CVPR 2026 Papers

Recently, reinforcement learning (RL) has been employed for improving generative image super-resolution (ISR) performance. However, the current efforts are focused on multi-step generative ISR, while one-step generative ISR remains underexplored due to its limited stochasticity. In addition, RL methods such as Direct Preference Optimization (DPO) require the generation of positive and negative sample pairs offline, leading to a limited number of samples, while Group Relative Policy Optimization (GRPO) only calculates the likelihood of the entire image, ignoring local details that are crucial for ISR. In this paper, we propose Group Direct Preference Optimization (GDPO), a novel approach to integrate RL into one-step generative ISR model training. First, we introduce a noise-aware one-step diffusion model that can generate diverse ISR outputs. To prevent performance degradation caused by noise injection, we introduce an unequal-timestep strategy to decouple the timestep of noise addition from that of diffusion. We then present the GDPO strategy, which integrates the principle of GRPO into DPO, to calculate the group-relative advantage of each online generated sample for model optimization. Meanwhile, an attribute-aware reward function is designed to dynamically evaluate the score of each sample based on its statistics of smooth and texture areas. Experiments demonstrate the effectiveness of GDPO in enhancing the performance of one-step generative ISR models. Code: https://github.com/Joyies/GDPO.

Graphical User Interface (GUI) agents require effective utilization of historical context to perform sequential navigation tasks. While incorporating past actions and observations can significantly improve decision-making, naively using full history leads to excessive computational overhead and potential distraction from irrelevant information. In this work, we introduce ****HiconAgent****, a GUI agent trained with ****History Context-aware Policy Optimization (HCPO)**** for effective and efficient utilization of historical information. HCPO explicitly optimizes history usage in both sampling and policy updates by integrating two complementary components: ****(1) Dynamic Context Sampling (DCS)**** presents the agent with variable-length histories during sampling, enabling adaptive use of the most relevant historical context to improve sequential decision quality; ****(2) Anchor-guided History Compression (AHC)**** refines the policy update phase via a dual-branch optimization strategy, where the compressed branch drops history observations while keeping history actions as information flow anchors. The compressed and uncompressed branches are coupled through a history-enhanced alignment loss to enforce consistent history usage, achieving efficiency with minimal performance degradation. Extensive experiments on mainstream GUI navigation benchmarks demonstrate the strong performance of our model. Despite its smaller size, HiconAgent-3B outperforms GUI-R1-7B by ****+8.46****% grounding and ****+11.32****% step successful rate on GUI-Odyssey, while achieving comparable results on AndroidControl and AITW, with up to ****2.47x computational speedup**** and ****60% FLOPs reduction****.

Recent advances in image-to-video (I2V) generation have achieved remarkable progress in synthesizing high-quality, temporally coherent videos from static images. Among all the applications of I2V, human-centric video generation includes a large portion. However, existing I2V models encounter difficulties in maintaining identity consistency between the input human image and the generated video, especially when the person in the video exhibits significant expression changes and movements. This issue becomes critical when the human face occupies merely a small fraction of the image. Since humans are highly sensitive to identity variations, this poses a critical yet under-explored challenge in I2V generation. In this paper, we propose Identity-Preserving Reward-guided Optimization (IPRO), a novel video diffusion framework based on reinforcement learning to enhance identity preservation. Instead of introducing auxiliary modules or altering model architectures, our approach introduces a direct and effective tuning algorithm that optimizes diffusion models using a face identity scorer. To improve performance and accelerate convergence, our method backpropagates the reward signal through the last steps of the sampling chain, enabling richer gradient feedback. We also propose a novel facial scoring mechanism that treats faces in ground-truth videos as facial feature pools, providing multi-angle facial information to enhance generalization. A KL-divergence regularization is further incorporated to stabilize training and prevent overfitting to the reward signal. Extensive experiments on Wan 2.2 I2V model and our in-house I2V model demonstrate the effectiveness of our method. Our project is available at https://dof-gaussian.github.io/.

