Dataset distillation compresses large training sets into compact synthetic datasets while preserving downstream performance. As modern systems increasingly operate on paired vision-language inputs, multimodal distillation must preserve representation quality and cross-modal alignment under tight compute and memory budgets, yet prior methods often require heavy computes and overlook their correlations. To address this, we present Multimodal Distribution Matching (MDM), a geometry-aware framework for efficient and generalizable multimodal distillation. Specifically, MDM integrates complementary components at the data, model, and loss levels. At the data level, it initializes synthetic image-text pairs by sampling from clusters in the joint embedding space. At the model level, it forms a mixed teacher by interpolating independently fine-tuned models in weight space according to their angular deviation from the pretrained anchor. At the loss level, it matches joint distributions on the unit hypersphere using a geometry-aware matching objective that exploits the joint features in the cross-modal alignment and discrepancy directions along with symmetric contrastive learning. Across image-text retrieval benchmarks with cross-architecture evaluation, MDM yields compact synthetic sets that preserve multimodal semantics, substantially reduce distillation cost, and remain robust across architectures.
论文检索
输入标题、作者或关键词，从 3,042 篇学术成果中精准定位
Vision-Language Models (VLMs) have achieved remarkable progress across tasks such as visual question answering and image captioning. Yet, the extent to which these models perform visual reasoning as opposed to relying on linguistic priors remains unclear. To address this, we introduce VisRes Bench, a benchmark designed to study visual reasoning in naturalistic settings without contextual language supervision. Analyzing model behavior across three levels of complexity, we uncover clear limitations in perceptual and relational visual reasoning capacities. VisRes isolates distinct reasoning abilities across its levels. Level 1 probes perceptual completion and global image matching under perturbations such as blur, texture changes, occlusion, and rotation; Level 2 tests rule-based inference over a single attribute (e.g., color, count, orientation); and Level 3 targets compositional reasoning that requires integrating multiple visual attributes. Across more than 19,000 controlled task images, we find that state-of-the-art VLMs perform near random under subtle perceptual perturbations, revealing limited abstraction beyond pattern recognition. We conclude by discussing how VisRes provides a unified framework for advancing abstract visual reasoning in multimodal research (https://visres-bench.github.io).
Network pruning is an effective technique for enabling lightweight Large Vision-Language Models (LVLMs), which primarily incorporates both weights and activations into the importance metric. However, existing efforts typically process calibration data from different modalities in a unified manner, overlooking modality-specific behaviors. This raises a critical challenge: how to address the divergent behaviors of textual and visual tokens for accurate pruning of LVLMs. To this end, we systematically investigate the sensitivity of visual and textual tokens to the pruning operation by decoupling their corresponding weights, revealing that: (i) the textual pathway should be calibrated via text tokens, since it exhibits higher sensitivity than the visual pathway; (ii) the visual pathway exhibits high redundancy, permitting even 50% sparsity. Motivated by these insights, we propose a simple yet effective Asymmetric Text-Visual Weight Pruning method for LVLMs, dubbed ATV-Pruning, which establishes the importance metric for accurate weight pruning by selecting the informative tokens from both textual and visual pathways. Specifically, ATV-Pruning integrates two primary innovations: first, a calibration pool is adaptively constructed by drawing on all textual tokens and a subset of visual tokens; second, we devise a layer-adaptive selection strategy to yield important visual tokens. Finally, extensive experiments across standard multimodal benchmarks verify the superiority of our ATV-Pruning over state-of-the-art methods.
