CVPR 2026 Papers

Adversarial attacks have evolved from simply disrupting predictions on conventional task-specific models to the more complex goal of manipulating image semantics in Large Vision-Language Models (LVLMs). However, existing methods struggle with controllability and cannot precisely manipulate the semantics of specific concepts in an image. We attribute this limitation to semantic entanglement in the patch-token representations that adversarial attacks typically operate on: global context aggregated by self-attention in the vision encoder dominates patch features, making them unreliable for precise local semantic manipulation. Our systematic investigation reveals a key insight: value features, computed within the transformer attention block, provide much more precise handles for manipulation. We show that these value features suppress global-context channels, allowing them to retain high-entropy, disentangled local semantic information. Building on this discovery, we propose V-Attack, a novel method for precise local semantic attacks. V-Attack targets value features and introduces two core components: (1) a Self-Value Enhancement module to refine the intrinsic semantic richness of value features, and (2) a Text-Guided Value Manipulation module that uses text prompts to locate the source concept and optimize it toward a target concept. By bypassing entangled patch features, V-Attack achieves highly effective semantic control. Extensive experiments across diverse LVLMs, including LLaVA, InternVL, DeepSeek-VL, and GPT-4o, show that V-Attack improves attack success rate by an average of 36% over state-of-the-art methods, exposing critical vulnerabilities in modern vision-language understanding.

Vision-Language Navigation (VLN) aims to enable agents to navigate to a target location based on language instructions. Traditional VLN often follows a close-set assumption, i.e., training and test data share the same style of the input images and instructions. However, the real world is open and filled with various unseen environments, posing enormous difficulties for close-set methods.To this end, we focus on the General Scene Adaptation (GSA-VLN) task, aiming to learn generalized navigation ability by introducing diverse environments and inconsistent intructions.Towards this task, when facing unseen environments and instructions, the challenge mainly lies in how to enable the agent to dynamically produce generalized strategies during the navigation process. Recent research indicates that by means of fast and slow cognition systems, human beings could generate stable policies, which strengthen their adaptation for open world. Inspired by this idea, we propose the slow4fast-VLN, establishing a dynamic interactive fast-slow reasoning framework.The fast-reasoning module, an end-to-end strategy network, outputs actions via real-time input. It accumulates execution records in a history repository to build memory.The slow-reasoning module analyze the memories generated by the fast-reasoning module. Through deep reflection, it extracts experiences that enhance the generalization ability of decision-making. These experiences are structurally stored and used to continuously optimize the fast-reasoning module. Unlike traditional methods that treat fast-slow reasoning as independent mechanisms, our framework enables fast-slow interaction. By leveraging the experiences from slow reasoning, it continually improves the accuracy and generalization ability of fast decisions. This interaction allows the system to continuously adapt and efficiently execute navigation tasks when facing unseen scenarios. Extensive experiments demonstrate the superiorities of our method.

We present Franca (pronounced Fran-ka): free one; the first fully open-source (data, code, weights) vision foundation model that matches and in many cases surpasses the performance of state-of-the-art proprietary models, e.g., DINOv2, CLIP, SigLIPv2, etc. Our approach is grounded in a transparent training pipeline inspired by Web-SSL and uses publicly available data: ImageNet-21K and a subset of ReLAION-2B. Beyond model release, we tackle critical limitations in self-supervised learning clustering methods. Existing approaches assign image features to large codebooks via clustering algorithms such as Sinkhorn-Knopp, but they often overlook the inherent ambiguity in cluster semantics. To address this, we introduce a multi-head clustering projector based on nested Matryoshka representations. This design progressively refines features into increasingly fine-grained clusters without increasing the model size, producing higher-quality dense representations. Additionally, we propose a novel positional disentanglement strategy that explicitly removes positional biases from dense representations.This leads to consistent gains on several downstream benchmarks, demonstrating the utility of cleaner feature spaces. Our contributions establish a new standard for transparent, high-performance vision models and open a path toward more reproducible and generalizable foundation models for the broader AI community.

