CVPR 2026 Papers

X-ray computed tomography (CT) suffers from severe metal artifacts when high-attenuation objects such as dental fillings or orthopedic implants are present. These artifacts originate from the polychromatic nature of X-rays, where attenuation varies strongly with photon energy and material composition, breaking the monochromatic assumption used by conventional reconstruction algorithms. Recent neural rendering approaches attempt to address this mismatch through differentiable polychromatic projection models, but they still struggle with smoothness bias, loss of fine structures, and prohibitive computation when extended to large-scale cone-beam CT. We introduce a splat-based metal artifact reduction framework that incorporates a physically grounded polychromatic forward model into a continuous Gaussian representation for cone-beam CT. Each Gaussian encodes the energy-dependent attenuation of the underlying material using a compact material parameterization, which enables efficient joint optimization of geometric and material properties without relying on a metal mask. This compact attenuation formulation captures the essential variation across biological tissues and metallic implants, allowing our model to explain metal-induced nonlinearity while preserving high-frequency structure. Experiments on simulated and real cone-beam CT scans show that our method converges significantly faster and suppresses metal artifacts more effectively than existing reconstruction and neural field-based approaches.

While recent video generation models achieve impressive visual quality, generating physically plausible videos remains challenging, especially for fluid dynamics and rigid-body motions. To address this, we present **NS-Diff**, a physics-guided reinforcement learning framework for video diffusion. First, we design a noise-robust physical dynamics detector that distinguishes rigid and fluid regions by analyzing motion in noisy latent frames. Second, we introduce a Physics-Conditioned Latent Injection module, which encodes velocity fields, deformation gradients, and material masks, and injects them into the DiT denoiser via cross-attention. Third, we introduce a reinforcement learning optimization module that enforces simplified Navier-Stokes constraints on fluid dynamics and minimum-jerk principles on rigid bodies through policy gradients. Experiments on PhysVideoBench, UCF, and MSR-VTT show that our approach reduces jerk errors by 43%, decreases fluid divergence by 33%, and improves FVD by 22.7%, achieving higher physical plausibility and visual quality.

Text-to-image (T2I) generative models such as Stable Diffusion and FLUX can synthesize realistic, high-quality images directly from textual prompts. The resulting image quality depends critically on well-crafted prompts that specify both subjects and stylistic modifiers, which have become valuable digital assets. However, the rising value and ubiquity of high-quality prompts expose them to security and intellectual-property risks. One key threat is the prompt stealing attack, i.e., the task of recovering the textual prompt that generated a given image. Prompt stealing enables unauthorized extraction and reuse of carefully engineered prompts, yet it can also support beneficial applications such as data attribution, model provenance analysis, and watermarking validation. Existing approaches often assume white-box gradient access, require large-scale labeled datasets for supervised training, or rely solely on captioning without explicit optimization, limiting their practicality and adaptability. To address these challenges, we propose PROMPTMINER, a black-box prompt stealing framework that decouples the task into two phases: (1) a reinforcement learning-based optimization phase to reconstruct the primary subject, and (2) a VLM-guided search phase to recover stylistic modifiers. Experiments across multiple datasets and diffusion backbones demonstrate that PROMPTMINER achieves superior results, with CLIP similarity up to 0.958 and textual alignment with SBERT up to 0.751, surpassing all baselines. Even when applied to in-the-wild images with unknown generators, it outperforms the strongest baseline by 7.5% in CLIP similarity, demonstrating better generalization. Finally, PROMPTMINER maintains strong performance under defensive perturbations, highlighting remarkable robustness.

Subject-Driven Text-to-Image (T2I) Generation aims to preserve a subject's identity while editing its context based on a text prompt. A core challenge in this task is the "similarity-controllability paradox", where enhancing textual control often degrades the subject's fidelity, and vice-versa. We argue this paradox stems from the ambiguous role of text prompts, which are often tasked with describing both the subject and the desired modifications, leading to conflicting signals for the model. To resolve this, we propose DisCo, a novel framework that first Disentangles and then re-Couples visual and textual information. First, our textual-visual decoupling module isolates the sources of information: subject identity is extracted exclusively from the reference image with the entity word of the subject, while the text prompt is simplified to contain only the modification command, where the subject refers to general pronouns, eliminating descriptive ambiguity. However, this strict separation can lead to unnatural compositions between the subject and its contexts. We address this by designing a dedicated reward signal and using reinforcement learning to seamlessly recouple the visually-defined subject and the textually-generated context. Our approach effectively resolves the paradox, enabling simultaneous high-fidelity subject preservation and precise textual control. Extensive experiments demonstrate that our method achieves state-of-the-art performance, producing highly realistic and coherent images.

