CVPR 2026 论文列表

Recent advancements in reinforcement learning with verifiable rewards (RLVR) have significantly improved the complex reasoning ability of vision-language models (VLMs). However, its outcome-level supervision is too coarse to diagnose and correct errors within the reasoning chain. To this end, we propose Perceval, a process reward model (PRM) that enables token-level error grounding, which can extract image-related claims from the response and compare them one by one with the visual evidence in the image, ultimately returning claims that contain perceptual errors. Perceval is trained with perception-intensive supervised training data. We then integrate Perceval into the RL training process to train the policy models. Specifically, compared to traditional GRPO, which applies sequence-level advantages, we apply token-level advantages by targeting penalties on hallucinated spans identified by Perceval, thus enabling fine-grained supervision signals. In addition to augmenting the training process, Perceval can also assist VLMs during the inference stage. Using Perceval, we can truncate the erroneous portions of the model's response, and then either have the model regenerate the response directly or induce the model to reflect on its previous output. This process can be repeated multiple times to achieve test-time scaling. Experiments show significant improvements on benchmarks from various domains across multiple reasoning VLMs trained with RL. For test-time scaling, it also demonstrates consistent performance gains over other strategies, such as major voting. Our code and data will be publicly released.

Effective human behavior modeling requires a representation of the human body movement that capitalizes on its compositionality. We propose a hierarchical representation consisting of Action Atoms that capture the atomic joint movements and Action Motifs that are formed by their temporal compositions and encode similar body movements found across different overall human actions. We derive A4Mer, a nested latent Transformer to learn this hierarchical representation from human pose data in a fully self-supervised manner. A4Mer splits a 3D pose sequence into variable-length segments and represents each segment as a single latent token (Action Atoms). Through bottom-up representation learning, temporal patterns composed of these Action Atoms, which capture meaningful temporal spans of reusable, semantic segments of body movements, naturally emerge (Action Motifs). A4Mer achieves this with a unified pretext task of masked token prediction in their respective latent spaces. We also introduce Action Motif Dataset (AMD), a large-scale dataset of multi-view human behavior videos with full SMPL annotations. We introduce a novel use of cameras by mounting them on the feet to achieve their frame-wise annotations despite frequent and heavy body occlusions. Experimental results demonstrate the effectiveness of A4Mer for extracting meaningful Action Motifs, which significantly benefit human behavior modeling tasks including action recognition, motion prediction, and motion interpolation.

Vision-Language Models (VLMs) have achieved remarkable progress in multimodal understanding, yet their positional encoding mechanisms remain suboptimal. Existing approaches uniformly assign positional indices to all tokens, overlooking variations in information density within and across modalities, which leads to inefficient attention allocation where redundant visual regions dominate while informative content is underrepresented. We identify positional granularity as an implicit resource and propose MODIX (Multimodal Information-Driven Positional IndeX Scaling), a training-free framework that dynamically adapts positional strides based on modality-specific contributions. MODIX jointly models intra-modal density via covariance-based entropy and inter-modal interaction via cross-modal alignment to derive unified scores, which rescale positional indices to allocate finer granularity to informative modalities while compressing redundant ones, without requiring any modification to model parameters or architecture. Experiments across diverse architectures and benchmarks demonstrate that MODIX consistently improves multimodal reasoning and adaptively reallocates attention according to task-dependent information distributions, suggesting that positional encoding should be treated as an adaptive resource in Transformers for multimodal sequence modeling.

Conventional vision-language models (VLMs) struggle to interpret scenes captured under adverse conditions (e.g., low light, high dynamic range, or fast motion) because standard RGB images degrade in such environments. Event cameras provide a complementary modality: they asynchronously record per-pixel brightness changes with high temporal resolution and wide dynamic range, preserving motion cues where frames fail. We propose RE-VLM, the first dual-stream vision-language model that jointly leverages RGB images and event streams for robust scene understanding across both normal and challenging conditions. RE-VLM employs parallel RGB and event encoders together with a progressive training strategy that aligns heterogeneous visual features with language. To address the scarcity of RGB-Event-Text supervision, we further propose a graph-driven pipeline that converts synchronized RGB-Event streams into verifiable scene graphs, from which we synthesize captions and question-answer (QA) pairs. To develop and evaluate RE-VLM, we construct two datasets: PEOD-Chat, targeting illumination-challenged scenes, and RGBE-Chat, covering diverse scenarios. On captioning and VQA benchmarks, RE-VLM consistently outperforms state-of-the-art RGB-only and event-only models with comparable parameter counts, with particularly large gains under challenging conditions. These results demonstrate the effectiveness of event-augmented VLMs in achieving robust vision-language understanding across a wide range of real-world environments. Code and datasets are available at https://github.com/bupt-ai-cz/RE-VLM.