Feature matching is a critical task in geometric perception, yet existing methods often struggle under domain shifts and illumination changes due to reliance on visual-only learning and expensive 3D supervision. In this paper, we present TextFM, the first language-guided feature matching framework that incorporates domain-invariant semantic information from vision-language models (VLMs). Built upon a detector-free architecture, TextFM leverages textual embeddings as instance-level queries to provide global semantic context during coarse-level matching, enhancing robustness in challenging scenarios such as textureless surfaces and cross-domain shifts. Additionally, we integrate illumination-invariant physical priors and apply Low-Rank Adaptation (LoRA) to efficiently fine-tune Vision Foundation Models (VFMs) for more robust visual feature extraction. Extensive experiments on outdoor and indoor datasets show that our method outperforms other state-of-the-art methods. In addition, we contribute a synthetic day-night matching benchmark for rigorous evaluation under extreme lighting conditions. Together, our method and dataset establish a strong foundation for robust and generalizable feature matching under real-world constraints.

Recent advances in CLIP-based continual learning have shown the potential of leveraging pre-trained vision-language models for sequential tasks. However, existing methods overlook a key problem we call Asymmetric Drift. In unimodal CLIP-based continual learning, the visual branch undergoes stronger adaptation because the visual distribution shifts significantly, whereas the text branch remains relatively stable due to the low variance of textual prompts. This imbalance increases the modality distance and degrades cross-modal alignment over time. To address this issue, we propose CCA-CL, a framework that accumulates visual-textual covariance statistics across tasks and solves Canonical Correlation Analysis to compute a shared subspace. In this subspace, the distance between visual and textual features is minimized, enabling better alignment without modifying CLIP parameters. This also makes our method naturally compatible with exemplar-free CL settings. To further capture nonlinear relationships that linear Canonical Correlation Analysis is hard to model, we introduce Random Fourier Projection as an extension. Experimental results demonstrate that CCA-CL effectively mitigates the asymmetric drift problem and achieves state-of-the-art performance on several benchmarks.

Counting is a core capability for multimodal large language models (MLLMs), yet there is no unified counting dataset to rigorously evaluate this ability across image, text, and audio. We present UNICBench, a unified multimodal, multi-level counting benchmark and evaluation toolkit with accurate ground truth, deterministic numeric parsing, and stratified reporting. The corpus comprises 5,300 images (5,508 QA), 872 documents (5,888 QA), and 2,069 audio clips (2,905 QA), annotated with a three-level capability taxonomy and difficulty tags. Under a standardized protocol with fixed splits/prompts/seeds and modality-specific matching rules, we evaluate 45 state-of-the-art MLLMs across modalities. Results show strong performance on some basic counting tasks but significant gaps on reasoning and the hardest partitions, highlighting long-tail errors and substantial headroom for improving general counting. UNICBench offers a rigorous and comparable basis for measurement and a public toolkit to accelerate progress.

Generating realistic and physically plausible 3D Human-Object Interactions (HOI) remains a key challenge in motion generation. One primary reason is that describing these physical constraints with words alone is difficult. To address this limitation, we propose a new paradigm: extracting rich interaction priors from easily accessible 2D images. Specifically, we introduce ViHOI, a novel framework that enables diffusion-based generative models to leverage rich, task-specific priors from 2D images to enhance generation quality. We utilize a large Vision-Language Model (VLM) as a powerful prior-extraction engine and adopt a layer-decoupled strategy to obtain visual and textual priors. Concurrently, we design a Q-Former-based adapter that compresses the VLM's high-dimensional features into compact prior tokens, which significantly facilitates the conditional training of our diffusion model. Our framework is trained on motion-rendered images from the dataset to ensure strict semantic alignment between visual inputs and motion sequences. During inference, it leverages reference images synthesized by a text-to-image generation model to improve generalization to unseen objects and interaction categories. Experimental results demonstrate that ViHOI achieves state-of-the-art performance, outperforming existing methods across multiple benchmarks and demonstrating superior generalization. The code for this work will be released at https//github.com/MPI-Lab/ViHOI.

