CVPR 2026 Papers

Test-time adaptation for video understanding, which enables vision-language models (VLMs) to generalize to downstream tasks such as action recognition, has demonstrated substantial value in real-world applications. Existing memory-based methods typically build a visual cache from high-confidence test videos and perform inference via a two-step modality mapping chain, i.e., vision-vision and vision-text. However, due to the asymmetry of the two mappings, the chain exhibits non-transitivity, hindering the generalization of VLMs. To this end, we propose a novel training-free Condensed Dynamic Adapter ConDA for action recognition, which leverages vision-text alignment to guide vision-vision alignment. It first selects semantic patches based on the semantic activation probability obtained from the vision-text alignment (Probability-based Semantic Patch Selection, PSPS), and then adaptively constructs spatial-temporal video tubes based on patch-level visual similarity (Adaptive Tube Construction, ATC). We conduct extensive experiments on seven benchmarks with different backbones and baselines. The quantitative results demonstrate that ConDA is compatible with arbitrary VLM and generalizes well across complex scenarios, such as long-term and egocentric scenarios. In addition, qualitative analyses showcase the interpretability of ConDA in terms of capturing semantic cues.

Recent advances in vision-language models (VLMs) have made them highly effective at reasoning tasks. However, the principles underlying the construction of performant VL reasoning training datasets remain poorly understood. In this work, we introduce several data curation approaches and study their impacts on VL reasoning capabilities by carefully controlling training and evaluation setups. We analyze the effects of context (image and question pair) sources, implement targeted data interventions, and explore scaling up images, questions, and chain-of-thought (CoT) solutions. Our findings reveal that (a) context source strategies significantly affect VLM performance, (b) interventions such as auxiliary signals from image captions and the inclusion of text-only reasoning yield substantial gains, and (c) scaling all data dimensions (e.g., unique questions per image and unique CoTs per image-question pair) consistently improves reasoning capability. Motivated by these insights, we introduce HoneyBee, a large-scale, high-quality CoT reasoning dataset with 2.5M examples consisting 350K image-question pairs. VLMs trained with HoneyBee outperform state-of-the-art models across model sizes. For instance, a HoneyBee-trained VLM with 3B parameters outperforms the SOTA model and the base model by 7.8% and 24.8%, respectively, on MathVerse. Furthermore, we propose a test-time scaling strategy that reduces decoding cost by 73% without sacrificing accuracy. Overall, this work presents improved strategies for VL reasoning dataset curation research.

We present a novel unsupervised framework to unlock vast unlabeled human demonstration data from continuous industrial video streams for Vision-Language-Action (VLA) model pre-training. Our method first trains a lightweight motion tokenizer to encode motion dynamics, then employs an unsupervised action segmenter leveraging a novel "Latent Action Energy" metric to discover and segment semantically coherent action primitives. The pipeline outputs both segmented video clips and their corresponding latent action sequences, providing structured data directly suitable for VLA pre-training. Evaluations on public benchmarks and a proprietary electric motor assembly dataset demonstrate effective segmentation of key tasks performed by humans at workstations. Further clustering and quantitative assessment via a Vision-Language Model confirm the semantic coherence of the discovered action primitives. To our knowledge, this is the first fully automated end-to-end system for extracting and organizing VLA pre-training data from unstructured industrial videos, offering a scalable solution for embodied AI integration in manufacturing.

Human perception of visual similarity is inherently adaptive and subjective, depending on the users' interests and focus. However, most image retrieval systems fail to reflect this flexibility, relying on a fixed, monolithic metric that cannot incorporate multiple conditions simultaneously. To address this, we propose CLAY, an adaptive similarity computation method that reframes the embedding space of pretrained Vision-Language Models (VLMs) as a text-conditional similarity space without additional training. This design separates the textual conditioning process and visual feature extraction, allowing highly efficient and multi-conditioned retrieval with fixed visual embeddings. We also construct a synthetic evaluation dataset CLAY-EVAL, for comprehensive assessment under diverse conditioned retrieval settings. Experiments on standard datasets and our proposed dataset show that CLAY achieves state-of-the art retrieval accuracy and notable computational efficiency compared to previous works.