ReaGEN: Adaptive Generation of Structured Chains-of-Thought for Efficient Multimodal Reasoning
PDF ↗Large Vision Language Models (LVLMs) exhibit strong perceptual and linguistic capabilities yet struggle with complex visual reasoning tasks that require structured, compositional, and adaptive inference. Existing approaches either rely on costly inference-time exploration--such as multi-path or tree-based Chain-of-Thought (CoT) search--or on expensive post-training with large curated CoT datasets. We propose ReaGEN, a lightweight framework for the adaptive generation of structured reasoning chains that enhances reasoning without modifying the underlying vision-language model (VLM). ReaGEN first employs a teacher-guided evolutionary search to collect sample specific CoT structure, leveraging attention-derived stage importance to capture how information flows across reasoning stages. These adaptive CoT structures are then used to train a compact generator (GEN) that learns to refine and improve CoT structures by reflecting on attention feedback from the reasoning process. At inference, the GEN dynamically produces question-adaptive structured CoTs, and can be iteratively invoked to refine them based on the VLM's internal state--achieving the flexibility of deep search with single-path efficiency. Across diverse multimodal reasoning benchmarks, ReaGEN achieves up to +26 accuracy points over test-time scaling methods while reducing the average inference-time token usage by 79%, making it a scalable approach for structured reasoning generation in self-hosted or open-source VLMs with access to internal attention signals.
Edge-based representations are fundamental cues for visual understanding, a principle rooted in early vision research and still central today. We extend this principle to vision-language alignment, showing that isolating and aligning structural cues across modalities can greatly benefit fine-tuning on long, detail-rich captions, with a specific focus on improving cross modal retrieval. We introduce StructXLIP, a fine-tuning alignment paradigm that extracts edge maps (e.g., Canny), treating them as proxies for the visual structure of an image, and filters the corresponding captions to emphasize structural cues, making them "structure-centric". Fine-tuning augments the standard alignment loss with three structure-centric losses: (i) aligning edge maps with structural text, (ii) matching local edge regions to textual chunks, and (iii) connecting edge maps to color images to prevent representation drift. From a theoretical standpoint, while standard CLIP maximizes the mutual information between visual and textual embeddings, StructXLIP additionally maximizes the mutual information between multimodal structural representations. This auxiliary optimization is intrinsically harder, guiding the model toward more robust and semantically stable minima, enhancing vision-language alignment. Beyond outperforming current competitors on cross-modal retrieval on both general and specialized domains, our method serves as a general boosting recipe that can be integrated into future approaches in a plug-and-play manner. Code and pretrained models are available at: https://github.com/intelligolabs/StructXLIP.
CausalLens: Sensitivity-Guided Multi-Head Causal Intervention for Hallucination Mitigation in Large Vision-Language Models
PDF ↗Recent Large Vision-Language Models (LVLMs) have shown impressive capabilities in multimodal understanding and generation. Despite this progress, they remain prone to *hallucination*, where model outputs conflict with the visual input due to an over-reliance on textual priors. Existing inference-time mitigation approaches frequently depend on multi-pass or contrastive decoding, which increases latency and limits their applicability in real-time settings. To address this limitation, we propose **CausalLens**, a training-free and single-pass intervention that directly adjusts the decoder hidden states to strengthen visual grounding. By decomposing attention heads into visual, textual, and system prompt pathways, CausalLens identifies visually reliable heads using a sensitivity measure and selectively adjusts their mid-layer hidden-state contributions. A projection-aligned correction further stabilizes these adjusted states after multi-head fusion, ensuring that the enhanced visual information is preserved throughout decoding. Extensive experiments across multiple hallucination benchmarks and LVLM architectures demonstrate that CausalLens consistently improves visual fidelity while adding negligible computational overhead. The method requires no fine-tuning or architectural changes, making it well-suited for practical, latency-sensitive applications.