Multimodal large language models (MLLMs) that "think with images" can interactively use tools to reason about visual inputs, but current approaches often rely on a narrow set of tools with limited real-world necessity and scalability. In this work, we first reveal a critical and previously overlooked weakness: even state-of-the-art MLLMs are surprisingly brittle, showing significant performance degradation on images with simple orientation changes, underscoring the need for more robust tool-based reasoning. To address this, we propose CodeVision, a flexible and scalable "code-as-tool" framework where the model generates code as a universal interface to invoke any image operation, moving beyond fixed tool registries. We train our model using a two-stage methodology, beginning with Supervised Fine-Tuning (SFT) on a high-quality dataset curated for complex, multi-turn tool composition and error recovery, followed by Reinforcement Learning (RL) with a dense process reward function to encourage strategic and efficient tool use. To facilitate this research, we construct new SFT and RL datasets and introduce a challenging new benchmark suite designed to rigorously evaluate robustness to orientation changes and multi-tool reasoning. Experiments on Qwen2.5-VL and Qwen3-VL series show that our approach significantly improves model performance and fosters emergent capabilities such as flexible tool composition, efficient chained execution, and robust error recovery from runtime feedback. Code is available at https://github.com/ByteDance-BandAI/CodeVision.

Classifying fine-grained visual concepts under open-world settings, i.e., without a predefined label set, demands models to be both accurate and specific. Recent reasoning Large Multimodal Models (LMMs) exhibit strong visual understanding capability but tend to produce overly generic predictions when performing fine-grained image classification. Our preliminary analysis reveals that models do possess the intrinsic fine-grained domain knowledge. However, promoting more specific predictions (specificity) without compromising correct ones (correctness) remains a non-trivial and understudied challenge. In this work, we investigate how to steer reasoning LMMs toward predictions that are both correct and specific. We propose a novel specificity-aware reinforcement learning framework, SpeciaRL, to fine-tune reasoning LMMs on fine-grained image classification under the open-world setting. SpeciaRL introduces a dynamic, verifier-based reward signal anchored to the best predictions within online rollouts, promoting specificity while respecting the model's capabilities to prevent incorrect predictions. Our out-of-domain experiments show that SpeciaRL delivers the best trade-off between correctness and specificity across extensive fine-grained benchmarks, surpassing existing methods and advancing open-world fine-grained image classification. Code and model are publicly available at https://github.com/s-angheben/SpeciaRL.

While large language models provide strong compositional reasoning, existing reasoning segmentation pipelines fail to transparently connect this reasoning to visual perception. Current methods, such as latent query alignment, are end-to-end yet opaque "black boxes". Conversely, textual localization readout is merely readable, not truly interpretable, often functioning as an unconstrained post-hoc step. To bridge this interpretability gap, we propose SegCompass, an end-to-end model that leverages a Sparse Autoencoder (SAE) to forge an explicit, interpretable, and differentiable alignment pathway. Given an image-instruction pair, SegCompass first generates a chain-of-thought (CoT) trace. The core of our method is an SAE that maps both the CoT and visual tokens into a shared, high-dimensional sparse concept space. A query codebook selects salient concepts from this space, which are then spatially grounded by a slot mapper into a multi-slot heatmap that guides the final mask decoder. The entire model is trained jointly, unifying reinforcement learning for the reasoning path with standard segmentation supervision. This SAE-driven interface provides a "white-box" connection that is significantly more traceable than latent queries and more coherent than textual readouts. Extensive experiments on five challenging benchmarks demonstrate that SegCompass matches or surpasses state-of-the-art performance. Crucially, our visual and quantitative analyses show a strong correlation between the quality of the learned sparse concepts and final mask accuracy, confirming that SegCompass achieves superior results through its enhanced and inspectable alignment. Code is available at https://github.com/ZhenyuLU-Heliodore/SegCompass.

