Large Vision-Language Models (LVLMs) have advanced multimodal learning but face high computational cost issues due to the input of large number of visual tokens, motivating token pruning to improve inference efficiency.The key challenge lies in identifying which tokens are truly important.Most existing approaches rely on attention- or similarity-based criteria to estimate token importance.However, they inherently suffer from certain limitations, such as being task-agnostic and exhibiting positional bias.In this work, we explore a new perspective on token importance assignment based on token transitions in LVLMs, where token transitions are defined as the changes in token representations occurring as they propagate through the model's modules.We observe that the transition of token representations provides a meaningful signal of semantic information.Based on this insight, we propose TransPrune, a training-free and efficient token pruning method.Specifically, TransPrune progressively prunes tokens by assessing their importance through a combination of Token Transition Variation (TTV), which measures changes in both the magnitude and direction of token representations; as well as Instruction-Guided Attention (IGA), which measures how strongly the instruction attends to visual tokens via attention.Extensive experiments on various LVLM architectures, such as LLaVA-v1.5, LLaVA-Next and Qwen2.5-VL, demonstrate that TransPrune maintains comparable multimodal performance while reducing inference TFLOPs by more than half. The code is available at https://github.com/liaolea/TransPrune.
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
Understanding high-resolution (HR) images remains a critical challenge for multimodal large language models (MLLMs). Recent approaches leverage vision-based retrieval-augmented generation (RAG) to retrieve query-relevant crops from HR images, improving understanding capacity of MLLMs. However, this paradigm often leads to object fragmentation, resulting in semantic bias and incomplete retrieval, while also introducing false positives from irrelevant background patches. To address these issues, we propose Multi-resolution Retrieval-Detection (MRD), a training-free framework that enhances HR image understanding from both local and global perspectives. Locally, MRD enforces cross-scale semantic consistency via multi-resolution semantic fusion to mitigate single-resolution bias and alleviate object fragmentation. Globally, it integrates open-vocabulary object detection (OVD) as localization priors within a unified framework. Extensive experiments across multiple MLLMs on HR image benchmarks demonstrate that MRD achieves state-of-the-art (SOTA) performance on both single-object and multi-object understanding tasks. Code will be available at: https://github.com/yf0412/MRD.
Establishing dense correspondences between shapes is a crucial task in computer vision and graphics, while prior approaches depend on near-isometric assumptions and homogeneous subject types (i.e., only operate for human shapes). However, building semantic correspondences for cross-category objects remains challenging and has received relatively little attention. To achieve this, we propose UniMatch, a semantic-aware, coarse-to-fine framework for constructing dense semantic correspondences between strongly non-isometric shapes without restricting object categories. The key insight is to lift "coarse" semantic cues into "fine" correspondence, which is achieved through two stages. In the "coarse" stage, we perform class-agnostic 3D segmentation to obtain non-overlapping semantic parts and prompt multimodal large language models (MLLMs) to identify part names. Then, we employ pretrained vision language models (VLMs) to extract text embeddings, enabling the construction of matched semantic parts. In the "fine" stage, we leverage these coarse correspondences to guide the learning of dense correspondences through a dedicated rank-based contrastive scheme. Thanks to class-agnostic segmentation, language guiding, and rank-based contrastive learning, our method is versatile for universal object categories and requires no predefined part proposals, enabling universal matching for inter-class and non-isometric shapes. Extensive experiments demonstrate UniMatch consistently outperforms competing methods in various challenging scenarios.
Large vision-language models (LVLMs) have demonstrated remarkable capabilities in multimodal understanding tasks. However, the increasing demand for high-resolution image and long-video understanding results in substantial token counts, consequently leading to reduced inference efficiency. Token compression offers a direct solution by reducing the number of tokens to be processed, thereby improving computational efficiency without architectural changes. Through extensive analysis, we identify two critical limitations in existing inner-LLM token compression methods: positional bias and incompatibility with efficient operators, which critically hinder their practical deployment for LVLM acceleration. This paper presents the first approach from a dynamic token variation perspective, revealing that visual token variations within LLMs exhibit task-agnostic properties. We propose Variation-aware Vision Token Dropping (i.e., V2Drop), which progressively removes visual tokens with minimal variation during LVLM inference, thereby enhancing computational efficiency. Extensive experiments across multiple models and benchmarks consistently demonstrate that V2Drop maintains 94.0% and 98.6% of the original performance for image and video understanding tasks respectively, while reducing LLM generation latency by 31.5% and 74.2%.