Reinforcement Learning with Verifiable Reward (RLVR) has substantially advanced the video understanding capabilities of Multimodal Large Language Models (MLLMs). However, the rapid progress of MLLMs is outpacing the complexity of existing video datasets, while the manual annotation of new, high-quality data remains prohibitively expensive.This work investigates a pivotal question: Can the rich, intrinsic information within videos be harnessed to self-generate high-quality, verifiable training data?To investigate this problem, we first introduce three self-supervised pretext tasks for video understanding: Anomaly Grounding, Object Counting, and Temporal Jigsaw. To validate the difficulty of these tasks, we construct the Video Intrinsic Understanding Benchmark (VIUBench), revealing that current state-of-the-art MLLMs struggle significantly on these tasks. Building upon these pretext tasks, we develop the VideoSSR-30K dataset and propose VideoSSR, a novel video self-supervised reinforcement learning framework for RLVR. Extensive experiments across 17 benchmarks, spanning four major video domains (General Video QA, Long Video QA, Temporal Grounding, and Complex Reasoning), demonstrate that our VideoSSR consistently enhances model performance, yielding an average improvement of over 5%. These results establish VideoSSR as a potent foundational framework for developing more advanced video understanding in MLLMs.

Recently, GRPO-based reinforcement learning has shown remarkable progress in optimizing flow-matching models, effectively improving their alignment with task-specific rewards. Within these frameworks, the policy update relies on importance-ratio clipping to constrain overconfident positive and negative gradients. However, in practice, we observe a systematic shift in the importance-ratio distribution--its mean falls below 1 and its variance differs substantially across timesteps. This left-shifted and inconsistent distribution prevents positive-advantage samples from entering the clipped region, causing the mechanism to fail in constraining overconfident positive updates. As a result, the policy model inevitably enters an **implicit over-optimization stage**--while the proxy reward continues to increase, essential metrics such as image quality and text-prompt alignment deteriorate sharply, ultimately making the learned policy impractical for real-world use. To address this issue, we introduce **GRPO-Guard**, a simple yet effective enhancement to existing GRPO frameworks. Our method incorporates ratio normalization, which restores a balanced and step-consistent importance ratio, ensuring that PPO clipping properly constrains harmful updates across denoising timesteps. In addition, a gradient reweighting strategy equalizes policy gradients over noise conditions, preventing excessive updates from particular timestep regions. Together, these designs act as a regulated clipping mechanism, stabilizing optimization and substantially mitigating implicit over-optimization without relying on heavy KL regularization. Extensive experiments on multiple diffusion backbones (e.g., SD3.5M, Flux.1-dev) and diverse proxy tasks demonstrate that GRPO-Guard significantly reduces over-optimization while maintaining or even improving generation quality. We provide detailed demonstrations of the over-optimization process and corresponding visualizations in https://jingw193.github.io/GRPO-Guard/.

Charts are high-density visual carriers of complex data and medium for information extraction and analysis. Due to the need for precise and complex visual reasoning, automated chart understanding poses a significant challenge to existing Multimodal Large Language Models (MLLMs). Many MLLMs trained with reinforcement learning (RL) face the challenge of credit assignment. Their advantage estimation, typically performed at the trajectory level, cannot distinguish between correct and incorrect reasoning steps within a single generated response. To address this limitation, we introduce SketchVL, a novel MLLM that optimized with FinePO, a new RL algorithm designed for fine-grained credit assignment within each trajectory. SketchVL's methodology involves drawing its intermediate reasoning steps as markers on the image and feeding the annotated image back to itself, creating a robust, multi-step reasoning process. During training, the FinePO algorithm leverages a Fine-grained Process Reward Model (FinePRM) to score each drawing action within a trajectory, thereby precisely assigning credit for each step. This mechanism allows FinePO to more strongly reward correct tokens when a trajectory is globally successful, and more heavily penalize incorrect tokens when the trajectory is globally suboptimal, thus achieving fine-grained reinforcement signals. Experiments show that SketchVL learns to align its step-level behavior with the FinePRM, achieving an average performance gain of 7.23% over its base model across chart datasets, natural image datasets, and mathematics, providing a promising new direction for training powerful reasoning models.