Recent progress in multimodal large language models (MLLMs) has highlighted the challenge of efficiently bridging pre-trained Vision-Language Models (VLMs) with Diffusion Models. While methods using a fixed number of learnable query tokens offer computational efficiency, they suffer from task generalization collapse, failing to adapt to new tasks that are distant from their pre-training tasks. To overcome this, we propose Noisy Query Tokens, which learn a distributed representation space between the VLM and Diffusion Model via end-to-end optimization, enhancing continual learning. Additionally, we introduce a VAE branch with linear projection to recover fine-grained image details. Experimental results confirm our approach mitigates generalization collapse and enables stable continual learning across diverse tasks.

Reinforcement Learning with Verifiable Rewards (RLVR) traditionally relies on a sparse, outcome-based signal. Recent work shows that providing a fine-grained, model-intrinsic signal--rewarding the confidence growth in the ground-truth answer--effectively improves language reasoning training by providing step-level guidance without costly external models. While effective for unimodal text, we find that naively applying this global reward to vision-language (V-L) reasoning is a suboptimal strategy, as the task is a heterogeneous mix of sparse visual perception and dense textual reasoning. This global normalization creates mixture-induced signal degradation, where the training signal for visual steps is statistically distorted by the predominant textual steps. We propose Perception-Decomposed Confidence Reward (PDCR), a framework that solves this by aligning the reward structure with the task's heterogeneous nature. PDCR first performs an unsupervised skill decomposition, introducing a model-internal Visual Dependence Score to quantify visual reliance and applying a clustering algorithm to separate perception and reasoning steps. Based on this, PDCR computes a decomposed advantage by normalizing confidence gains within each skill cluster. This intra-cluster normalization provides a stable, correctly-scaled signal for both perception and reasoning. We demonstrate that PDCR outperforms the naive, global-reward formulation and sparse-reward baselines on key V-L reasoning benchmarks.

Large Vision-Language Models (LVLMs) excel in visual understanding and reasoning, but the excessive visual tokens lead to high inference costs. Although recent token reduction methods mitigate this issue, they mainly target single-turn Visual Question Answering (VQA), leaving the more practical multi-turn VQA (MT-VQA) scenario largely unexplored. MT-VQA introduces additional challenges, as subsequent questions are unknown beforehand and may refer to arbitrary image regions, making existing reduction strategies ineffective. Specifically, current approaches fall into two categories: prompt-dependent methods, which bias toward the initial text prompt and discard information useful for subsequent turns; prompt-agnostic ones, which, though technically applicable to multi-turn settings, rely on heuristic reduction metrics such as attention scores, leading to suboptimal performance. In this paper, we propose a learning-based prompt-agnostic method, termed MetaCompress, overcoming the limitations of heuristic designs. We begin by formulating token reduction as a learnable compression mapping, unifying existing formats such as pruning and merging into a single learning objective. Upon this formulation, we introduce a data-efficient training paradigm capable of learning optimal compression mappings with limited computational costs. Extensive experiments on MT-VQA benchmarks and across multiple LVLM architectures demonstrate that MetaCompress achieves superior efficiency-accuracy trade-offs while maintaining strong generalization across dialogue turns. Our code is available at https://github.com/MArSha1147/MetaCompress.

Modern generative models can propose striking 3D vehicle shapes from text and images, but turning these sketches into aerodynamically efficient, regulation compliant designs still requires weeks of high-fidelity computational fluid dynamics (CFD) and manual iteration. As a result, fast 3D generation without trustworthy physics in the loop does little to reduce end-to-end design time. We study how an AI agent can close this loop under a strict CFD budget. We introduce AeroAgent, a vision-physics-decision framework built around a single 3D, editable surface representation for vehicle shapes. A vision module turns text and 2D references into diverse, standardized 3D candidates and supports image-level edits. A physics module, AeroFormer is a geometry-guided Transformer surrogate trained on a large-scale vehicle aerodynamics dataset of roughly 50k CFD simulations; three task-specific heads predict drag Cd, surface pressure, and velocity fields on shared 3D grids. A decision module encodes regulatory size limits and aesthetic constraints as feasibility tests, uses prototype priors and surrogate sensitivities to guide free-form deformation edits, and runs a budget-aware propose evaluate refine loop in which only the final top K shapes are confirmed by high-fidelity CFD. In extensive experiments across five common vehicle classes, running only five propose evaluate refine iterations per vehicle reduces drag by an average of 2-12% and cuts high-fidelity CFD calls by 50-80% compared to baseline workflows, while preserving or improving styling quality.