Training vision-language models (VLMs) for complex reasoning remains a challenging task, i.a. due to the scarcity of high-quality image-text reasoning data. Conversely, text-based reasoning resources are abundant and scalable, but it is still an open question how to leveraging them for VLM reasoning. To address this problem, we propose VOLD, a framework to transfer reasoning capabilities from text-only teacher models to VLM student models. To this end, VOLD combines reinforcement learning via Group Relative Policy Optimization (GRPO) with on-policy distillation, which allows the student reasoning traces to be guided by the teacher model, resulting in a significant gain over using GRPO alone.We further show that a cold-start alignment is essential for an effective transfer during the online training phase in this scenario and that without sufficient distributional alignment between teacher and student, on-policy distillation fails to provide meaningful guidance.We evaluate VOLD across diverse benchmarks including MMMU-Pro, MathVision, MathVista, and LogicVista, showing that VOLD outperforms the baseline model significantly and improves over the state of the art by a margin.Our ablation shows the importance of a cold-start alignment via SFT for on-policy distillation with a text-only teacher

Vision-Language-Action (VLA) models built upon Chain-of-Thought (CoT) have achieved remarkable success in advancing general-purpose robotic agents, owing to its significant perceptual comprehension. Recently, since text-only CoT struggles to adequately capture scene details in complex spatial environments, a highly promising strategy involves leveraging visual priors to guide robotic action generation. Nevertheless, these strategies face two inherent challenges: (i) a modality gap between visual observations and low-level actions, and (ii) unstable training due to competing objectives between visual prediction and action generation. To address these challenges, we propose a Vision-Integrated Trajectory Alignment (VITA) framework that learns a shared discrete latent space for vision and action, enabling joint modeling of perception and motor control. VITA introduces a implicit visual CoT: autoregressively generated tokens is simultaneously decoded into future frames predictions and robot actions, thereby internalizing visual dynamics as an inductive bias for motion planning. Extensive experiments on simulated and real-world environments demonstrate state-of-the-art performance. VITA improves 14.5%, 9.6% and 12.1% over existing baselines on CALVIN, LIBERO and SimplerEnv. Furthermore, VITA attains an average success rate of 80.5% across six real-world tasks, demonstrating its potential as a generalist robotic manipulation model.

What does it mean for a visual system to truly understand affordance? We argue that this understanding hinges on two complementary capacities: geometric perception, which identifies the structural parts of objects that enable interaction, and interaction perception, which models how an agent's actions engage with those parts. To test this hypothesis, we conduct a systematic probing of Visual Foundation Models (VFMs). We find that models like DINO inherently encode part-level geometric structures, while generative models like Flux contain rich, verb-conditioned spatial attention maps that serve as implicit interaction priors. Crucially, we demonstrate that these two dimensions are not merely correlated but are composable elements of affordance. By simply fusing DINO's geometric prototypes with Flux's interaction maps in a training-free, zero-shot manner, we achieve affordance estimation competitive with weakly-supervised methods. This final fusion experiment confirms that geometric and interaction perception are the fundamental building blocks of affordance understanding in VFMs, providing a mechanistic account of how perception grounds action.