Most of the multi-agent video understanding frameworks adopt static and non-learnable tool invocation mechanisms, which limit the discovery of diverse clues essential for robust perception and reasoning regarding temporally or spatially complex videos. To address this challenge, we propose a novel Multi-agent system for video understanding, namely VideoChat-M1. Instead of using a single or fixed policy, we adopt a distinct Collaborative Policy Planning (CPP) paradigm with multiple policy agents, which comprises three key processes. (1)Policy Generation: Each agent generates its unique tool invocation policy tailored to the user's query. (2) Policy Execution: Each agent sequentially invokes relevant tools to execute its policy and explore the video content. (3) Policy Communication: During the intermediate stages of policy execution, agents interact with one another to update their respective policies. Through this collaborative framework, all agents work in tandem, dynamically refining their preferred policies based on contextual insights from peers. Moreover, we equip our CPP paradigm with Multi-Agent Reinforcement Learning (MARL). Consequently, policy agents can be jointly optimized to enhance the performance, guided by both the final answer reward and intermediate collaborative process feedback. Extensive experiments demonstrate that VideoChat-M1 achieves SOTA performance across eight benchmarks on four tasks. Notably, on LongVideoBench, our method outperforms Gemini 2.5 pro by 3.6% and GPT-4o by 15.6%.

Video-based spatial reasoning -- such as estimating distances, judging directions, or understanding layouts from multiple views -- requires selecting informative frames and, when needed, actively seeking additional viewpoints during inference. Existing multimodal large language models (MLLMs) consume a fixed set of uniformly sampled frames and cannot request new views once reasoning begins, often missing the geometric cues necessary for reliable spatial judgments. We present EagleVision, a dual-stage framework that combines geometry-aware frame selection with active, Bird's-Eye-View (BEV)-grounded reasoning. In the first stage (macro perception), a semantics-perspective-fusion determinantal point process (SPF-DPP) selects a compact set of keyframes that jointly maximize semantic relevance and viewpoint diversity under a fixed token budget. In the second stage (micro verification), the model performs iterative spatial Chain-of-Thought: at each step it can either reason in text or predict a pose on the BEV plane to retrieve the nearest real frame, forming a closed-loop hypothesize-look-verify cycle. The querying policy is trained purely via reinforcement learning with a spatial grounding reward, requiring no human-annotated reasoning traces. On VSI-Bench and SQA3D, EagleVision achieves state-of-the-art performance among open-source vision-language models.

Reasoning segmentation increasingly employs reinforcement learning to generate explanatory reasoning chains that guide Multimodal Large Language Models. While these geometric rewards are primarily confined to guiding the final localization, they are incapable of discriminating whether the reasoning process remains anchored on the referred region or strays into irrelevant context. Lacking this discriminative guidance, the model's reasoning often devolves into unfocused and verbose chains that ultimately fail to disambiguate and perceive the target in complex scenes. This suggests a need to complement the RL objective with Discriminative Perception, an ability to actively distinguish a target from its context. To realize this, we propose DPAD to compel the model to generate a descriptive caption of the referred object, which is then used to explicitly discriminate by contrasting the caption's semantic relevance to the referred object against the wider context. By optimizing for this discriminative capability, the model is forced to focus on the unique attributes of the target, leading to a more converged and efficient reasoning chain. The descriptive caption also serves as an interpretability rationale that aligns with the segmentation. Experiments on the benchmarks confirm the validity of our approach, delivering substantial performance gains, with the cIoU on ReasonSeg increasing by 3.09% and the reasoning chain length decreasing by approximately 42%. Code is available at https://github.com/mrazhou/DPAD.

Reinforcement learning (RL) has demonstrated significant potential for post-training language models and autoregressive visual generative models, but adapting RL to masked generative models (MGMs) remains challenging. The core factor is that policy optimization requires the probability likelihood of each step due to its multi-step iteration process. This reliance on entire sampling trajectories introduces high computational cost, whereas natively optimizing random steps often yields suboptimal results. In this paper, we present MaskFocus, a novel RL framework that achieves effective policy optimization for MGMs by focusing on critical steps. Specifically, we determine the step-level information gain by measuring the similarity between the intermediate images at each sampling step and the final generated image. Crucially, we leverage this to identify the most critical and valuable steps and execute policy optimization on them. Furthermore, we design a dynamic routing sampling based on entropy to encourage the model to explore more valuable masking strategies. Extensive experiments on multiple Text-to-Image benchmarks validate the effectiveness of our method.

