This paper introduces a spatial-multimedia view of BIM-to-robot execution. Rather than treating BIM as static design data, the system compiles Rhino-authored BIM, QR-linked identifiers, Mixed Reality overlays, and runtime state into an Action Skill Graph (ASG) that functions as executable spatial media. The same compiled graph drives a Unity digital twin and an eight-cable parallel robot through six reusable primitives: approach, grasp, align, insert, release, and verify. Under a scoped zero-code contract, new projects reuse a standardized BIM schema, rule templates, and a robot adapter instead of per-project motion scripts. In a controlled ten-scene feasibility study, the compiled graphs executed 5,016 logged primitive steps and transferred to a small-object hardware trial. Companion uncertainty benchmarks further showed that ASG-governed execution increased mean completion from 56.8% to 83.4% under a high-precision profile and from 31.6% to 73.7% under a lower-precision profile, with only modest runtime overhead. The contribution is therefore a proof-of-concept for multimedia-grounded robotic assembly rather than a full obstacle-aware motion-planning solution.

Abstract:
AI-generated image detection has become a critical challenge as generative models produce increasingly realistic fake content. Existing methods, including CNN-based detectors and VLM-based approaches, typically overfit to training-domain patterns and struggle to generalize to unseen generative models. Recent RAG-augmented detection methods alleviate this by retrieving similar images as references to reduce VLMs' reliance on purely semantic context; however, they treat all retrieved samples uniformly based on semantic similarity, overlooking a key asymmetry: a retrieval gallery contains both real and fake samples, which carry inherently different discriminative signals. We argue that deviating from real image patterns is not equivalent to sharing forgery clues with fake samples, and that treating them uniformly discards the asymmetry that enables reliable detection. To address this, we propose AEC, an asymmetric evidence comparison framework for RAG-based AI-generated image detection, which explicitly converts retrieved real and fake samples into asymmetric discriminative evidence, rather than simply appending them as context. Specifically, AEC retrieves reference samples from complementary semantic, texture, and frequency spaces, and evaluates multi-stream reliability through an entropy-based adaptive weighting mechanism. The Evidence Comparison Module (ECM) then processes them via two asymmetric branches: the authenticity deviation branch computes residuals against real samples to capture discrepancies from real image distributions, while the forgery alignment branch applies the Hadamard product with fake samples to amplify shared forgery clues in specific feature dimensions. AEC organizes these signals into a structured evidence template guiding VLM inference, achieving interpretable detection and strong cross-domain generalization. Experiments on multiple benchmarks demonstrate superior performance against unseen generative models. The code and models are at https://anonymous.4open.science/r/AEC-85D7.

Abstract:
The rapid evolution of Large Multimodal Models (LMMs) has enabled agents to perform complex digital and physical tasks, yet their deployment as autonomous decision-makers introduces substantial unintentional behavioral safety risks. Existing behavioral-safety evaluations remain fragmented across domains and environment types. To support consistent cross-domain analysis, we present BeSafe-Bench (BSB), a benchmark for exposing behavioral safety risks of situated agents in functional environments, covering four representative domains: Web, Mobile, Embodied VLM, and Embodied VLA. Using functional environments, we construct a diverse instruction space by augmenting tasks with nine categories of safety-critical risks, and adopt a hybrid evaluation framework that combines rule-based checks with LLM-as-a-judge reasoning to assess real environmental impacts. Evaluating 13 popular agents reveals a concerning trend: even the best-performing agent completes fewer than 40% of tasks while fully adhering to safety constraints, and strong task performance frequently coincides with severe safety violations. These findings underscore the urgent need for improved safety alignment before deploying agentic systems in real-world settings.