With the rapid development of pre-training technologies, adapting large-scale Vision-Language Models (VLMs) for video understanding i.e. image-to-video transfer learning has become a dominant paradigm. To achieve superior performance, it raises as an effective strategy among recent advances to employ Mixture-of-Experts (MoE) to enhance VLMs' temporal modeling capabilities. However, conventional MoE designs suffer from expert homogenization, where all experts act as identical generalists, inefficiently learning spatio-temporal features from undifferentiated video streams.To overcome this problem, we propose VidPrism, a novel heterogeneous temporal Mixture-of-Experts framework. VidPrism pioneers a division of labor by deploying functionally specialized experts, each assuming a role ranging from spatial understanding to temporal modeling.To feed these specialists appropriately, we introduce a content-aware, multi-rate sampling module that dynamically generates streams ranging from semantically rich to motion-focused representations, providing specialized inputs for experts.Furthermore, a dynamic, bidirectional fusion mechanism enables synergistic information exchange between these pathways, leading to a comprehensive video representation. Extensive experiments on various video recognition benchmarks demonstrate that VidPrism achieves state-of-the-art performance and effectively fosters expert specialization.

Multi-view image compression (MIC) aims to achieve high compression efficiency by exploiting inter-image correlations, playing a crucial role in 3D applications. As a subfield of MIC, distributed multi-view image compression (DMIC) offers performance comparable to MIC while eliminating the need for inter-view information at the encoder side.However, existing methods in DMIC typically treat all images equally, overlooking the varying degrees of correlation between different views during decoding, which leads to suboptimal coding performance. To address this limitation, we propose a novel OmniParallax Attention Mechanism (OPAM), which is a general mechanism for explicitly modeling correlations and aligned features between arbitrary pairs of information sources.Building upon OPAM, we propose a Parallax Multi Information Fusion Module (PMIFM) to adaptively integrate information from different sources. PMIFM is incorporated into both the joint decoder and the entropy model to construct our end-to-end DMIC framework, ParaHydra.Extensive experiments demonstrate that ParaHydra is the first DMIC method to significantly surpass state-of-the-art MIC codecs, while maintaining low computational overhead. Performance gains become more pronounced as the number of input views increases. Compared with LDMIC, ParaHydra achieves bitrate savings of 19.72\% on WildTrack(3) and up to 24.18\% on WildTrack(6), while significantly improving coding efficiency (as much as 65xin decoding and 34xin encoding).

Multimodal Web Search Agents demonstrate a practically valuable capability by fusing information from diverse modalities (e.g., text and vision), retrieved iteratively from the internet, to address complex user queries. However, the visual modality is prone to information overload, and the noise contained within it--such as irrelevant background details or complex structures--can disrupt the model's attention, misdirecting its operational focus toward an erroneous path. To address the aforementioned challenge, we propose ReFAct (Reasoning, Focusing, and Acting), a novel framework that empowers the agent to actively manage its cross-modal context. This allows the agent to adjust its operational focus, thereby mitigating the impact of noise on multimodal Web Search Agents. Specifically, ReFAct employs a Grounding tool for active visual perception to dynamically filter information. We also design external memory-based Defocus/Refocus operations for selective information retention, further modulating information density within the multimodal context. Ultimately, this ensures the agent maintains focus during problem-solving. To evaluate and enhance agent capabilities in complex and noisy multimodal contexts, we first propose a pipeline for constructing datasets with flexible complexity. We introduce a new open-source benchmark: GroundedVQA. Finally, we experimentally demonstrate the effectiveness of our proposed method on GroundedVQA and other widely-used benchmarks.

Vision-Language-Action Models (VLAs) have demonstrated significant promise in generalizing to complex, long-horizon robotic manipulation tasks. However, their performance remains brittle, as they are typically trained on trajectory-monotonic, failure-free demonstrations. This reliance on "perfect" data leaves them unable to recover from common execution errors, such as a missed grasp, a dropped object, or an unexpected collision. In this paper, we propose FLARE, a novel framework that endows VLAs with robust error recovery capabilities through a "Retry" and "Reset" paradigm. First, we introduce a "Retry" mechanism by injecting perturbation and bridging segments that decouple robot pose from environment state into demonstrations, enabling the policy to autonomously handle execution deviations. Second, to address critical, state-breaking (OOD) failures, we introduce a "Reset" pipeline. We leverage an MLLM for offline failure analysis to automatically identify OOD states from execution videos. This analysis enables the efficient, targeted collection of a small library of object-centric "Reset" skills, which are trained to restore the environment to a task-valid state. Our full framework integrates these learned policies. At inference, an online MLLM monitor arbitrates between task execution and "Reset" skills. Experiments on challenging, contact-rich manipulation tasks show our approach significantly improves task success and robustness.