Exploration capacity shapes both inference-time performance and reinforcement learning (RL) training for large (vision-) language models, as stochastic sampling often yields redundant reasoning paths with little high-level diversity. This paper proposes Reasoning Palette, a novel latent-modulation framework that endows the model with a stochastic latent variable for strategic contextualization, guiding its internal planning prior to token generation. This latent context is inferred from the mean-pooled embedding of a question-answer pair via a variational autoencoder (VAE), where each sampled latent potentially encodes a distinct reasoning context. During inference, a sampled latent is decoded into learnable token prefixes and prepended to the input prompt, modulating the model's internal reasoning trajectory. In this way, the model performs internal sampling over reasoning strategies prior to output generation, which shapes the style and structure of the entire response sequence. A brief supervised fine-tuning (SFT) warm-up phase allows the model to adapt to this latent conditioning. Within RL optimization, Reasoning Palette facilitates structured exploration by enabling on-demand injection for diverse reasoning modes, significantly enhancing exploration efficiency and sustained learning capability. Experiments across multiple reasoning benchmarks demonstrate that our method enables interpretable and controllable control over the (vision-) language model's strategic behavior, thereby achieving consistent performance gains over standard RL methods.

Recent advances in Vision-Language Models (VLMs) have benefited from Reinforcement Learning (RL) for enhanced reasoning. However, existing methods still face critical limitations, including the lack of low-level visual information and effective visual feedback. To address these problems, this paper proposes a unified multimodal interleaved reasoning framework ForeSight, which enables VLMs to See Further with low-level visual cues and Think Deeper with effective visual feedback. First, it introduces a set of low-level visual tools to integrate essential visual information into the reasoning chain, mitigating the neglect of fine-grained visual features. Second, a mask-based visual feedback mechanism is elaborated to incorporate visual reflection into the thinking process, enabling the model to dynamically re-examine and update its answers. Driven by RL, ForeSight learns to autonomously decide on tool invocation and answer verification, with the final answer accuracy as the reward signal. To evaluate the performance of the proposed framework, we construct a new dataset, Character and Grounding SalBench (CG-SalBench), based on the SalBench dataset. Experimental results demonstrate that the ForeSight-7B model significantly outperforms other models with the same parameter scale, and even surpasses the current SOTA closed-source models on certain metrics.

Recent studies have demonstrated that Reinforcement Learning (RL), notably Group Relative Policy Optimization (GRPO), can intrinsically elicit and enhance the reasoning capabilities of Vision-Language Models (VLMs). However, despite the promise, the underlying mechanisms that drive the effectiveness of RL models as well as their limitations remain underexplored. In this paper, we highlight a fundamental behavioral distinction between RL and base models, where the former engages in deeper yet narrow reasoning, while base models, despite less refined along individual path, exhibit broader and more diverse thinking patterns. Through further analysis of training dynamics, we show that GRPO is prone to diversity collapse, causing models to prematurely converge to a limited subset of reasoning strategies while discarding the majority of potential alternatives, leading to local optima and poor scalability. To address this, we propose Multi-Group Policy Optimization (MUPO), a simple yet effective approach designed to incentivize divergent thinking across multiple solutions, and demonstrate its effectiveness on established benchmarks.

Text-to-Image (T2I) generation has achieved remarkable progress in recent years. Meanwhile, reinforcement learning methods, particularly those based on Group Relative Policy Optimization (GRPO), have attracted widespread attention and been successfully applied to T2I tasks. However, the uniform sampling strategy commonly used during training often ignores the match between sample difficulty and the model's current learning capability, leading to low training efficiency. We argue that improving training efficiency requires continuously prioritizing prompts that match the model's evolving capability and remain actively learnable. To this end, we propose Curriculum Group Policy Optimization (CGPO), an adaptive curriculum training framework. During training, each prompt produces a group of images scored by a reward model. We use the variance of group rewards as an online proxy for prompt inconsistency. A higher variance suggests that the model has partially captured the prompt requirements but has not yet achieved stable mastery. Such prompts are more likely to provide useful learning signals, so we increase their sampling probabilities accordingly. Additionally, to address data imbalance in multi-category datasets, we design a category calibration method based on proportional fairness optimization, which balances training difficulty across categories. Experiments on GenEval, T2I-CompBench++, and DPG Bench demonstrate that our framework effectively improves generation performance.

