CVPR 2026 Papers

The video reasoning ability of multimodal large language models (MLLMs) is crucial for downstream tasks like video question answering and temporal grounding. While recent approaches have explored text-based chain-of-thought (CoT) reasoning for MLLMs, these methods often suffer from limited cross-modal interaction and increased hallucination, especially with longer videos or reasoning chains. To address these challenges, we propose Video Intelligence via Tool-Augmented Learning (VITAL), a novel end-to-end agentic video reasoning framework. With a visual toolbox, the model can densely sample new video frames on demand and generate multimodal CoT for precise long video reasoning. We observe that temporal grounding and question answering are mutually beneficial for video understanding tasks. Therefore, we construct two high-quality multi-task video reasoning datasets MTVR-CoT-72k for supervised fine-tuning and MTVR-RL-110k for reinforcement learning. Moreover, we propose a Difficulty-aware Group Relative Policy Optimization algorithm (DGRPO) to mitigate difficulty imbalance in multi-task reinforcement learning. Extensive experiments on eleven challenging video understanding benchmarks demonstrate the advanced reasoning ability of VITAL, outperforming existing methods in video question answering and temporal grounding tasks, especially in long video scenarios. Code is available at https://github.com/zhang9302002/ThinkingWithVideos.

Long video understanding is essential for human-like intelligence, enabling coherent perception and reasoning over extended temporal contexts. While the emerging thinking-with-frames paradigm--which alternates between global temporal reasoning and local frame examination--has advanced the reasoning capabilities of video multi-modal large language models (MLLMs), it suffers from a significant efficiency bottleneck due to the progressively growing and redundant multi-modal context. To address this, we propose SpecTemp, a reinforcement learning-based Speculative Temporal reasoning framework that decouples temporal perception from reasoning via a cooperative dual-model design. In SpecTemp, a lightweight draft MLLM rapidly explores and proposes salient frames from densely sampled temporal regions, while a powerful target MLLM focuses on temporal reasoning and verifies the draft's proposals, iteratively refining its attention until convergence. This design mirrors the collaborative pathways of the human brain, balancing efficiency with accuracy. To support training, we construct the SpecTemp-80K dataset, featuring synchronized dual-level annotations for coarse evidence spans and fine-grained frame-level evidence. Experiments across multiple video understanding benchmarks demonstrate that SpecTemp not only maintains competitive accuracy but also significantly accelerates inference compared with existing thinking-with-frames methods.

Zero-shot unsupervised reinforcement learning (URL) offers a promising direction for building generalist agents capable of generalizing to unseen tasks without additional supervision. Among existing approaches, successor representations (SR) have emerged as a prominent paradigm due to their effectiveness in structured, low-dimensional settings. However, SR methods struggle to scale to high-dimensional visual environments. Through empirical analysis, we identify two key limitations of SR in visual URL: (1) SR objectives often lead to suboptimal representations that attend to dynamics-irrelevant regions, resulting in inaccurate successor measures and degraded task generalization; and (2) these flawed representations hinder SR policies from modeling multi-modal skill-conditioned action distributions and ensuring skill controllability. To address these limitations, we propose Saliency-Guided Representation with Consistency Policy Learning (SRCP), a novel framework that improves zero-shot generalization of SR methods in visual URL. SRCP decouples representation learning from successor training by introducing a saliency-guided dynamics task to capture dynamics-relevant representations, thereby improving successor measure and task generalization. Moreover, it integrates a fast-sampling consistency policy with URL-specific classifier-free guidance and tailored training objectives to improve skill-conditioned policy modeling and controllability. Extensive experiments on 16 tasks across 4 datasets from the ExORL benchmark demonstrate that SRCP achieves state-of-the-art zero-shot generalization in visual URL and is compatible with various SR methods.

Video Temporal Grounding (VTG), the task of localizing video segments from text queries, struggles in open-world settings due to limited dataset scale and semantic diversity, causing performance gaps between common and rare concepts. To overcome these limitations, we introduce OmniVTG, a new large-scale dataset for open-world VTG, coupled with a Self-Correction Chain-of-Thought (CoT) training paradigm designed to enhance the grounding capabilities of Multimodal Large Language Models (MLLMs). Our OmniVTG is constructed via a novel Semantic Coverage Iterative Expansion pipeline, which first identifies gaps in the vocabulary of existing datasets and collects videos that are highly likely to contain these target concepts. For high-quality annotation, we leverage the insight that modern MLLMs excel at dense captioning more than direct grounding and design a caption-centric data engine to prompt MLLMs to generate dense, timestamped descriptions. Beyond the dataset, we observe that simple supervised finetuning (SFT) is insufficient, as a performance gap between rare and common concepts still persists. We find that MLLMs' video understanding ability significantly surpasses their direct grounding ability. Based on this, we propose a Self-Correction Chain-of-Thought (CoT) training paradigm. We train the MLLM to first predict, then use its understanding capabilities to reflect on and refine its own predictions. This capability is instilled via a three-stage pipeline of SFT, CoT finetuning, and reinforcement learning. Extensive experiments show our approach not only excels at open-world grounding in our OmniVTG dataset but also achieves state-of-the-art zero-shot performance on four existing VTG benchmarks. Code is available at https://github.com/oceanflowlab/OmniVTG.