Large vision-language models (VLMs) are highly vulnerable to multimodal jailbreak attacks that exploit visual-textual interactions to bypass safety guardrails. In this paper, we present DTR, a novel inference-time defense that mitigates multimodal jailbreak attacks through optimizing the model's key-value (KV) caches. Rather than relying on curated safety-specific data or costly image-to-text conversion, we introduce a new formulation of the safety-relevant distributional shift induced by the visual modality. This formulation enables DTR to dynamically adjust visual token weights, minimizing the impact of adversarial visual inputs while preserving the model's general capabilities and inference efficiency. Extensive evaluation across diverse VLMs and attack benchmarks demonstrates that DTR outperforms existing defenses in both attack robustness and benign-task performance, marking the first successful application of KV cache optimization for safety enhancement in multimodal foundation models. The code for replicating DTR is available at: https://github.com/TanqiuJiang/DTR.

Multi-Label Recognition (MLR) based on Vision-Language Models (VLMs) aims to leverage their pre-trained knowledge to better adapt complex recognition scenarios, thereby enhancing model robustness. However, for realistic decentralized applications requiring federated learning, adapting VLMs to each client that possesses private and heterogeneous data can cause the model to overfit spurious label correlations, consequently triggering irrelevant categories when encountering new samples. To tackle this problem, we reconsider the federated learning for MLR with a causal model, in which we adopt a front-door adjustment and decouple the MLR modeling process by intermediate variables that magnify the oracle label co-occurrence. Guided by our analysis, we propose our FedMPT, the first method specifically designed for federated MLR. The core idea of FedMPT is to leverage generalizable conditions to steer federated MLR to mitigate erroneous label activations. To achieve this, FedMPT introduces an Large Language Model (LLM)-driven pipeline to decipher the underlying conditions that govern label dependencies. Furthermore, we introduce an optimal transport between the condition-enriched prompts and the image patches to uncover multiple region-level semantics. Finally, we generate synergistic predictions from different conditions with a crafted gating mechanism. Experiments on multiple benchmark datasets show that our proposed approach achieves competitive results and outperforms SOTA methods under varied settings.

Understanding the physical world is essential for generalist AI agents. However, it remains unclear whether state-of-the-art vision perception models (e.g., large VLMs) can perform quantitative physical reasoning tasks. Existing evaluations are predominantly VQA-based and qualitative, offering limited insight into whether these models can infer the kinematic quantities of moving objects from videos. To address this, we present QuantiPhy, the first benchmark designed to quantitatively measure a VLM's physical reasoning ability. Comprising more than 3.3K video-text instances with numerical ground truth, QuantiPhy evaluates a VLM's performance on estimating an object's size, velocity, and acceleration at a given timestamp, using one of these properties as an input prior. The benchmark standardizes prompts and scoring to assess numerical accuracy, enabling fair comparisons across models. Our experiments on state-of-the-art VLMs reveal a consistent gap between their qualitative plausibility and actual numerical correctness. We further provide an in-depth analysis of key factors like background noise, counterfactual priors, and strategic prompting and find that state-of-the-art VLMs lean heavily on pre-trained world knowledge rather than faithfully using the provided visual and textual inputs when reasoning about objects' kinematic properties. QuantiPhy offers the first rigorous, scalable testbed to move VLMs beyond mere verbal plausibility toward a numerically grounded physical understanding.

Spatial intelligence in vision-language models (VLMs) attracts research interest with the practical demand to reason in the 3D world. Despite promising results, most existing methods follow the conventional 2D pipeline in VLMs and use pixel-aligned representations for the vision modality. However, correspondence-based models with implicit 3D scene understanding often fail to achieve spatial consistency, and representation-based models with 3D geometric priors lack efficiency in vision sequence serialization. To address this, we propose a Proxy3D method with compact yet comprehensive 3D proxy representations for the vision modality. Given only video frames as input, we employ semantic and geometric encoders to extract scene features and then perform their semantic-aware clustering to obtain a set of proxies in the 3D space. For representation alignment, we further curate the SpaceSpan dataset and apply multi-stage training to adopt the proposed 3D proxy representations with the VLM. When using shorter sequences for vision information, our method achieves competitive or state-of-the-art performance in 3D visual question answering, visual grounding and general spatial intelligence benchmarks.