Large Vision-Language Models (LVLMs) are foundational to modern multimodal applications, yet their susceptibility to adversarial attacks remains a critical concern. Prior white-box attacks rarely generalize across tasks, and black-box methods depend on expensive transfer, which limits efficiency. The vision encoder, standardized and often shared across LVLMs, provides a stable gray-box pivot with strong cross-model transfer. Building on this premise, we introduce PA-Attack (Prototype-Anchored Attentive Attack). PA-Attack begins with a prototype-anchored guidance that provides a stable attack direction towards a general and dissimilar prototype, tackling the attribute-restricted issue and limited task generalization of vanilla attacks. Building on this, we propose a two-stage attention enhancement mechanism: (i) leverage token-level attention scores to concentrate perturbations on critical visual tokens, and (ii) adaptively recalibrate attention weights to track the evolving attention during the adversarial process. Extensive experiments across diverse downstream tasks and LVLM architectures show that PA-Attack achieves an average 75.1% score reduction rate (SRR), demonstrating strong attack effectiveness, efficiency, and task generalization in LVLMs.
Vision-based end-to-end (E2E) driving has garnered interest in the research community due to its scalability and synergy with multimodal large language models (MLLMs). However, current E2E driving benchmarks primarily feature nominal scenarios paired with existing open-loop evaluation metrics that fall short in capturing the multimodal nature of driving or effectively evaluating performance in long-tail scenarios. To address these gaps, we introduce the Waymo Open Dataset for End-to-End Driving (WOD-E2E). WOD-E2E contains 4,021 driving segments (approximately 12 hours), specifically curated for challenging long-tail scenarios that are rare in daily life with an occurring frequency of less than 0.03%. Each segment in WOD-E2E includes the high-level routing information, ego states, and 360-degree camera views from 8 surrounding cameras. To evaluate E2E driving performance on these long-tail situations, we propose a novel open-loop evaluation metric: Rater Feedback Score (RFS). Unlike conventional distance-based metrics, RFS measures how closely a predicted trajectory matches rater-annotated trajectory preference labels. WOD-E2E includes rater preference labels for validation, and a separate held out test set is used for the 2025 WOD-E2E Challenge benchmark leaderboard.
Medical vision-language models can automate the generation of radiology reports but struggle with accurate visual grounding and factual consistency. Existing models often misalign textual findings with visual evidence, leading to unreliable or weakly grounded predictions. We present "CURE", an error-aware curriculum learning framework that improves grounding and report quality without any additional data. CURE tunes a multimodal instructional model on phrase grounding, grounded report generation, and anatomy-grounded report generation using public datasets. The method dynamically adjusts sampling based on model performance emphasizing on harder samples to improve spatial and textual alignment. CURE improves grounding accuracy by +0.35 IoU, boosts report quality by +0.192 CXRFEScore, and reduces hallucinations by 18.6%. CURE is a data-efficient framework that enhances both grounding accuracy and report reliability.
HUMAPS-4D: A Multimodal Dataset for HUman Motion Analysis with Physiological and Semantic informations
PDF ↗Current advancements in human motion understanding are strongly reliant on video data. Nevertheless, privacy regulations and operational constraints increasingly restrict the use of visual data in real-world scenarios. Inferring posture through wearable sensors, such as instrumented insoles measuring plantar activation, presents itself as a promising alternative. However, the absence of large-scale multimodal datasets hinders the rigorous benchmarking of these methodologies. We introduce HUMAPS-4D, a novel multimodal dataset designed for human motion analysis, effectively bridging computer vision and biomechanics. This dataset integrates synchronized motion capture, multi-view video, IMUs, plantar pressure signals, sEMG activation patterns, and high-level semantic annotations. The data was collected from 32 subjects performing 30 actions over a total duration of 14 hours. Participants demonstrate substantial anthropometric variability (age, body proportions, and morphology), which supports robust generalization across diverse body types. Distinct from existing resources, this collection offers a unique pairing of low-level physiological signals and high-level human motor descriptors. This capability enables the development of generative and inference models conditioned by both physical and semantic constraints, while simultaneously reducing the reliance on personally identifiable visual data. We establish benchmark tasks specifically targeting posture reconstruction from plantar pressure, semantic motion segmentation, physics-informed motricity analysis, and multimodal fusion under privacy-preserving conditions. The dataset, along with its associated annotation tools and visualization utilities, is scheduled for online release soon.