Visual reinforcement learning has achieved remarkable progress in visual control and robotics, but its vulnerability to adversarial perturbations remains underexplored. Most existing black-box attacks focus on vector-based or discrete-action RL, and their effectiveness on image-based continuous control is limited by the large action space and excessive environment queries. We propose SEBA, a sample-efficient framework for black-box adversarial attacks on visual RL agents. SEBA integrates a shadow Q model that estimates cumulative rewards under adversarial conditions, a generative adversarial network that produces visually imperceptible perturbations, and a world model that simulates environment dynamics to reduce real-world queries. Through a two-stage iterative training procedure that alternates between learning the shadow model and refining the generator, SEBA achieves strong attack performance while maintaining efficiency. Experiments on MuJoCo and Atari benchmarks show that SEBA significantly reduces cumulative rewards, preserves visual fidelity, and greatly decreases environment interactions compared to prior black-box and white-box methods. The code is available at https://github.com/tairanhuang/seba online.

Generating accurate glyphs for visual text rendering is essential yet challenging. Existing methods typically enhance text rendering by training on a large amount of high-quality scene text images, but the limited coverage of glyph variations and excessive stylization often compromise glyph accuracy, especially for complex or out-of-domain characters. Some methods leverage reinforcement learning to alleviate this issue, yet their reward models usually depend on text recognition systems that are insensitive to fine-grained glyph errors, so images with incorrect glyphs may still receive high rewards. Inspired by Direct Preference Optimization (DPO), we propose ***GlyphPrinter***, a preference-based text rendering method that eliminates reliance on explicit reward models. However, the standard DPO objective only models overall preference between two samples, which is insufficient for visual text rendering where glyph errors typically occur in localized regions. To address this issue, we construct the ***GlyphCorrector*** dataset with region-level glyph preference annotations and propose ***Region-Grouped DPO*** (***R-GDPO***), a region-based objective that optimizes inter- and intra-sample preferences over annotated regions, substantially enhancing glyph accuracy. Furthermore, we introduce ***Regional Reward Guidance***, an inference strategy that samples from an optimal distribution with controllable glyph accuracy. Extensive experiments demonstrate that the proposed GlyphPrinter outperforms existing methods in glyph accuracy while maintaining a favorable balance between stylization and precision.

This paper proposes the Continual Dual-Critic with Cross-Attention (CD-CCA) framework for visual reinforcement learning to address the plasticity-stability conflict. Our method introduces continual learning techniques into the visual RL architecture, constructing two complementary critics using Continual Backpropagation (CBP) and Elastic Weight Consolidation (EWC) -- one for maintaining representational plasticity for rapid environmental adaptation, and the other for preserving knowledge stability to prevent catastrophic forgetting. Furthermore, we design a cross-attention based fusion mechanism that balances the value estimates from the dual critics according to observation characteristics. Experimental results on DeepMind Control and CARLA benchmarks show that CD-CCA effective mitigates issues of representation drift and policy degradation. Compared to existing visual RL methods, our approach exhibits enhanced robustness and adaptability in non-stationary environments and long-horizon decision-making tasks, providing a new architectural paradigm for the advancement of continual reinforcement learning.

Recently, Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as an effective approach to incentivizing reasoning capability in Large Multimodal Models (LMMs), while the underlying mechanisms behind this post-training paradigm are poorly understood. We begin by exploring how input activations are affected by RLVR through the perspective of logit lens. Our systematic investigations across multiple post-trained LMMs suggest that RLVR shifts low-entropy activations unexpectedly, while high-entropy ones are less affected. We further demonstrate that such phenomena are associated with LMM reasoning by controlled experiments, suggesting a potentially beneficial role of modulating low-entropy activations. To this end, we propose Activation Replay, a novel simple yet effective training-free approach that boosts multimodal reasoning of post-trained LMMs without requiring expensive policy optimization. Our design involves manipulation of visual tokens at test time, replaying low-entropy activations from the input context of base LMMs to regulating the RLVR counterparts. Activation Replay triggers better reasoning across diverse scenarios, including mathematics, o3-like visual agents, and video reasoning. We further show that Activation Replay boosts Pass@K and mitigates narrower reasoning coverage of RLVR. Our design is compared against alternative choices, such as replaying high-entropy activations instead of low-entropy ones, or direct cross-model intervention instead of manipulating input tokens, demonstrating the superiority of our implementation. Codes will be made publicly available.