Abstract:
With the widespread adoption of Text-to-Image (T2I) Diffusion Models (DMs) across various applications, their security has attracted increasing attention. In particular, these models are highly vulnerable to backdoor attacks, where adversaries implant hidden triggers during training to manipulate model outputs at inference time. To deal with the backdoor attacks, existing defense methods typically require substantial training data and employ coarse-grained alignment, leading to catastrophic forgetting and semantic degradation, respectively. To defend against backdoor attacks while preventing performance degradation, this paper first analyzes the behavior of poisoned models and finds that i) the neuron activation variation distributions of trigger tokens are significantly different from those of normal tokens, and ii) the embeddings of non-semantic trigger tokens are far from the manifold space constructed by all the normal tokens in the vocabulary. Based on the findings, we propose ManiEdit, a novel model-edit-based backdoor elimination framework for T2I DMs. Specifically, we propose an activation-based module localization methodology to pinpoint the backdoored module by calculating the Neuron Activation Variations (NAV), enabling precise parameter recoupling within restricted areas to safeguard benign parameters from collateral damage. To sever malicious associations and restore the original semantics of triggers, we present a manifold-driven trigger categorization strategy, which detects triggers via NAV within the localized modules, subsequently applying syntactic and geometric constraints to distinguish non-semantic from semantic triggers. By aligning non-semantic triggers with empty prompts and realigning semantic triggers with denoising-optimized semantic anchors, ManiEdit can achieve efficient elimination without expensive retraining. Comprehensive experiments demonstrate the effectiveness of ManiEdit in defending against various state-of-the-art backdoor attack methods.

Abstract:
As a pivotal source of global plant protein and vegetable oil, soybean breeding in agricultural science is currently constrained by the lack of non-destructive, in-situ dynamic monitoring methods for seed development, which limits the efficiency and accuracy of phenotyping analysis. To address this long-standing challenge, we propose an in vivo monitoring scheme using time-series CT scans and construct SoySeedCT, the first temporal CT dataset for seed development with expert-annotated images of pod-enclosed seeds throughout development. Parsing growth patterns from these time-series CT scans requires precise seed segmentation, which is hindered by the indistinct boundaries caused by the tight connection between pod and seed tissues. In this work, we propose a model-agnostic Uncertainty Labeling Strategy (ULS) for 3D segmentation, which marks boundary voxels as "uncertainty" during training and dynamically reassigns them during inference. Benchmarking on SoySeedCT shows that ULS achieves state-of-the-art results, outperforming the baseline models. By integrating computer vision techniques with agricultural phenotyping, our study not only redefines the paradigm of seed phenotyping analysis, but also provides a practical technical path for AI to support agricultural science research.

Abstract:
Multimodal deep learning has become a powerful paradigm for cancer prognosis by integrating heterogeneous biomedical data such as whole-slide histopathology images and molecular profiles. However, existing models typically produce deterministic survival risk scores that conceal the intrinsic ambiguity present in real-world clinical cases. Tumors often exhibit substantial intra-slide heterogeneity, ambiguous morphological patterns, and inconsistent prognostic signals across imaging and molecular modalities, leading to structural uncertainty that conventional models fail to capture. We introduce SURE-MIL (Structural Uncertainty Reasoning in Multiple Instance Learning), a framework for quantifying and visualizing structural uncertainty in multimodal cancer prognosis. Unlike prior approaches that assume deterministic slide-level predictions, SURE-MIL explicitly models three sources of ambiguity: instance selection uncertainty from heterogeneous patch contributions, semantic confounding uncertainty from ambiguous morphology, and modality conflict uncertainty across imaging and molecular data. The framework decomposes these uncertainties and provides interpretable visualizations alongside survival risk predictions. Experiments on multimodal cancer prognosis datasets demonstrate that modeling structural uncertainty improves predictive robustness, generalization, and interpretability. High-uncertainty cases correspond to clinically challenging borderline patients, highlighting the potential of SURE-MIL for reliable, uncertainty-aware decision support in high-stakes clinical settings.

Abstract:
In order to gain fresh insights about the information processing characteristics of different audio classification models, we propose \emphtransferability analysis. Given a minimal, sufficient signal for a classification on a model f, transferability analysis asks whether other models accept this minimal signal as having the same classification as it did on f. We define what it means for a sufficient signal to be transferable and perform a large study over 3 different classification tasks: music genre, emotion recognition and deepfake detection. We find that transferability rates vary depending on the task, with sufficient signals for music genre being transferable \approx26% of the time. The other tasks reveal much higher variance in transferability and reveal that some models, in particular on deepfake detection, have different transferability behavior. We call these models `flat-earther' models. We investigate deepfake audio in more depth, and show that transferability analysis also allows us to discover information theoretic differences between the models which are not captured by the more familiar metrics of accuracy and precision.