Self-supervised Vision Transformers (ViTs) like DINO show an emergent ability to discover objects, typically observed in \texttt [CLS] token attention maps of the final layer. However, these maps often contain spurious activations resulting in poor localization of objects. This is because the \texttt [CLS] token, trained on an image-level objective, summarizes the entire image instead of focusing on objects. This aggregation dilutes the object-centric information existing in the local, patch-level interactions. We analyze this by computing inter-patch similarity using patch-level attention components (query, key, and value) across all layers. We find that: (1) Object-centric properties are encoded in the similarity maps derived from all three components (q, k, v), unlike prior work that uses only key features or the \texttt [CLS] token. (2) This object-centric information is distributed across the network, not just confined to the final layer. Based on these insights, we introduce Object-DINO, a training-free method that extracts this distributed object-centric information. Object-DINO clusters attention heads across all layers based on the similarities of their patches and automatically identifies the object-centric cluster corresponding to all objects. We demonstrate Object-DINO's effectiveness on two applications: enhancing unsupervised object discovery (+3.6 to +12.4 CorLoc gains) and mitigating object hallucination in Multimodal Large Language Models by providing visual grounding. Our results demonstrate that using this distributed object-centric information improves downstream tasks without additional training.
While Large Multimodal Models (LMMs) have made significant progress, they remain largely text-centric, relying on language as their core reasoning modality. As a result, they are limited in their ability to handle reasoning tasks that are predominantly visual. Recent approaches have sought to address this by supervising intermediate visual steps with helper images, depth maps, or image crops. However, these strategies impose restrictive priors on what "useful" visual abstractions look like, add heavy annotation costs, and struggle to generalize across tasks. To address this critical limitation, we propose a task-agnostic mechanism that trains LMMs to discover and use visual reasoning tokens without explicit supervision. These tokens attend globally and re-encode the image in a task-adaptive way, enabling the model to extract relevant visual information without hand-crafted supervision. Our approach outperforms direct fine-tuning and achieves state-of-the-art results on a diverse range of vision-centric tasks -- including those where intermediate abstractions are hard to specify -- while also generalizing to multi-task instruction tuning.
While a general embodied agent must function as a unified system, current methods are built on isolated models for understanding, world modeling, and control. This fragmentation prevents unifying multimodal generative capabilities and hinders learning from large-scale, heterogeneous data. In this paper, we propose Motus, a unified latent action world model that leverages existing general pretrained models and rich, sharable motion information. Motus introduces a Mixture-of-Transformer (MoT) architecture to integrate three experts (ie, understanding, video generation, and action) and adopts a UniDiffuser-style scheduler to enable flexible switching between different modeling modes (ie, world models, vision-language-action models, inverse dynamics models, video generation models, and video-action joint prediction). Motus further leverages the optical flow to learn latent actions and adopts a recipe with three-phase training pipeline and six-layer data pyramid, thereby extracting pixel-level "delta action" and enabling large-scale action pretraining. Experiments show that Motus achieves superior performance against state-of-the-art methods in both simulation (a +15% improvement over X-VLA and a +45% improvement over Pi-0.5) and real-world scenarios(improved by +11 48%), demonstrating unified modeling of all functionalities and priors significantly benefits downstream robotic tasks.
RoboWheel: A Data Engine from Real-World Human Demonstrations for Cross-Embodiment Robotic Learning
PDF ↗We introduce Robowheel, a data engine that converts human hand-object interaction (HOI) videos into training-ready supervision for cross-morphology robotic learning. From monocular RGB/RGB-D inputs, we perform high-precision HOI reconstruction and enforce physical plausibility via a reinforcement learning (RL) optimizer that refines hand-object relative poses under contact and penetration constraints. The reconstructed, contact-rich trajectories are then retargeted to cross-embodiments, robot arms with simple end-effectors, dexterous hands, and humanoids, yielding executable actions and rollouts. To scale coverage, we build a simulation-augmented framework on Isaac Sim with diverse domain randomization (embodiments, trajectories, object retrieval, background textures, hand motion mirroring), which enriches the distributions of trajectories and observations while preserving spatial relationships and physical plausibility. The entire data pipeline forms an end-to-end pipeline from video - reconstruction - retargeting - augmentation - data acquisition.We validate the data on mainstream vision-language-action (VLA) and imitation learning architectures, demonstrating that trajectories produced by our pipeline are as stable as those from teleoperation and yield comparable continual performance gains. To our knowledge, this provides the first quantitative evidence that HOI modalities can serve as effective supervision for robotic learning. Compared with teleoperation, Robowheel is lightweight: a single monocular RGB(D) camera is sufficient to extract a universal, embodiment-agnostic motion representation that could be flexibly retargeted across embodiments. We further assemble a large-scale multimodal dataset combining multi-camera captures, monocular videos, and public HOI corpora for training and evaluating embodied models.