Multimodal Large Language Models (MLLMs) have demonstrated promising advancements in augmenting the capabilities of LLMs to comprehend visual input. However, modality misalignment between vision and text remains a key challenge in MLLM, which can be attributed to two aspects: misalignment of modality-specific representations and depletion of modality-specific details. To address the issue of modality misalignment, we propose DeepAlign, a novel multimodal alignment framework to mitigate modality conflict, which employs representation intervention and structure-induced knowledge distillation to prevent the misalignment and depletion of modality-specific information. Extensive experiments demonstrate that DeepAlign significantly mitigates modality conflicts, leading to substantial performance improvements compared to backbone models across multiple vision-language tasks. It also stimulates some emergent abilities in MLLMs, such as multimodal in-context learning on interleaved text-image sequences.
Despite remarkable advancements in multimodal large language models (MLLMs), their fine-grained visual understanding is constrained by a primary reliance on sparse textual supervision. Existing efforts to introduce visual supervision typically do so during post-training, when visual representations have already been largely fixed, causing such signals to act mainly as auxiliary constraints rather than as a primary force for shaping perceptual features. In this paper, we aim to fundamentally reshape the model's perceptual backbone by incorporating vision supervision directly into the pre-training stage. We observe that pixel-level image patches and textual tokens naturally coexist in a shared, raw high-dimensional space characterized by an inherent input symmetry. Leveraging this insight, we propose UVU, a novel vision-language unified autoregressive framework that eschews vector quantization. It uniquely employs continuous visual encoding for lossless representation of visual inputs and proposes a large-scale iterative hierarchical clustering algorithm to construct a pixel-level visual codebook, thereby extending the vocabulary for unified supervision and enabling autoregressive generation of pixel-level image tokens alongside textual tokens. UVU effectively synergizes pixel-level visual perception with semantic-level visual understanding, internalizing visual reconstruction capabilities and unlocking the facilitative role of visual supervision in enhancing understanding in the pre-training stage. Extensive experiments across multiple tasks demonstrate that MLLMs are capable of achieving superior multimodal understanding performance under the supervised learning paradigm of UVU.
UARE: A Unified Vision-Language Model for Image Quality Assessment, Restoration, and Enhancement
PDF ↗Image quality assessment (IQA) and image restoration are fundamental problems in low-level vision. Although IQA and restoration are closely connected conceptually, most existing work treats them in isolation. Recent advances in unified multimodal understanding-generation models demonstrate promising results and indicate that stronger understanding can improve generative performance. This motivates a single model that unifies IQA and restoration and explicitly studies how IQA can guide restoration, a setting that remains largely underexplored yet highly valuable. In this paper, we propose UARE, to our knowledge the first Unified vision-language model for image quality Assessment, Restoration, and Enhancement. Built on pretrained unified understanding and generation models, we introduce a two-stage training framework. First, a progressive, easy-to-hard schedule expands from single-type distortions to higher-order mixed degradations, enabling UARE to handle multiple degradations. Second, we perform unified fine-tuning of quality understanding and restoration with interleaved text-image data, aligning IQA signals with restoration objectives. Through multi-task co-training, UARE leverages IQA to boost restoration and enhancement performance. Extensive experiments across IQA, restoration, and enhancement tasks demonstrate the effectiveness of UARE. The code and models will be made available.