Conventional evaluation methods for multimodal LLMs (MLLMs) lack interpretability and are often insufficient to fully disclose significant capability gaps across models. To address this, we introduce AuditDM, an automated framework that actively discovers and rectifies MLLM failure modes by auditing their divergence. AuditDM fine-tunes an MLLM as an auditor via reinforcement learning to generate challenging questions and counterfactual images that maximize disagreement among target models. Once trained, the auditor uncovers diverse, interpretable exemplars that reveal model weaknesses and serve as annotation-free data for rectification. When applied to SoTA models like Gemma-3 and PaliGemma-2, AuditDM discovers more than 20 distinct failure types. Fine-tuning on these discoveries consistently improves all models across 16 benchmarks, and enables a 3B model to surpass its 28B counterpart. Our results suggest that as data scaling hits diminishing returns, targeted model auditing offers an effective path to model diagnosis and improvement.

Vision-Language-Action (VLA) models excel in robotic manipulation but are constrained by their heavy reliance on expert demonstrations, leading to demonstration bias and limiting performance. Reinforcement learning (RL) is a vital post-training strategy to overcome these limits, yet current VLA-RL methods, including group-based optimization approaches, are crippled by severe reward sparsity. Relying on binary success indicators wastes valuable information in failed trajectories, resulting in low training efficiency. To solve this, we propose Self-Referential Policy Optimization (SRPO), a novel VLA-RL framework. SRPO eliminates the need for external demonstrations or manual reward engineering by leveraging the model's own successful trajectories, generated within the current training batch, as a self-reference. This allows us to assign a progress-wise reward to failed attempts. A core innovation is the use of Latent World Representations to measure behavioral progress robustly. Instead of relying on raw pixels or requiring domain-specific fine-tuning, we utilize the compressed, transferable encodings from a world model's latent space. These representations naturally capture progress patterns across environments, enabling accurate, generalized trajectory comparison. Empirical evaluations on the LIBERO benchmark demonstrate SRPO's efficiency and effectiveness. Starting from a supervised baseline with 48.9% success, SRPO achieves a new state-of-the-art success rate of 99.2% in just 200 RL steps, representing a 103% relative improvement without any extra supervision. Furthermore, SRPO shows substantial robustness, achieving a 167% performance improvement on the LIBERO-Plus benchmark.

Graphical User Interface (GUI) Agents, benefiting from recent advances in multimodal large language models (MLLM), have achieved significant development. However, due to the frequent updates of GUI applications, adapting to new tasks without forgetting old tasks in GUI continual learning remains an open problem. In this work, we reveal that while Supervised Fine-Tuning (SFT) facilitates fast adaptation, it often triggers knowledge overwriting, whereas Reinforcement Learning (RL) demonstrates an inherent resilience that shields prior interaction logic from erasure. Based on this insight, we propose a Continual GUI Learning (CGL) framework that dynamically balances adaptation efficiency and skill retention by enhancing the synergy between SFT and RL. Specifically, we introduce an SFT proportion adjustment mechanism guided by policy entropy to dynamically control the weight allocation between the SFT and RL training phases. To resolve explicit gradient interference, we further develop a specialized gradient surgery strategy. By projecting exploratory SFT gradients onto GRPO-based anchor gradients, our method explicitly clips the components of SFT gradients that conflict with GRPO. On top of that, we establish an AndroidControl-CL benchmark, which divides GUI applications into distinct task groups to effectively simulate and evaluate the performance of continual GUI learning. Experimental results demonstrate the effectiveness of our proposed CGL framework across continual learning scenarios. The benchmark, code, and model will be made publicly available.

Efficiently understanding long-form videos remains a fundamental challenge for Multimodal Large Language Models (MLLMs). In this paper, we present MLLM-Sampler Joint Evolution (MSJoE), a novel framework that jointly evolves the MLLM and a lightweight key-frame sampler for efficient long-form video understanding. MSJoE builds upon a key assumption that only a small subset of key-frames is truly informative for answering each question to a video. Specifically, MSJoE first reasons out several queries, which describe diverse visual perspectives relevant to the question. Then, these queries interact with a frozen CLIP model to produce a query-frame similarity matrix. Finally, A lightweight sampler predicts key-frame sampling weights from this matrix, selecting informative frames, which are then fed into the MLLM for answer generation. Both the MLLM and sampler are jointly optimized through reinforcement learning, enabling co-adaptation of query-reasoning, frame-sampling, and key-frame understanding. A new long-video QA dataset containing 2.8k videos with 7k question-answer pairs is collected to support the training process. Extensive experiments on VideoMME, LongVideoBench, LVBench, and MLVU show that MSJoE achieves 8.0% accuracy gain upon the base MLLM, and 1.1% higher accuracy than strongest baseline method.