Vision-Language-Action (VLA) models trained via imitation learning suffer from significant performance degradation in data-scarce scenarios due to their reliance on large-scale demonstration datasets. Although reinforcement learning (RL)-based post-training has proven effective in addressing data scarcity, its application to VLA models is hindered by the non-resettable nature of real-world environments. This limitation is particularly critical in high-risk domains such as industrial automation, where interactions often induce state changes that are costly or infeasible to revert. Furthermore, existing VLA approaches lack a reliable mechanism for detecting task completion, leading to redundant actions that reduce overall task success rates. To address these challenges, we propose RehearseVLA, an RL-based post-training framework that replaces physical interaction with a low-cost world model-based virtual simulator. RehearseVLA consists of two key components: (1) a physically-consistent world simulator that generates temporally consistent future visual observations, and (2) a vision-language model (VLM)-guided instant reflector that provides continuous reward signals and predicts action termination. This simulated environment enables VLA models to safely explore and generalize beyond their initial imitation learning distribution. Our method achieves notable performance gains with as few as five expert demonstrations per task. Experiments on complex robotic manipulation tasks demonstrate that RehearseVLA effectively overcomes the data inefficiency, safety constraints, and inefficient execution of conventional VLA models that rely on real-world interaction, offering a practical and scalable solution for post-training in resource-constrained settings. Our code is available at https://github.com/iSEE-Laboratory/RehearseVLA.

Reinforcement learning from human feedback (RLHF) has proven effective in aligning large language models with human preferences, inspiring the development of reward-centric diffusion reinforcement learning (RDRL) to achieve similar alignment and controllability. While diffusion models can generate high-quality outputs, RDRL remains susceptible to reward hacking, where the reward score increases without corresponding improvements in perceptual quality. We demonstrate that this vulnerability arises from the non-robustness of reward model gradients, particularly when the reward landscape with respect to the input image is sharp. To mitigate this issue, we introduce methods that exploit gradients from a robustified reward model, without requiring its retraining. Specifically, we employ gradients from a flattened reward model, obtained through parameter perturbations of the diffusion model and perturbations of its generated samples. Empirically, each method independently alleviates reward hacking and improves robustness, while their joint use amplifies these benefits. Our resulting framework, RSA-FT (Reward Sharpness-Aware Fine-Tuning), is simple, broadly compatible, and consistently enhances the reliability of RDRL.

The learning order of semantic classes significantly impacts unsupervised domain adaptation for semantic segmentation, especially under adverse weather conditions. Most existing curricula rely on handcrafted heuristics (e.g., fixed uncertainty metrics) and follow a static schedule, which fails to adapt to a model's evolving, high-dimensional training dynamics, leading to category bias. Inspired by Reinforcement Learning, we cast curriculum learning as a sequential decision problem and propose an autonomous class scheduler. This scheduler consists of two components: (i) a high-dimensional state encoder that maps the model's training status into a latent space and distills key features indicative of progress, and (ii) a category-fair policy-gradient objective that ensures balanced improvement across classes. Coupled with mixed source-target supervision, the learned class rankings direct the network's focus to the most informative classes at each stage, enabling more adaptive and dynamic learning. It is worth noting that our method achieves state-of-the-art performance on three widely used benchmarks (e.g., ACDC, Dark Zurich, and Nighttime Driving) and shows generalization ability in synthetic-to-real semantic segmentation.

Consistent image generation requires faithfully preserving identities, styles, and logical coherence across multiple images,which is essential for applications such as storytelling and character design.Supervised training approaches struggle with this task due to the lack of large-scale datasets capturing visual consistency and the complexity of modeling human perceptual preferences.In this paper, we argue that reinforcement learning (RL) offers a promising alternative by enabling models to learn complex and subjective visual criteria in a data-free manner.To achieve this, we introduce PaCo-RL, a comprehensive framework that combines a specialized consistency reward model with an efficient RL algorithm.The first component, PaCo-Reward, is a pairwise consistency evaluator trained on a large-scale dataset constructed via automated sub-figure pairing.It evaluates consistency through a generative, autoregressive scoring mechanism enhanced by task-aware instructions and CoT reasons.The second component, PaCo-GRPO,leverages a novel resolution-decoupled optimization strategy to substantially reduce RL cost,alongside a log-tamed multi-reward aggregation mechanism that ensures balanced and stable reward optimization.Extensive experiments across the two representative subtasks show that PaCo-Reward significantly improves alignment with human perceptions of visual consistency,and PaCo-GRPO achieves state-of-the-art consistency performance with improved training efficiency and stability.Together, these results highlight the promise of PaCo-RL as a practical and scalable solution for consistent image generation.We will release the code, dataset, and models.