On Token's Dilemma: Dynamic MoE with Drift-Aware Token Assignment for Continual Learning of Large Vision Language Models
PDF ↗Multimodal Continual Instruction Tuning aims to continually enhance Large Vision Language Models (LVLMs) by learning from new data without forgetting previously acquired knowledge. Mixture of Experts (MoE) architectures naturally facilitate this by incrementally adding new experts and expanding routers while keeping the existing ones frozen. However, despite expert isolation, MoE-based continual learners still suffer from forgetting due to routing-drift: old-task tokens become mistakenly attracted to newly added experts, degrading performance on prior tasks. We analyze the failure mode at the token level and reveal the token's dilemma: ambiguous and old tokens in new-task data offer minimal learning benefit yet induce forgetting when routed to new experts, due to their ambiguous routing assignment during training. Motivated by this, we propose LLaVA-DyMoE, a dynamic MoE framework that incrementally expands the MoE with drift-aware token assignment. We characterize token types via their routing score distributions and apply targeted regularization. Specifically, a token-level assignment guidance steers ambiguous and old tokens away from new experts to preserve established routing patterns and alleviate routing-drift, while complementary routing score regularizations enforce expert-group separation and promote new-expert specialization. Extensive experiments demonstrate that our LLaVA-DyMoE effectively mitigates routing-drift-induced forgetting, achieving over a 7% gain in mean final accuracy and a 12% reduction in forgetting compared to baselines. The project page is zhaoc5.github.io/DyMoE.
Image memorability, i.e., how likely an image is to be remembered, has traditionally been studied in computer vision either as a passive prediction task, with models regressing a scalar score, or with generative methods altering the visual input to boost the image likelihood of being remembered. Yet, none of these paradigms supports users at capture time, when the crucial question is how to improve a photo memorability. We introduce the task of **Mem**orability **Feed**back (**MemFeed**), where an automated model should provide actionable, human-interpretable guidance to users with the goal to enhance an image future recall. We also present **MemCoach**, the first approach designed to provide concrete suggestions in natural language for memorability improvement (e.g., "emphasize facial expression," "bring the subject forward"). Our method, based on Multimodal Large Language Models (MLLMs), is training-free and employs a teacher-student steering strategy, aligning the model internal activations toward more memorable patterns learned from a teacher model progressing along least-to-most memorable samples. To enable systematic evaluation on this novel task, we further introduce **MemBench**, a new benchmark featuring sequence-aligned photoshoots with annotated memorability scores. Our experiments, considering multiple MLLMs, demonstrate the effectiveness of MemCoach, showing consistently improved performance over several zero-shot models. The results indicate that memorability can not only be predicted but also taught and instructed, shifting the focus from mere prediction to actionable feedback for human creators. Dataset and code will be publicly released upon publication.
Multimodal Large Language Models (MLLMs) have shown strong potential in remote sensing (RS) through multi-task reasoning and cross-modal generalization.However, existing RS-MLLMs mainly rely on a single shared expert for all tasks, making it hard to produce reliable results. Meanwhile, the intrinsic redundancy and homogeneity of RS images bring substantial difficulties for both training and inference. These challenges directly conflict with the demands of remote sensing, which values task precision and trustworthy reasoning.To address these limitations, we propose GeoCoT, a manifold-driven mixture-of-experts (MoE) system with Chain-of-Thought (CoT) reasoning. GeoCoT introduces Mani-MoE, a sparse expert architecture grounded in local manifold mapping. It projects high-dimensional tokens onto low-rank subspaces adaptively to eliminate redundancy and uncover intrinsic structure, and then routes them through a sparse expert pathway, where gating decisions are guided by the manifold structure of the input.To optimize this architecture, we adopt a CoT-driven multi-stage training strategy. It leverages a cold-start phase for domain adaptation, followed by our RS Vision Group Relative Policy Optimization (RSV-GRPO) to systematically strengthen structured reasoning from global to objectives. Furthermore, we innovatively build *RS-CoT-20k* dataset for task-specific supervision.Extensive experiments on multi-task datasets demonstrate that GeoCoT outperforms prior approaches, achieving 5.27 \% higher average accuracy than the state-of-the-art method. Our code will be available.