The rapid advancement of Vision-Language Models (VLMs) has brought their safety vulnerabilities into sharp focus. However, existing red teaming methods are fundamentally constrained by an inherent linear exploration paradigm, confining them to optimizing within a predefined strategy set and preventing the discovery of novel, diverse exploits. To transcend this limitation, we introduce TreeTeaming, an automated red teaming framework that reframes strategy exploration from static testing to a dynamic, evolutionary discovery process. At its core lies a strategic Orchestrator, powered by a Large Language Model (LLM), which autonomously decides whether to evolve promising attack paths or explore diverse strategic branches, thereby dynamically constructing and expanding a strategy tree. A multimodal actuator is then tasked with executing these complex strategies. In the experiments across 12 prominent VLMs, TreeTeaming achieves state-of-the-art attack success rates on 11 models, outperforming existing methods and reaching up to 87.60% on GPT-4o. The framework also demonstrates superior strategic diversity over the union of previously public jailbreak strategies. Furthermore, the generated attacks exhibit an average toxicity reduction of 23.09%, showcasing their stealth and subtlety. Our work introduces a new paradigm for automated vulnerability discovery, underscoring the necessity of proactive exploration beyond static heuristics to secure frontier AI models. The code and data are available at: https://github.com/ChunXiaostudy/TreeTeaming. Warning: This paper contains examples of harmful texts and images, and reader discretion is recommended.

Recent progress in deep learning has been driven by increasingly large-scale models, but the resulting computational cost has become a critical bottleneck. Sparse Mixture of Experts (MoE) offers an effective solution by activating only a small subset of experts for each input, achieving high scalability without sacrificing inference speed. Although effective, sparse MoE training exhibits characteristic optimization difficulties. Because the router receives informative gradients only through the experts selected in the forward pass, it suffers from gradient blocking and obtains little information from unselected routes. This limited, highly localized feedback makes it difficult for the router to learn appropriate expert-selection scores and often leads to unstable routing dynamics, such as fluctuating expert assignments during training. To address this issue, we propose TGR-MoE: Teacher-Guided Routing for Sparse Vision Mixture-of-Experts, a simple yet effective method that stabilizes router learning using supervision derived from a pretrained dense teacher model. TGR-MoE constructs a teacher router from the teacher's intermediate representations and uses its routing outputs as pseudo-supervision for the student router, suppressing frequent routing fluctuations during training and enabling knowledge-guided expert selection from the early stages of training. Extensive experiments on ImageNet-1K and CIFAR-100 demonstrate that TGR consistently improves both accuracy and routing consistency, while maintaining stable training even under highly sparse configurations.

We present DrivoR, a simple transformer-based architecture for end-to-end autonomous driving. Our approach builds on pretrained Vision Transformers (ViTs) and introduces camera-aware register tokens that compress multi-camera features into a compact scene representation, significantly reducing downstream computation without sacrificing accuracy. These tokens drive two lightweight transformer decoders that generate and then score candidate trajectories. The scoring decoder learns to mimic an oracle and predicts interpretable sub-scores e.g., safety or efficiency, enabling behavior-conditioned driving at inference. Despite its minimal design, DrivoR outperforms or matches strong baselines across NAVSIM-v1/v2, and closed-loop HUGSIM benchmarks. Our results show that a pure-transformer architecture, combined with targeted token compression, is sufficient for accurate, efficient, and adaptive end-to-end driving. Code and checkpoints are available via the project page.

Vision-language models like CLIP have achieved remarkable progress in cross-modal representation learning, yet suffer from systematic misclassifications among visually and semantically similar categories. We observe that such confusion patterns are not random but persistently occur between specific category pairs, revealing the model's intrinsic bias and limited fine-grained discriminative ability. To address this, we propose CAPT, a Confusion-Aware Prompt Tuning framework that enables models to learn from their own misalignment. Specifically, we construct a Confusion Bank to explicitly model stable confusion relationships across categories and misclassified samples. On this basis, we introduce a Semantic Confusion Miner (SEM) to capture global inter-class confusion through semantic difference and commonality prompts, and a Sample Confusion Miner (SAM) to retrieve representative misclassified instances from the bank and capture sample-level cues through a Diff-Manner Adapter that integrates global and local contexts. To further unify confusion information across different granularities, a Multi-Granularity Difference Expert (MGDE) module is designed to jointly leverage semantic- and sample-level experts for more robust confusion-aware reasoning. Extensive experiments on 11 benchmark datasets demonstrate that our method significantly reduces confusion-induced errors while enhancing the discriminability and generalization of both base and novel classes, successfully resolving 50.72 percent of confusable sample pairs. Code will be released at https://github.com/greatest-gourmet/CAPT.