Vision-language models (VLMs) have achieved strong performance in visual question answering (VQA), yet they remain constrained by static training data. Retrieval-Augmented Generation (RAG) mitigates this limitation by enabling access to up-to-date, culturally grounded, and multilingual information; however, multilingual multimodal RAG remains largely underexplored. We introduce M4-RAG, a massive-scale benchmark spanning 42 languages, 56 regional dialects and registers, and 189 countries, comprising over 80,000 culturally diverse image-question pairs for evaluating retrieval-augmented VQA across languages and modalities. To balance realism with reproducibility, we build a controlled retrieval environment containing millions of carefully curated multilingual documents relevant to the query domains, approximating real-world retrieval conditions while ensuring consistent experimentation. Our systematic evaluation reveals that although RAG consistently benefits smaller VLMs, it fails to scale to larger models and often even degrades their performance, exposing a critical mismatch between model size and current retrieval effectiveness. Our cross-lingual evaluations also reveal significant performance degradation when prompts or retrieved context are provided in non-English languages. The code, datasets, and evaluation protocols for M4-RAG are available as open-source at https://github.com/davidanugraha/M4-RAG.
Achieving human-level performance in Vision-and-Language Navigation (VLN) requires an embodied agent to understand textual instructions, perceive visual observations, and reason over long action sequences. Recent works, such as NavCoT and NavGPT-2, demonstrate the potential of Chain-of-Thought (CoT) reasoning for improving interpretability and long-horizon planning. Moreover, multimodal extensions like OctoNav-R1 and CoT-VLA further validate CoT as a promising pathway toward human-like navigation reasoning. However, existing approaches face critical drawbacks: purely textual CoTs lack visual perception and easily overfit to sparse annotated reasoning steps, while multimodal CoTs incur severe token inflation by generating imagined visual observations, making real-time navigation impractical. In this work, we propose FantasyVLN, a unified implicit reasoning framework that preserves the benefits of CoT reasoning without explicit token overhead. Specifically, imagined visual tokens are encoded into a compact latent space using a pretrained Visual AutoRegressor (VAR) during CoT reasoning training, and the model jointly learns from textual, visual, and multimodal CoT modes under a unified multi-CoT strategy. At inference, our model performs direct instruction-to-action mapping while still enjoying reasoning-aware representations. Extensive experiments on LH-VLN show that our approach achieves reasoning-aware yet real-time navigation, improving success rates and efficiency while reducing inference latency by an order of magnitude compared to explicit CoT methods. Code is available at https://github.com/Fantasy-AMAP/fantasy-vln.
Most vision-language models assume images have a single literal meaning, even though images are inherently polysemous. We propose a retrieval paradigm that models many-to-many relationships between images and text using interpretive lenses and introduce Lenses, a multi-prompt embedding model and dataset for polysemous image-text retrieval. The Lenses dataset contains 105,669 images and 732,405 captions, with each image paired with multiple captions and image-side prompts annotated across five categories: Literal, Figurative, Abstract, Background, and Emotional. Building on a multimodal large language model, the Lenses model uses learned lens tokens to extract lens-specific embeddings for every image and caption and compares these using a lens-masking similarity function with a global fallback that prioritizes same-lens matches while retaining a global pathway. Training uses a category-aware multi-positive contrastive loss and intra-set diversity regularization to align corresponding perspectives while preventing semantic collapse across lenses. We further propose lens-aware evaluation protocols, including category-aware ranking, that better reflect how humans match images and text. Experiments on the Lenses dataset and public benchmarks show that our model outperforms baselines on literal and non-literal retrieval and reduces over-reliance on literal cues.