End-to-end driving frameworks, which directly map raw sensor data to vehicle control commands, have shown remarkable potential. However, their performance often deteriorates in sparse-critical scenarios, where rare but safety-sensitive events occur. To address this problem, we propose Explorer-Expert Reinforcement Learning (EE-RL), a novel end-to-end framework that integrates an RL-based explorer, a fine-tuned vision-language model (VLM)-based expert, and a dual replay buffer. EE-RL adopts a collaborative learning strategy in which an explorer and two experts jointly generate experiences from regular driving scenarios to guide policy learning. As training progresses, one of the VLMs focuses on reasoning about sparse-critical scenarios, thereby enhancing learning efficiency and policy optimization in both scenarios. Additionally, the StateHash algorithm is designed to separately measure RGB image similarity and vehicle kinematic similarity, thereby skipping unnecessary VLM reasoning and enabling denser, more effective expert experience generation. Experiments on the CARLA Leaderboard demonstrate that EE-RL significantly outperforms state-of-the-art (SOTA) baselines, achieving +19.82% and +20.98% improvements in driving and infraction scores on Town03, respectively. The EE-RL further achieves 0% accident probability for red-light running and an average driving score of 80.09 in Town05-06, demonstrating robustness, generalizability, and the capability to handle sparse-critical scenarios.

The Aerial Object Goal Navigation, a challenging frontier in Embodied AI, requires an Unmanned Aerial Vehicle (UAV) agent to autonomously explore, reason, and identify a specific target using only visual perception and language description. However, existing methods struggle with the memorization of complex spatial representations in aerial environments, reliable and interpretable action decision-making, and inefficient exploration and information gathering. To address these challenges, we introduce **APEX** (Aerial Parallel Explorer), a novel hierarchical agent designed for efficient exploration and target acquisition in complex aerial settings. APEX is built upon a modular, three-part architecture: 1) Dynamic Spatio-Semantic Mapping Memory, which leverages the zero-shot capability of a Vision-Language Model (VLM) to dynamically construct high-resolution 3D Attraction, Exploration, and Obstacle maps, serving as an interpretable memory mechanism. 2) Action Decision Module, trained with reinforcement learning, which translates this rich spatial understanding into a fine-grained and robust control policy. 3) Target Grounding Module, which employs an open-vocabulary detector to achieve definitive and generalizable target identification. All these components are integrated into a hierarchical, asynchronous, and parallel framework, effectively bypassing the VLM's inference latency and boosting the agent's proactivity in exploration. Extensive experiments show that APEX outperforms the previous state of the art by +4.2% SR and +2.8% SPL on challenging UAV-ON benchmarks, demonstrating its superior efficiency and the effectiveness of its hierarchical asynchronous design.

Reinforcement learning (RL) provides a principled framework for improving vision-language models (VLMs) on complex reasoning tasks. However, existing RL approaches often depend on human-annotated labels or task-specific heuristics to define verifiable rewards--both costly and limited in scalability. We introduce VisPlay, a self-evolving RL framework that enables VLMs to autonomously improve their reasoning capabilities from massive unlabeled image data. Starting from a single base VLM, VisPlay assigns the model into two interacting roles: an Image-Conditioned Questioner that formulates challenging yet answerable visual questions, and a Multimodal Reasoner that generates silver responses. These roles are jointly trained using Group Relative Policy Optimization (GRPO), which uses diversity and difficulty rewards to balance the difficulty of generated questions with the quality of silver answers. VisPlay scales efficiently across two model families. Trained on Qwen2.5-VL and MiMo-VL, VisPlay achieves consistent improvements in visual reasoning, compositional generalization, and hallucination reduction across eight benchmarks including MM-Vet and MMMU, and establishes a scalable path toward self-evolving multimodal intelligence.