The aesthetic quality of a scene depends strongly on camera viewpoint. Existing approaches for aesthetic viewpoint suggestion are either single-view adjustments, predicting limited camera adjustments from a single image without understanding scene geometry, or 3D exploration approaches, which rely on dense captures or prebuilt 3D environments coupled with costly reinforcement learning (RL) searches. In this work, we introduce the notion of 3D aesthetic field that enables geometry-grounded aesthetic reasoning in 3D with sparse captures, allowing efficient viewpoint suggestions in contrast to costly RL searches. We opt to learn this 3D aesthetic field using a feedforward 3D Gaussian Splatting network that distills high-level aesthetic knowledge from a pretrained 2D aesthetic model into 3D space, enabling aesthetic prediction for novel viewpoints from only sparse input views. Building on this field, we propose a two-stage search pipeline that combines coarse viewpoint sampling with gradient-based refinement, efficiently identifying aesthetically appealing viewpoints without dense captures or RL exploration. Extensive experiments show that our method consistently suggests viewpoints with superior framing and composition compared to existing approaches, establishing a new direction toward 3D-aware aesthetic modeling.

We present UIKA, a feed-forward animatable Gaussian head model from an arbitrary number of pose-free inputs, including a single image, multi-view captures, and smartphone-captured videos. Unlike the traditional avatar method, which requires a studio-level multi-view capture system and reconstructs a human-specific model through a long-time optimization process, we rethink the task through the lenses of model representation, network design, and data preparation. First, we introduce a UV-guided avatar modeling strategy, in which each input image is associated with a pixel-wise facial correspondence estimation. Such correspondence estimation allows us to reproject each valid pixel color from screen space to UV space, which is independent of camera pose and character expression. Furthermore, we design learnable UV tokens on which the attention mechanism can be applied at both the screen and UV levels. The learned UV tokens can be decoded into canonical Gaussian attributes using aggregated UV information from all input views. To train our large avatar model, we additionally prepare a large-scale, identity-rich synthetic training dataset. Our method significantly outperforms existing approaches in both monocular and multi-view settings. See more details in our project page: https://zijian-wu.github.io/uika-page/.

Neural networks are often treated as monolithic black boxes that process all inputs uniformly through all layers. However, researchers intuitively wonder: do simple images require all 50 layers of ResNet-50, or is the prediction effectively decided much earlier? We investigate when pretrained models make up their minds during a forward pass by training linear probes at each layer of ResNet variants on ImageNet, without modifying the base model. Our findings reveal substantial computational heterogeneity across architectures. We discover pronounced bimodal patterns with distinct populations of early and late deciders: in deeper ResNets, 33--39% of samples stabilize within the first third of the network, while 39--42% are assigned to the final 30% of depth; across all three ResNets, the late fraction ranges from 39% to 54%. The decision layer is highly sensitive to stability criteria, with mean depths increasing from 2.6--4.1 (k=1) to 9.0--10.0 (k=4). Linear probe accuracy exhibits sharp jumps in final residual stages, reaching 73--75% for ResNet-50/101 and 65% for ResNet-18, indicating that semantic consolidation occurs late. These findings expose computational heterogeneity in standard inference and clarify both the promise and the limits of stability-based early exiting.