Vision-Language-Action (VLA) models have emerged as a powerful paradigm for general robotic manipulation. However, existing approaches typically omit intermediate reasoning steps and directly regress actions, limiting reasoning interpretability and performance in long-horizon or compositional tasks. Although recent studies introduce Chain-of-Thought (CoT) reasoning into VLA models, their effectiveness remains suboptimal due to two key issues: (1) generating a full reasoning trajectory at every timestep introduces substantial redundancy, thereby hinders real-time deployment and (2) reasoning is performed independently, neglecting temporal consistency, which leads to planning conflicts.We propose TRM-VLA, a temporal-aware reasoning and memorization framework that integrates explicit temporal modeling into the VLA reasoning process. TRM-VLA consists of two core components: (1) Keyframe-Triggered Reasoning (KTR), which identifies task progress and performs hierarchical CoT reasoning only at key decision points to reduce redundant inference; and (2) Granularity-adaptable Context Memory (GCM), which dynamically stores and retrieves historical reasoning trajectories to maintain inter-frame coherence and global context. Built upon a dual-system architecture--combining a multimodal foundation model for slow reasoning (System 2) with a diffusion-based policy for fast execution (System 1)--TRM-VLA learns to plan and act efficiently in a unified manner. Extensive experiments on LIBERO-90, SIMPLER, and four real-world robotic tasks demonstrate that TRM-VLA achieves state-of-the-art performance while improving reasoning efficiency.

Existing humanoid control systems often rely on teleoperation or modular generation pipelines that separate language understanding from physical execution. However, the former is entirely human-driven, and the latter lacks tight alignment between language commands and physical behaviors. In this paper, we present SENTINEL, a fully end-to-end language-action framework for humanoid whole-body control. We construct a large-scale dataset by tracking human motions in simulation using a pretrained whole body controller, combined with their text annotations. The model directly maps language commands and proprioceptive inputs to low-level actions without any intermediate representation. The model generates action chunks using flow matching, which can be subsequently refined by a residual action head for real-world deployment. Our method exhibits strong semantic understanding and stable execution on humanoid robots in both simulation and real-world deployment, and also supports multi-modal extensions by converting inputs into texts.

Vision-language models (VLMs) struggle with 3D-related tasks such as spatial cognition and physical understanding, which are crucial for real-world applications like robotics and embodied agents. We attribute this to a modality gap between the 3D tasks and the 2D training of VLM, which led to inefficient retrieval of 3D information from 2D input. To bridge this gap, we introduce SandboxVLM, a simple yet effective framework that leverages abstract bounding boxes to encode geometric structure for VLMs. Specifically, we design a 3D Sandbox reconstruction and perception pipeline comprising four stages: generating multi-view priors with abstract control, proxy elevation, multi-view voting and clustering, and 3D-aware reasoning. Evaluated in zero-shot settings across multiple benchmarks and VLM backbones, our approach consistently improves spatial intelligence, achieving an 8.3% gain on SAT Real compared with baseline methods for instance. These results demonstrate that presenting VLMs with 3D abstractions substantially enhances their 3D reasoning ability without additional training, suggesting new possibilities for general-purpose embodied intelligence.

Humans often resolve visual uncertainty by comparing an image with relevant examples, but ViTs lack the ability to identify which examples would improve their predictions. We present Task-Aligned Context Selection (TACS), a framework that learns to select paired examples which truly improve task performance rather than those that merely appear similar. TACS jointly trains a selector network with the task model through a hybrid optimization scheme combining gradient-based supervision and reinforcement learning, making retrieval part of the learning objective. By aligning selection with task rewards, TACS enables discriminative models to discover which contextual examples genuinely help. Across 18 datasets covering fine-grained recognition, medical image classification, and medical image segmentation, TACS consistently outperforms similarity-based retrieval, particularly in challenging or data-limited settings.