Finetuning Large Vision-Language Models with reinforcement learning has emerged as a promising approach to enhance their capability in object-level grounding. However, existing methods, mainly based on GRPO, assign rewards at the response level. Such sparse reward leads to minimal learning signals when all candidate responses are failed in challenging scenarios.In this work, we propose a group-revision optimisation paradigm that enhances learning on hard cases. It begins with a sampled initial response and generates a set of revised candidates to explore improved grounding outcomes. Inspired by reward shaping, we introduce a consolidation process that quantifies each candidate's improvement over the initial attempt and converts it into informative shaping signals.These signals are used to both refine the reward and modulate the advantage, amplifying the influence of high-quality revisions.Our method achieves consistent gains across referring and reasoning segmentation, REC, and counting benchmarks compared with prior GRPO-based models. Code is available at https://github.com/yyliu01/GroupRevision.

Electromagnetic Inverse Scattering Problems (EISP) seek to reconstruct relative permittivity from scattered fields and are fundamental to applications like medical imaging. This inverse process is inherently ill-posed and highly nonlinear, making it particularly challenging, especially under sparse transmitter setups, e.g., with only one transmitter. While recent machine learning-based approaches have shown promising results, they often rely on time-consuming, case-specific optimization and perform poorly under sparse transmitter setups. To address these limitations, we revisit EISP from a data-driven perspective. The scarcity of transmitters leads to an insufficient amount of measured data, which fails to capture adequate physical information for stable inversion. Accordingly, we propose a fully end-to-end and data-driven framework that predicts the relative permittivity of scatterers from measured fields, leveraging data distribution priors to compensate for the incomplete information from sparse measurements. This design enables data-driven training and feed-forward prediction of relative permittivity while maintaining strong robustness to transmitter sparsity. Extensive experiments show that our method outperforms state-of-the-art approaches in reconstruction accuracy and robustness. Notably, we demonstrate, for the first time, high-quality reconstruction from a single transmitter. This work advances practical electromagnetic imaging by providing a new, cost-effective paradigm to inverse scattering. Code and models are released at https://gomenei.github.io/SingleTX-EISP/.

Multimodal reasoning over long-horizon video is challenging due to the need for precise spatiotemporal fusion and alignment across modalities. While recent methods such as Group Relative Policy Optimization (GRPO) have shown promise in this domain, they suffer from three key limitations: (1) data inefficiency from their on-policy design, (2) a vanishing advantage problem, where identical or near-identical rewards within a group eliminate the learning signal by producing zero-valued advantages, and (3) uniform credit assignment that fails to emphasize critical reasoning steps. We introduce AVATAR (Audio-Video Agent for Alignment and Reasoning, a framework that addresses these limitations through two core components: (1) an off-policy training architecture that improves sample efficiency and resolves vanishing advantages by reusing past experiences with greater reward diversity, and (2) Temporal Advantage Shaping (TAS), a credit assignment strategy that emphasizes early (planning) and late (synthesis) reasoning phases. AVATAR achieves strong performance across various benchmarks, outperforming the Qwen2.5-Omni baseline by +5.4 on MMVU, +4.9 on OmniBench, and +4.5 on Video-Holmes. Furthermore, it surpasses standard GRPO by +3.7 on OmniBench and +1.9 on Video-Holmes, while demonstrating 5 times sample efficiency, requiring 80 fewer generated completions to reach target performance.

While Reinforcement Learning from Human Feedback (RLHF) has become a pivotal paradigm for text-to-image generation, its application to image editing remains largely unexplored. A key bottleneck is the lack of a robust general reward model for all editing tasks. Existing edit reward models usually give overall scores without detailed checks, ignoring different instruction requirements and causing biased rewards. To address this, we propose the Verifier-Based Reasoning Reward Model (RRM), which breaks instructions into verifiable principles, evaluates the edited images against each principle, and aggregates fine-grained scores to reduce hallucinations and provide more interpretable criteria. To address this, we argue the key is to move from a simple scorer to a reasoning verifier. We introduce Edit-R1, a framework to build a chain-of-thought (CoT) verifier-based reasoning reward model (RRM) and leverage it into the downstream editing task. The Edit-RRM breaks instructions into distinct principles, evaluates the edited image against each, and aggregates these checks to provide an interpretable, fine-grained reward. To build such an RRM, we first apply supervised fine-tuning (SFT) as a "cold-start" to generate CoT reward trajectories. Then, we introduce Group Contrastive Preference Optimization (GCPO), a reinforcement learning algorithm that leverages human pairwise preference data to reinforce our pointwise RRM. After building the RRM, we use GRPO to train editing models with this non-differentiable yet powerful reward model. Extensive experiments demonstrate that our Edit-RRM surpasses powerful VLMs such as Seed-1.5-VL and Seed-1.6-VL as an editing-specific reward model, and we observe a clear scaling trend, with performance consistently improving from 3B to 7B parameters. Moreover, Edit-R1 delivers gains to editing models like FLUX.1-kontext, highlighting its effectiveness in enhancing image editing.