Reliable generalization metrics are fundamental to the evaluation of machine learning models. Especially in high-stakes applications where labeled target data are scarce, evaluation of models' generalization performance under distribution shift is a pressing need. We focus on two practical scenarios: (1) Before deployment, how to select the best model for unlabeled target data? (2) After deployment, how to monitor model performance under distribution shift? The central need in both cases is a reliable and label-free proxy metric. Yet existing proxy metrics, such as model confidence or accuracy-on-the-line, are often unreliable as they only assess model output while ignoring the internal mechanisms that produce them. We address this limitation by introducing a new perspective: using the inner workings of a model, i.e., circuits, as a predictive metric of generalization performance. Leveraging circuit discovery, we extract the causal interactions between internal representations as a circuit, from which we derive two metrics tailored to the two practical scenarios. (1) Before deployment, we introduce Dependency Depth Bias, which measures different models' generalization capability on target data. (2) After deployment, we propose Circuit Shift Score, which predicts a model's generalization under different distribution shifts. Across various tasks, both metrics demonstrate significantly improved correlation with generalization performance, outperforming existing proxies by an average of 13.4% and 34.1%, respectively. Our code is available at https://github.com/deep-real/GenCircuit.

Visual-Language-Action (VLA) models report impressive success rates exceeding 95% on robotic manipulation benchmarks, yet these results may mask fundamental weaknesses in robustness. Current simulation-based robustness evaluations suffer from narrow perturbation coverage, manual design constraints, and coarse-grained analysis that fails to reveal when and how models fail. To address this gap, we propose LIBERO-Plus, a comprehensive, automatic, and fine-grained evaluation framework with controlled perturbations across seven dimensions: object layouts, camera viewpoints, robot initial states, language instructions, lighting conditions, background textures, and sensor noise. Our systematic analysis of ten state-of-the-art models reveals consistent brittleness beneath apparent competence, with performance dropping from 95% to below 30% under modest perturbations. Our findings challenge the assumption that high benchmark scores equate to true competency and highlight the need for evaluation practices that assess reliability under realistic variation.

Adapting vision-language models to remote sensing imagery remains challenging due to two key factors: limited semantic coverage in textual representations and insufficient adaptability of visual features. These issues are particularly significant in aerial scenes, which involve various visual appearances and fine-grained object distinctions. We propose AVION, a knowledge distillation framework tailored for remote sensing adaptation of vision-language models. The teacher module constructs semantically rich textual prototypes by collecting descriptions from a large language model and verifying validity using remote sensing image features. The student module integrates lightweight and learnable prompts into both vision and language encoders, guided by the teacher to align embeddings and their cross-modal relationships. Once trained, the student operates independently during inference. Experiments on six optical remote sensing benchmarks show that AVION improves few-shot classification and base-class accuracy without degrading generalization to novel categories. It also enhances mean recall for cross-modal retrieval, with minimal additional trainable parameters.

Vision-Language Pre-training (VLP) models, while achieving state-of-the-art performance on various multimodal tasks, exhibit significant vulnerability to multimodal adversarial examples. In black-box attack scenarios of VLP models, a key challenge lies in the limited transferability of these adversarial examples. Existing methods to enhance transferability often suffer from an excessive dependence on the source model and a reliance on limited and fixed transformation techniques. To overcome these limitations, we propose a novel Transform to Transfer Attack (TTA) method. Our approach introduces a learnable transformation mechanism that adaptively selects optimal combinations of transformations to maximize input diversity, and incorporates integrated gradients to mitigate over-reliance on the source model, thereby refining the attack optimization process. Extensive experiments demonstrate that TTA achieves outstanding attack performance in downstream tasks, outperforming current state-of-the-art attack methods across different VLP architectures.