Human mesh reconstruction (HMR) provides direct insights into body-environment interaction, enabling various immersive applications. However, existing large-scale HMR benchmarks largely rely on line-of-sight RGB sensing, causing HMR systems to inherit the limitations of vision-based systems, including sensitivity to occlusion, lighting variation, and privacy concerns. These limitations have motivated growing interest in radio-frequency (RF) mmWave radar as a privacy-preserving and robust modality for human sensing. Despite this promise, current radar datasets remain limited by sparse skeleton annotations, small scale, and simple in-place actions. To address this gap, we introduce M4Human, the largest-scale multimodal benchmark to date for radar-based HMR, featuring 661K frames---9 times larger than the previous largest---with high-resolution mmWave radar, RGB, and depth data. M4Human provides both raw radar tensors (RT) and processed radar point clouds (RPC), enabling research across different levels of RF signal granularity. It also includes high-quality motion capture (MoCap) annotations with 3D meshes and global trajectories, covering 20 subjects and 50 diverse actions, including in-place, seated, and free-space sports or rehabilitation movements. We establish benchmarks on RT and RPC modalities, as well as multimodal fusion with RGB-D inputs. Extensive results demonstrate the value of M4Human for radar-based human modeling while revealing persistent challenges under fast and unconstrained motion. Our benchmark is publicly available at \href https://fanjunqiao.github.io/M4Human-site/ https://fanjunqiao.github.io/M4Human-site/ .
Envision, Attend, Then Respond: Counterfactual Hallucination Mitigation in Large Vision-Language Models
PDF ↗Large Vision-Language Models (LVLMs) often hallucinate when visual evidence conflicts with world knowledge, i.e., in counterfactual scenarios. We propose Envision-Attend-Respond (EnAR), a training-free framework that leverages visual priors to steer the model's attention toward counterfactual elements in the image. The Envision stage constructs a visual impression by invoking a diffusion prior to perform latent perturbations, yielding a prior-consistent counterpart of the input image. The Attend stage processes the original image and its visual impression through the LVLM's vision encoder to localize counterfactual elements, forming a corresponding padded input. The Respond stage performs contrastive decoding between the original and padded inputs to suppress bias and enhance visual understanding. Empirically, EnAR consistently mitigates hallucinations and improves response fidelity, achieving a 10.82% gain on VLMBias and an average 6.9% improvement on POPE, demonstrating robustness across both counterfactual and general hallucination settings. Moreover, the framework remains effective across heterogeneous LVLM architectures, offering a new perspective for hallucination governance in multimodal reasoning.
Multimodal Chain-of-Thought (MCoT) models have demonstrated impressive capability in complex visual reasoning tasks. Unfortunately, recent studies reveal that they suffer from severe hallucination problems due to diminished visual attention during the generation process.However, visual attention decay is a well-studied problem in Large Vision-Language Models (LVLMs). Considering the fundamental differences in reasoning processes between MCoT models and traditional LVLMs, we raise a basic question: Whether MCoT models have unique causes of hallucinations? To answer this question, we systematically investigate the hallucination patterns of MCoT models and find that fabricated texts are primarily generated in associative reasoning steps, which we term divergent thinking. Leveraging these insights, we introduce a simple yet effective strategy that can effectively localize divergent thinking steps and intervene in the decoding process to mitigate hallucinations. Extensive experiments show that our method outperforms existing methods by a large margin. More importantly, our proposed method can be conveniently integrated with other hallucination mitigation methods and further boost their performance. The code is publicly available at https://github.com/ASGO-MM/MCoT-hallucination.