We present UniGen-1.5, a unified multimodal large language model (MLLM) for advanced image understanding, generation and editing. Building upon UniGen, we comprehensively enhance the model architecture and training pipeline to strengthen the image understanding and generation capabilities while unlocking strong image editing ability. Especially, we propose a unified Reinforcement Learning (RL) strategy that improves both image generation and image editing jointly via shared reward models. To further enhance image editing performance, we propose a light Edit Instruction Alignment stage that significantly improves the editing instruction comprehension that is essential for the success of the RL training. Experimental results show that UniGen-1.5 demonstrates competitive understanding and generation performance. Specifically, UniGen-1.5 achieves 0.89 and 4.31 overall scores on GenEval and ImgEdit benchmarks that surpass the state-of-the-art models such as BAGEL and reaching performance comparable to proprietary models such as GPT-Image-1.

Reinforcement learning (RL), particularly GRPO, improves image generation quality significantly by comparing the relative performance of images generated within the same group. However, in the later stages of training, the model tends to produce homogenized outputs, lacking creativity and visual diversity, restricting the application scenarios of the model. This issue can be analyzed from both reward modeling and generation dynamics perspectives. First, traditional GRPO relies on single sample quality as the reward signal, driving the model to converge toward a few high reward generation modes while neglecting distribution-level diversity. Second, conventional GRPO regularization neglects the dominant role of early stage denoising in preserving diversity, causing a misaligned regularization budget that limits the achievable quality-diversity trade-off.Motivated by these insights, we revisit the diversity degradation problem from both reward modeling and generation dynamics. At the reward level, we propose a distributional creativity bonus based on semantic grouping. Specifically, we construct a distribution-level representation via spectral clustering over samples generated from the same caption, and adaptively allocate exploratory rewards according to group sizes to encourage the discovery of novel visual modes. At the generation level, we introduce a structure-aware regularization, which enforces stronger early-stage constraints to preserve diversity without compromising reward optimization efficiency. Experiments demonstrate that our method achieves an 13%~18% improvement in semantic diversity under matched quality scores, establishing a new Pareto frontier between image quality and diversity for GRPO-based image generation.

Recent advances in 3D Gaussian Splatting (3DGS) enable photorealistic real-time rendering but also increase the risks of unauthorized copying and redistribution. Existing 3DGS watermarking methods typically rely on handcrafted thresholds or globally fixed hyperparameters to balance invisibility and robustness, making their embedding behavior static and scene-agnostic. We instead formulate 3DGS watermarking as a goal-directed decision process and introduce Write Where It Matters (W2M), the first reinforcement learning-based framework that adaptively learns where and how much to embed. By modeling the embedding process as a Markov Decision Process, W2M uses a lightweight policy network to allocate precise Gaussian updates directly from immediate reward feedback, iteratively. The reward incentivizes both rendering-space invisibility and decoding robustness under various image- and model-level distortions. To achieve efficiency, W2M operates on a structured 3DGS backbone organized around learnable anchors and applies policy-guided per-anchor gradient scaling. Extensive experiments across the Blender, LLFF, and Mip-NeRF 360 datasets demonstrate that W2M achieves state-of-the-art bit accuracy, strong perceptual fidelity, and structural consistency under both standard and adversarial conditions.

Recent progress in Multimodal Large Language Models (MLLMs) demonstrates that Chain-of-Thought (CoT) reasoning enables systematic solutions to complex understanding tasks. However, its extension to generation tasks remains nascent and limited by scenario-specific mechanisms that hinder generalization and adaptation. In this work, we present ThinkGen, the first think-driven visual generation framework that explicitly leverages MLLM's CoT reasoning in various generation scenarios. ThinkGen employs a decoupled architecture comprising a pretrained MLLM and a Diffusion Transformer (DiT), wherein the MLLM generates tailored instructions based on user intent, and DiT produces high-quality images guided by these instructions. We further propose a separable GRPO-based training paradigm (SepGRPO), alternating reinforcement learning between the MLLM and DiT modules. This flexible design enables joint training across diverse datasets, facilitating effective CoT reasoning for a wide range of generative scenarios. Extensive experiments demonstrate that ThinkGen achieves robust, state-of-the-art performance across multiple generation benchmarks.