VCP-Attack: Visual-Contrastive Projection for Transferable Black-Box Targeted Attacks on Large Vision-Language Models
PDF ↗Large vision-language models (LVLMs) have achieved impressive performance across a variety of multimodal tasks, yet remain vulnerable to targeted adversarial attacks, particularly in black-box settings. In this paper, we propose VCP-Attack, a transferable targeted attack framework that combines structured contrastive supervision with subspace-guided perturbation optimization. Specifically, we employ a dynamic PCA-based projection to constrain perturbations within semantically meaningful low-dimensional subspaces, and design a multi-sample contrastive loss to align adversarial features with target semantics while pushing them away from the source semantics. Extensive experiments on seven open-source and three proprietary LVLMs--including GPT-4o, Claude, and Gemini--show that VCP-Attack achieves state-of-the-art performance in black-box targeted attacks. Under a fixed perturbation budget (epsilon = 16/255), our method achieves an average attack success rate (ASR) of 94.2% on open-source models and 83.1% on proprietary models, surpassing the strongest baselines by 23.3% and 16.8%, respectively. Notably, VCP-Attack achieves a 95.6% ASR on GPT-4o. Comprehensive ablation studies and visualizations further validate the effectiveness of the dynamic subspace projection and semantic contrastive supervision. While evaluated on image captioning, our approach is model-agnostic and exhibits strong potential for broader applications to black-box adversarial settings in vision-language tasks.
BabyVLM-V2: Toward Developmentally Grounded Pretraining and Benchmarking of Vision Foundation Models
PDF ↗Early children's developmental trajectories set up a natural goal for sample-efficient pretraining of vision foundation models. We introduce BabyVLM-V2, a developmentally grounded framework for infant-inspired vision-language modeling that extensively improves upon BabyVLM-V1 through a longitudinal, multifaceted pretraining set, a versatile model, and, most importantly, DevCV Toolbox for cognitive evaluation. The pretraining set maximizes coverage while minimizing curation of a longitudinal, infant-centric audiovisual corpus, yielding video-utterance, image-utterance, and multi-turn conversational data that mirror infant experiences. DevCV Toolbox adapts all vision-related measures of the recently released NIH Baby Toolbox into a benchmark suite of ten multimodal tasks, covering spatial reasoning, memory, and vocabulary understanding aligned with early children's capabilities. Experimental results show that a compact model pretrained from scratch can achieve competitive performance on DevCV Toolbox, outperforming GPT-4o on some tasks. We hope the principled, unified BabyVLM-V2 framework will accelerate research in developmentally plausible pretraining of vision foundation models.
VRCLIP: Multimodal Canonical Correlation Alignment for CLIP-Driven Vision-Radio Person Re-Identification
PDF ↗Multimodal person Re-IDentification (ReID) aims to reliably associate specific individuals by utilizing complementary information from heterogeneous modalities. In contrast, low-frequency radio frequency (RF) signals, with their superior penetration capability and illumination invariance, provide ideal complementary information. However, a key challenge in fusing these heterogeneous modalities lies in the conventional approach that relies heavily on cross-modal distribution matching, which often over-regularizes and weakens the discriminative capacity within each modality. Rather than enforcing direct distribution alignment, canonical correlation analysis (CCA) constructs a shared subspace that maximizes cross-modal correlation, inherently balancing modality specificity and shared semantics. Inspired by this, we reformulate cross-modal alignment as a correlation maximization problem, avoiding direct constraints on feature distributions and guiding the model to harmonize intra-modal discriminative learning with cross-modal alignment. Specifically, VRCLIP first refines CLIP's visual encoder with illumination-disentangling objectives, then aligns RGB and RF embeddings in a canonical correlation subspace, and finally employs an RF-anchored reliability gate for adaptive fusion. To advance the area, we will release VRR, the first large-scale vision-radio ReID dataset with over 650K paired image-radar samples and position annotations for 31 participants. Extensive experiments show state-of-the-art 93.9% mAP and robust generalization across diverse lighting and occlusion conditions.