Supervised and unsupervised homography estimation methods depend on image pairs tailored to specific modalities to achieve high accuracy. However, their performance deteriorates substantially when applied to unseen modalities. To address this issue, we propose a training data synthesis method that generates unaligned image pairs with ground-truth offsets from a single input image. Our approach renders the image pairs with diverse textures and colors while preserving their structural information. These synthetic data empower the trained model to achieve greater robustness and improved generalization across various domains. Additionally, we design a network to fully leverage cross-scale information and decouple color information from feature representations, thus improving estimation accuracy. Extensive experiments show that our training data synthesis method improves generalization performance. The results also confirm the effectiveness of the proposed network.

Recent advances in Multimodal Large Language Models have greatly improved visual understanding and reasoning, yet their quadratic attention and offline training protocols make them ill-suited for streaming settings where frames arrive sequentially and future observations are inaccessible. We diagnose a core limitation of current Video-LLMs, namely Time-Agnosticism, in which videos are treated as an unordered bag of evidence rather than a causally ordered sequence, yielding two failures in streams: temporal order ambiguity, in which the model cannot follow or reason over the correct chronological order, and past-current focus blindness where it fails to distinguish present observations from accumulated history. We present WeaveTime, a simple, efficient, and model-agnostic framework that first teaches order and then uses order. We introduce a lightweight Temporal Reconstruction objective--our Streaming Order Perception enhancement--that instills order-aware representations with minimal finetuning and no specialized streaming data. At inference, a Past-Current Dynamic Focus Cache performs uncertainty-triggered, coarse-to-fine retrieval, expanding history only when needed. Plugged into exsiting Video-LLM without architectural changes, WeaveTime delivers consistent gains on representative streaming benchmarks, improving accuracy while reducing latency. These results establish WeaveTime as a practical path toward time-aware stream Video-LLMs under strict online, time-causal constraints. Code and weights will be made publicly available.

While model merging has demonstrated remarkable success for large language models (LLMs), its application to vision-language models (VLMs) remains largely underexplored. Recent methods attempt to enhance VLM reasoning capabilities by integrating specialized LLM parameters through layer-wise merging. However, existing paradigms suffer from two critical limitations: (1) strict positional correspondence, enforcing rigid one-to-one layer alignment, and (2) uniform merging weights applied across all layers. These constraints fail to account for substantial functional disparities between VLM and LLM layers, potentially misaligning incompatible layers and leading to detrimental parameter combinations. To address these, we propose Chain-of-Merging (CoM), a framework that adaptively adjusts merging plans for different images and questions, comprising two key stages: (1) Adaptive Layer Matching, which identifies optimal layer pairings based on structural and semantic matching scores while filtering incompatible pairings, and (2) Dynamic Weight Merging, which determines layer-specific merging weights based on matching scores and employs spherical linear interpolation to minimize memory overhead. Extensive experiments demonstrate that CoM achieves substantial performance improvements, with Qwen2.5-VL-7B + Qwen2.5-Math-7B attaining a 4.4% average improvement on mathematical reasoning benchmarks while enhancing general visual understanding, significantly outperforming existing training-free methods.

In Vision-Language-Action (VLA) models, action chunking (i.e., executing a sequence of actions without intermediate replanning) is a key technique to improve robotic manipulation abilities. However, a large chunk size reduces the model's responsiveness to new information, while a small one increases the likelihood of mode-jumping, jerky behavior resulting from discontinuities between chunks. Therefore, selecting the optimal chunk size is an urgent demand to balance the model's reactivity and consistency. Unfortunately, a dominant trend in current VLA models is an empirical fixed chunk length at inference-time, hindering their superiority and scalability across diverse manipulation tasks. To address this issue, we propose a novel Adaptive Action Chunking (AAC) strategy, which exploits action entropy as the cue to adaptively determine the chunk size based on current predictions. Extensive experiments on a wide range of simulated and real-world robotic manipulation tasks have demonstrated that our approach substantially improves performance over the state-of-the-art alternatives. The videos and source code are publicly available at https://lance-lot.github.io/adaptive-chunking.github.io.