The integration of vision and language in Vision-Language Models (VLMs), while enabling multimodal capabilities, inherently expands their attack surface. Among existing white-box jailbreak methods, suffix-optimization-based approaches often rely on gradient approximations over discrete token spaces, yielding insufficient guidance and causing optimization to stagnate in local optima, while image-perturbation-based ones frequently exhibit poor cross-model transferability. In this work, we introduce DGSIP, a Dissonance-Guided Suffix Optimization and Image-Phrase Injection framework. DGSIP leverages predictive dissonance between the target model and an unaligned model to identify tokens suppressed by safety alignment, using them as a more effective signal than gradient-based cues for suffix optimization. It further reinforces the attack by jointly optimizing the content and presentation of phrase embedded in images to leverage VLMs' cross-modal sensitivity. Our extensive experiments demonstrate that DGSIP outperforms prior baselines across multiple safety benchmarks and a range of open-source VLMs (e.g., MiniGPT-4, InstructBlip and LLaVA). Notably, compared to baselines, our method exhibits much stronger transferability to commercial black-box VLMs, such as GPT-4o-Mini, Gemini 2.0 Flash and Qwen 2.5-VL. Based upon DGSIP, we empirically reveal critical vulnerabilities in the safeguard mechanisms of current VLMs, highlighting the need for more robust defense strategies. The implementation is available on https://github.com/Trusted-LLM/DGSIP.

Developing a robust visual quality assessment (VQualA) large multi-modal model (LMM) requires achieving versatility, powerfulness, and transferability. However, existing VQualA LMMs typically focus on a single task and rely on full-parameter fine-tuning, which makes them prone to overfitting on specific modalities or task types, thereby limiting their generalization capacity and transferability. To address this, we propose a vision-encoder-centered generative pre-training pipeline and develop the VITAL-Series LMMs. (1) We adopt a machine-executed annotation-scrutiny paradigm, constructing over 4.5M vision-language (VL) pairs-the largest VQualA training dataset to date. (2) We employ a multi-task training workflow that simultaneously enhances the model's quantitative scoring precision and strengthens its capability for quality interpretation across both image and video modalities. (3) Building upon the vision encoder, we realize an efficient model zoo extension: the model zoo exhibits strong zero-shot performance, and each paired decoder requires only a swift warm-up using less than 1/1000 of the pre-training data to achieve performance comparable to the fully trained counterpart. Overall, our work lays a cornerstone for advancing toward the foundation LMM for VQualA. The project is at https://github.com/jzhws/VITAL-Series.

Multilingual document and scene text understanding plays an important role in applications such as search, finance, and public services. However, most existing benchmarks focus on high-resource languages and fail to evaluate models in realistic multilingual environments. In Southeast Asia, the diversity of languages, complex writing systems, and highly varied document types make this challenge even greater. We introduce SEA-Vision, a benchmark that jointly evaluates Document Parsing and Text-Centric Visual Question Answering (TEC-VQA) across 11 Southeast Asian languages. SEA-Vision contains 15,234 document parsing pages from nine representative document types, annotated with hierarchical page-, block-, and line-level labels. It also provides 7,496 TEC-VQA question-answer pairs that probe text recognition, numerical calculation, comparative analysis, logical reasoning, and spatial understanding. To make such multilingual, multi-task annotation feasible, we design a hybrid pipeline for Document Parsing and TEC-VQA. It combines automated filtering and scoring with MLLM-assisted labeling and lightweight native-speaker verification, greatly reducing manual labeling while maintaining high quality. We evaluate several leading multimodal models and observe pronounced performance degradation on low-resource Southeast Asian languages, highlighting substantial remaining gaps in multilingual document and scene text understanding. We believe SEA-Vision will help drive global progress in document and scene text understanding.