Abstract:
High-definition 3D LiDAR maps are important for autonomous driving and smart-city services, which require reliable detection of object-level changes in multi-temporal urban LiDAR to keep digital maps aligned with the physical world. Existing approaches from raster height differencing to depth image and point-cloud networks often remain tile-based and threshold-driven, yielding per-point scores without explicit detection limits or consistent object-level labels. We propose an object-level 3D change-detection pipeline that integrates detection-limitaware registration, geometry-driven object proxies with rule-based semantic and instance segmentation, and displacement cues in height, volume, and surface-normal direction to assign five change labels with confidence. By decoupling registration, geometry, and semantics, the pipeline propagates pose uncertainty into spatially varying detection limits, stabilizes cross-epoch correspondences, and suppresses false changes caused by residual misalignment and density variation. We also present LoDA, a level-of-detection (LoD) aware benchmark for Subiaco district with fused multi-temporal vehicle-LiDAR maps constructed with LiDAR, GNSS, and IMU support, semantic instances, and object-level annotations. On this benchmark, our method achieves 95.0% accuracy, 90.8% macro F1, and 83.0% macro IoU, exceeding the best baseline by 8.7 IoU points and 4.4 F1 points. On the public Urb3DCD-V2 benchmark evaluated under the official point-wise protocol, it reaches 96.81% mean accuracy and 89.52% mean change IoU, improving over the strongest reported baselines by 1.36 points in mAcc and 3.18 points in mIoUch.

Abstract:
Federated Continual Learning (FCL), extending Federated Learning (FL) across dynamically evolving tasks, is becoming increasingly important in multimodal applications. Given the established vulnerability of FL to backdoor attacks, FCL constitutes a natural target for similar threats. However, our experiments show that various backdoor attacks formulated for FL are ineffective when applied to FCL, resulting in an average decline of 89.42% in ASR. This counterintuitive phenomenon raises a critical question: does FCL truly provide inherent resistance to backdoor attacks? To answer this question, we conduct an investigation, which reveals that partial parameter sharing restricts the propagation of backdoors. More importantly, catastrophic forgetting caused by subsequent task updates erases injected malicious features in a manner similar to the degradation of clean task performance. Leveraging these insights, we propose AdvSE, the first backdoor attack specifically designed for FCL. Specifically, we employ a dynamic adversarial mechanism to navigate multiple potential backdoor activation pathways and isolate the optimal one, forging a robust malicious shortcut that is distinctly separated from benign feature mappings to maximize attack efficacy. To shield this optimized pathway from catastrophic forgetting during task transitions, we further introduce a selective backdoor embedding strategy that anchors malicious updates within shared stable parameters, preventing them from being overwritten by subsequent benign updates. Experiments on multimodal datasets demonstrate that AdvSE improves the ASR by approximately 90% compared to baseline methods. Notably, it consistently achieves an attack success rate above 92%, with a peak of 99.57%, even under robust defense mechanisms such as MultiKrum and Flame, without sacrificing the accuracy of the main task.

Abstract:
Multimodal Large Language Models (MLLMs) often adopt a hybrid Mamba-Transformer architecture to balance accuracy and inference efficiency. However, when processing long-video inputs, the interdependence between Mamba and Transformer layers renders traditional optimization techniques ineffective, as they are designed exclusively for single-architecture models. In this paper, we observe that Mamba and Transformer exhibit both consistent and complementary behaviors during inference. Building on these novel observations, we propose DeltaInfer, an attention-switch \Delta-guided collaborative compression framework for efficient inference. DeltaInfer operates in two stages. It first performs frame-level semantic condensation in Mamba layers to prune redundant backgrounds. It then selectively skips token-level Key-Value (KV) generation in Transformer layers for tokens whose semantics have been already encoded in Mamba's hidden states. This design jointly reduces prefill FLOPs and KV cache memory footprints. Both stages measure importance using the built-in parameter \Delta, which is natively available in Mamba layers, and therefore require neither auxiliary model inference nor attention-matrix materialization, unlike conventional attention-guided approaches. The importance threshold as a hyperparameter is adaptively determined in a data-driven manner, to ensure generalization. Extensive experiments on three hybrid MLLMs across five video understanding benchmarks demonstrate that DeltaInfer achieves 1.92× end-to-end acceleration and a 63.5% reduction in peak GPU memory, while reducing model-level average accuracy by less than 0.15 percentage points. DeltaInfer establishes a plug-and-play paradigm for efficient long-video inference in hybrid MLLMs.