Multimodal large language models (MLLMs) often suffer from perceptual impairments under extended reasoning modes, particularly in visual question answering (VQA) tasks. We identify attention dispersion as the underlying cause: during multi-step reasoning, model's visual attention becomes scattered and drifts away from question-relevant regions, effectively "losing focus" on the visual input. To better understand this phenomenon, we analyze the attention maps of MLLMs and observe that reasoning prompts significantly reduce attention to regions critical for answering the question. We further find a strong correlation between model's overall attention on image tokens and the spatial dispersiveness of model's attention within the image. Leveraging this insight, we propose a training-free Visual Region-Guided Attention (VRGA) framework that selects visual heads based on an entropy-focus criterion and reweights their attention, effectively guiding the model to focus on question-relevant regions during reasoning. Extensive experiments on vision-language benchmarks demonstrate that our method effectively alleviates perceptual degradation, leading to improvements in visual grounding and reasoning accuracy, while offering interpretable insights into how MLLMs process visual information.

Videos are unique in their ability to capture actions which transcend multiple frames. Accordingly, action recognition has long been a quintessential task for video models. Unfortunately, due to a lack of sufficiently diverse and challenging data, modern vision-language models (VLMs) are no longer evaluated on their action recognition capabilities. To revitalize action recognition in the era of VLMs, we advocate for a returned focus on domain-specific actions. To this end, we introduce VideoNet, a domain-specific action recognition benchmark covering 1,087 distinct actions from 38 domains. VLMs struggle immensely on VideoNet, with Gemini 2.5 Pro performing only 15.8 percentage points better than random chance. To improve model performance we provide in-context demonstrations, but only see a 3% boost in VLM performance compared to a 13% increase in non-expert human accuracy, suggesting that VLMs are poor few-shot learners. At last, we collect a large-scale training dataset containing nearly 500k video question-answer pairs. Fine-tuning an open-weight 4B model on our data, we surpass all Gemini models on the VideoNet benchmark. We release all of our data, inviting the community to explore new techniques to improve domain-specific action recognition capabilities and few-shot learning in video models.

Structure-from-Motion (SfM) is a cornerstone of 3D perception, yet current methods often fail when applied to complex videos involving challenging camera motions or dynamic scenes.Compounding the problem, the field lacks reliable ground-truth benchmarks for such difficult scenarios, making it hard to gauge real-world progress, or pinpoint where improvements are most needed.To address this gap, we introduce a new benchmark for evaluating camera pose estimation.Our key insight is to leverage online panoramic 360deg as a source of data from which to construct challenging clips, while still enabling robust ground-truth trajectory recovery.The panoramic nature of these videos provides richer visual context for tracking camera motion, even when parts of the view are affected by blur, motion, or dynamic objects.By tracking camera motion across full 360deg videos, we crop and reproject selected portions to generate perspective-view clips that serve as our benchmark---ORBIT---a diverse collection of 100 video clips.Experiments show that COLMAP and other state-of-the-art SfM methods struggle to accurately estimate camera positions on our benchmark, indicating that it remains a challenging and open problem space for future research.As a result, ORBIT provides a valuable testbed where researchers can meaningfully compete and measure progress on truly challenging, real-world SfM problems.

In this paper, we present a holistic multimodal benchmark that evaluates the reasoning capabilities of MLLMs with an explicit focus on reasoning width, a complementary dimension to the more commonly studied reasoning depth. Specifically, reasoning depth measures the model's ability to carry out long-chain, sequential reasoning in which each step is tightly and rigorously linked to the next. Reasoning width tends to focus more on the model's capacity for broad trial-and-error search or multi-constrained optimization: it must systematically traverse many possible and parallelized reasoning paths, apply diverse constraints to prune unpromising branches, and identify valid solution routes for efficient iteration or backtracking. To achieve it, we carefully curate 1200+ high-quality multimodal cases spanning heterogeneous domains, and propose a fine-grained tree-of-thought evaluation protocol that jointly quantifies reasoning width and depth. We evaluate 12 major model families (over 30 advanced MLLMs) across difficulty tiers, question types, and required skills. Results show that while current models exhibit strong performance on general or common-sense VQA tasks, they still struggle to combine deep sequential thought chains with wide exploratory search to perform genuine insight-based reasoning. Finally, we analyze characteristic failure modes to provide possible directions for building MLLMs that reason not only deeper but also wider.