Vision-language pre-training (VLP) has achieved remarkable performance across diverse multimodal learning (MML) tasks. Recently, many efforts have focused on reconstructing missing modalities to improve the adaptability of VLP models in incomplete MML scenarios. However, these approaches overlook the learning imbalance under severe missing-modality conditions, i.e., the optimization process is dominated by reconstructed samples, thereby weakening complete-sample representations. In this paper, we propose a novel ANchor-guided Gradient Alignment (ANGA) framework to address this issue. Specifically, we first retrieve similar instances to reconstruct the missing modalities, thereby alleviating information deficiency. We then introduce an entropy-driven curriculum that progressively incorporates reliable reconstructed samples together with complete ones to form an optimization anchor, which guides gradient alignment to mitigate learning imbalance. Furthermore, we design a semantic-enhanced adapter that leverages the retrieved instances to generate dynamic prompts, further enhancing the robustness of the VLP model. Extensive experiments on widely used datasets demonstrate the superiority of ANGA over state-of-the-art (SOTA) baselines across various missing-modality scenarios. The code is available at this repository.
Large multimodal 3D vision-language models show strong generalization across diverse 3D tasks, but their performance still degrades under domain shifts. This has motivated recent studies on test-time adaptation (TTA), which enables models to adapt online using test-time data. Among existing TTA methods, cache-based mechanisms are widely adopted to leverage previously observed samples for online prediction refinement. However, they store limited historical information, leading to progressive information loss as the test stream evolves. In addition, their prediction logits are fused heuristically, making adaptation unstable. To address these limitations, we propose BayesMM, a multimodal Bayesian distribution learning framework for test-time point cloud analysis. BayesMM models textual priors and streaming visual features of each class as Gaussian distributions. Textual parameters are derived from semantic prompts, while visual parameters are updated online with arriving samples. The two modalities are fused via Bayesian model averaging, which automatically adjusts their contributions based on posterior evidence. This yields a unified prediction that adapts to evolving test-time data without training. Extensive experiments on multiple point cloud benchmarks demonstrate that BayesMM maintains robustness under distributional shifts, yielding over 4 percent average improvement.
Understanding animal species from multimodal data poses an emerging challenge at the intersection of computer vision and ecology.While recent biological models, such as BioCLIP, have demonstrated strong alignment between images and textual taxonomic information for species identification, the integration of the audio modality remains an open problem. We propose BioVITA, a novel visual-textual-acoustic alignment framework for biological applications. BioVITA involves (i) a training dataset, (ii) a representation model, and (iii) a retrieval benchmark. First, we construct a large-scale training dataset comprising 1.3 million audio clips and 2.3 million images, covering 14,133 species annotated with 34 ecological trait labels. Second, building upon BioCLIP2, we introduce a two-stage training framework to effectively align audio representations with visual and textual representations. Third, we develop a cross-modal retrieval benchmark that covers all possible directional retrieval across the three modalities (i.e., image-to-audio, audio-to-text, text-to-image, and their reverse directions), with three taxonomic levels: Family, Genus, and Species. Extensive experiments demonstrate that our model learns a unified representation space that captures species-level semantics beyond taxonomy, advancing multimodal biodiversity understanding. The project page is available at: https://dahlian00.github.io/BioVITA_Page/
While recent Large Vision-Language Models (LVLMs) exhibit impressive multimodal reasoning abilities, they often produce ungrounded, hallucinated responses by over-relying on linguistic priors rather than visual evidence. This critical limitation arises from the lack of a quantitative measure of how much these models actually rely on visual inputs during reasoning. We propose Draft and Refine (DnR), an agent framework driven by a novel question-conditioned utilization metric. This metric quantifies the model's actual reliance on visual evidence by first constructing a query-conditioned relevance map to localize question-specific evidence, and then assessing dependence through relevance-based probabilistic masking. Guided by this metric, the DnR agent refines its initial "draft" through targeted feedback from external visual experts. Each expert's output (e.g., boxes, masks) is rendered as visual cues on the image, and the LVLM is re-queried to select the response that yields the greatest improvement in utilization. This process strengthens visual grounding of predictions without retraining or architectural changes. Experiments across a broad range of VQA and captioning benchmarks demonstrate consistent accuracy gains and reduced hallucination. These results show that quantifying visual utilization provides a principled path for designing more interpretable and evidence-driven multimodal agent systems that effectively leverage visual experts.