The Vision Transformer (ViT) has surpassed Convolutional Neural Networks (CNNs) in performance, becoming the de facto architecture in modern computer vision. However, despite its superior representational capacity, research on the adversarial robustness of ViTs remains limited, with most studies still biased toward CNN-based models. This work aims to address this architectural bias and conduct an in-depth analysis of the interaction between ViTs and adversarial training (AT).We first show that ViTs can identify semantic components of objects through their class attention maps, indicating that adversarially trained ViTs inherently encode strong semantic priors. Next, using the proposed Gradient Path Masking (GPM) analysis, we examine the internal information flow of ViTs and verify that the residual path serves as a major bottleneck that provides advantageous information to adversaries. Furthermore, our inter-patch relation analysis reveals that adversarially trained ViTs tend to rely more on global than local relationships in early layers--a novel observation suggesting a fundamental potential incompatibility between ViTs and hybrid architectures that inject CNN-style inductive biases.Building upon these findings, we design a simple yet effective two-stage AT scheme to mitigate this structural incompatibility, achieving simultaneous improvements in robustness and generalization across various ViT variants and training methods. The proposed method is compatible with a wide range of AT frameworks and models.

Multimodal Large Language Models (MLLMs) have achieved remarkable progress in vision-language understanding, yet how they internally integrate visual and textual information remains poorly understood. To bridge this gap, we perform a systematic layer-wise masking analysis across multiple architectures, revealing how visual-text fusion evolves within MLLMs. The results show that fusion emerges at several specific layers rather than being uniformly distributed across the network, and certain models exhibit a late-stage "review" phenomenon where visual signals are reactivated before output generation. Besides, we further analyze layer-wise attention evolution and observe persistent high-attention noise on irrelevant regions, along with gradually increasing attention on text-aligned areas. Guided by these insights, we introduce a training-free contrastive attention framework that models the transformation between early fusion and final layers to highlight meaningful attention shifts. Extensive experiments across various MLLMs and benchmarks validate our analysis and demonstrate that the proposed approach improves multimodal reasoning performance.Code will be released.

Planning-oriented end-to-end driving models show great promise, yet they fundamentally learn statistical correlations instead of true causal relationships. This vulnerability leads to causal confusion, where models exploit dataset biases as shortcuts, critically harming their reliability and safety in complex scenarios. To address this, we introduce CausalVAD, a de-confounding training framework that leverages causal intervention. At its core, we design the sparse causal intervention scheme (SCIS), a lightweight, plug-and-play module to instantiate the backdoor adjustment theory in neural networks. SCIS constructs a dictionary of prototypes representing latent driving contexts. It then uses this dictionary to intervene on the model's sparse vectorized queries. This step actively eliminates spurious associations induced by confounders, thereby eliminating spurious factors from the representations for downstream tasks. Extensive experiments on benchmarks like nuScenes show CausalVAD achieves state-of-the-art planning accuracy and safety. Furthermore, our method demonstrates superior robustness against both data bias and noisy scenarios configured to induce causal confusion.

Post-training quantization (PTQ) enables rapid deployment of deep pretrained models. In the low-bit regime, recent PTQ methods for vision models adopt asymmetric quantization (AsymQ), introducing zero-point offsets to mitigate quantization errors. However, these offsets impose substantial hardware overhead and fail to fully capture the non-symmetric structure of pretrained weight distributions, leaving many quantization levels unused.In this paper, we reveal a hidden symmetry in the pretrained weights: after removing a few sparse outliers, the distribution becomes nearly symmetric.Accordingly, we propose Dense and Additive Sparse Quantization (DASQ), which decomposes the weights into dense and sparse matrices.The dense component captures the symmetric structure around zero, while the sparse component models the removed outliers, and both can be processed in parallel and can be implemented with efficient zero-point-free computation.Experiments on image classification, object detection, and instance segmentation show that DASQ surpasses state-of-the-art PTQ methods with lower BOPs. On an FPGA, DASQ also demonstrates higher accuracy and lower power consumption than AsymQ at comparable throughput.