See It, Say It, Sorted: An Iterative Training-Free Framework for Visually-Grounded Multimodal Reasoning in LVLMs
PDF ↗Recent large vision-language models (LVLMs) have demonstrated impressive reasoning ability by generating long chain-of-thought (CoT) responses. However, CoT reasoning in multimodal contexts is highly vulnerable to visual hallucination propagation: once an intermediate reasoning step becomes inconsistent with the visual evidence, subsequent steps-even if logically valid-can still lead to incorrect final answers. Existing solutions attempt to mitigate this issue by training models to "think with images" via reinforcement learning (RL). While effective, these methods are costly, model-specific, and difficult to generalize across architectures. Differently, we present a lightweight method that bypasses RL training and provides an iterative, training-free, plug-and-play framework for visually-grounded multimodal reasoning. Our key idea is to supervise each reasoning step at test time with visual evidence, ensuring that every decoded token is justified by corresponding visual cues. Concretely, we construct a textual visual-evidence pool that guides the model's reasoning generation. When existing evidence is insufficient, a visual decider module dynamically extracts additional relevant evidence from the image based on the ongoing reasoning context, expanding the pool until the model achieves sufficient visual certainty to terminate reasoning and produce the final answer. Extensive experiments on multiple LVLM backbones and benchmarks demonstrate the effectiveness of our approach. Our method achieves 16.5%-29.5% improvements on TreeBench and 13.7% RH-AUC gains on RH-Bench, substantially reducing hallucination rates while improving reasoning accuracy without additional training.
Machine unlearning (MU) aims to remove sensitive or undesired content from pre-trained models. Existing MU methods are commonly characterized as gradually degrading model performance on undesired data to realize approximate forgetting. Despite their successes, the effectiveness in multimodal unlearning tasks remains largely unexplored. In this paper, we first conduct an in-depth analysis and reveal that traditional MU methods tend to disrupt cross-modal alignment, leading to incomplete forgetting in multimodal scenarios. To tackle this challenge, we propose VL-Eraser, a novel unlearning paradigm for VLM unlearning. VL-Eraser reformulates unlearning in VLMs as a two-stage process: distillation and deletion. Specifically, VL-Eraser first introduces a vacuum distillation that disentangles undesired knowledge from the intricate parameters of VLMs and transfers it into low-rank adapters (LoRA). After distillation, unlearning is efficiently achieved by deleting the LoRA parameters from the original model. Extensive experiments across multiple benchmarks demonstrate that VL-Eraser achieves superior unlearning performance while preserving utility compared to the state-of-the-art baselines.
Large-scale vision-language models (VLMs) have recently achieved remarkable multimodal understanding, but their massive size makes them impractical for deployment on mobile or edge devices. This raises the need for compact yet capable VLMs that can efficiently learn from powerful large teacher. However, distilling knowledge from large teacher to small student remains challenging due to their large size gap: the student often fails to reproduce the teacher's complex, high-dimensional representations, leading to unstable learning and degraded performance. To address this, we propose Masters (Masking teacher and reinforcing student), a mask-progressive reinforcement learning (RL) distillation framework. Masters first masks and non-dominant weights of the teacher to reduce unnecessary complexity, then progressively restores the teacher from mask to gradually increase the teacher capacity during training. This strategy allows the student to learn richer representations of teacher in a smooth and stable manner. To further refine knowledge transfer, Masters integrates an offline RL stage with two complementary rewards: an accuracy reward that measures the correctness of the generated responses, and a distillation reward that quantifies the ease of their responses' transferability from teacher to student. Unlike online think-answer RL paradigms that are computationally expensive and generate lengthy responses, our offline RL leverages pre-generated responses from masked teachers. These provide rich yet efficient guidance, enabling the students to achieve strong performance without requiring the think-answer process. Extensive experiments across diverse VLM benchmarks demonstrate that Masters outperforms existing compact VLMs and partially surpasses large ones, while being far more efficient. Moreover, gradually increasing the teacher sizes during distillation (e.g., from 14B to 38B) yields smoother convergence and stronger generalization than one-shot distillation (e.g., 38B), revealing a scalable path toward efficient and deployable VLMs.