Achieving efficient and robust whole-body control (WBC) is essential for enabling humanoid robots to perform complex tasks in dynamic environments. Despite the success of reinforcement learning (RL) in this domain, its sample inefficiency remains a significant challenge due to the intricate dynamics and partial observability of humanoid robots. To address this limitation, we propose **PvP**, a **P**roprioceptive-**P**rivileged contrastive learning framework that leverages the intrinsic complementarity between proprioceptive and privileged states. PvP learns compact and task-relevant latent representations without requiring hand-crafted data augmentations, enabling faster and more stable policy learning. To support systematic evaluation, we develop **SRL4Humanoid**, the first unified and modular framework that provides high-quality implementations of representative state representation learning (SRL) methods for humanoid robot learning. Extensive experiments on the LimX Oli robot across velocity tracking and motion imitation tasks demonstrate that PvP significantly improves sample efficiency and final performance compared to baseline SRL methods. Our study further provides practical insights into integrating SRL with RL for humanoid WBC, offering valuable guidance for data-efficient humanoid robot learning.

Despite Multimodal Large Language Models (MLLMs) having shown impressive capabilities, they may suffer from hallucinations. Empirically, we find that MLLMs attend disproportionately to task-irrelevant background regions compared with text-only LLMs, implying spurious background-answer correlations. We claim and analyze that (i) outcome-based rewards can be an important factor leading to spurious correlations, and (ii) spurious correlations can be an important factor leading to hallucinations. Based on these results, we propose Causal-Oriented Policy Optimization (COPO) to mitigate these spurious correlations, thus addressing the issue of hallucinations. It imposes token-level sufficiency and necessity constraints to measure each inference token's causal contribution, thus ensuring correct and evidence-grounded output. Specifically, we first evaluate each token's causal contribution via a newly proposed causal completeness reward. This reward is then used to construct a causally informed advantage function within the GRPO optimization framework, encouraging the model to focus on tokens that are causally sufficient and necessary for accurate generation. Experimental results across various benchmarks demonstrate the advantages of COPO.

Discrete Diffusion Models (DDMs) have shown great potential in image generation, especially when equipped with reinforcement learning (RL) techniques.However, a fundamental yet overlooked limitation is revealed in our experiments: severe imbalance of credit assignment across timesteps during training. As a result, early generation timesteps, which carry higher exploration potential and determine the global structure, provide a smaller contribution to policy optimization.To conquer this, we propose a simple-yet-effective approach, to Re-balance timestep credit of Reinforcement Learning (RebRL) for better exploration-exploitation trade-off and more efficient training of DDMs.RebRL is plug-and-play\textemdash simply replacing uniform temporal policy with strategic rebalancing along masking stages. RebRL is analytically plausible\textemdashderivation and analysis show that it enjoys a uniform token-level policy gradient, which benefits policy optimization.Experiments on text-to-image generation benchmarks show that RebRL achieves state-of-the-art performance on GenEval and improves human preference score by up to 3.40 while effectively reducing training steps by ~40\%.

Multimodal large language models (MLLMs) have rapidly advanced, yet their adoption in medicine remains limited by gaps in domain coverage, modality alignment, and grounded reasoning. In this work, we introduce MedMO, a medical foundation model built upon a generalized MLLM architecture and trained exclusively on large-scale, domain-specific data. MedMO follows a multi-stage training recipe: (i) cross-modal pretraining to align heterogeneous visual encoders with a medical language backbone; (ii) instruction tuning on multi-task supervision that spans captioning, VQA, report generation, retrieval, and grounded disease localization with bounding boxes; and (iii) reinforcement learning with verifiable rewards that combine factuality checks with a box-level GIoU reward to strengthen spatial grounding and reasoning in complex clinical scenarios. MedMO consistently outperforms strong open-source medical MLLMs across multiple modalities and tasks. On VQA benchmarks, MedMO achieves an average accuracy improvement of +21.3% over the baseline and performs within 0.6% of the SOTA Fleming-VL. For text-based QA, it attains +7.9% over the baseline and +15.6% over Fleming-VL. In medical report generation, MedMO delivers significant gains in both semantic and clinical accuracy. Moreover, it exhibits strong grounding capability, achieving an IoU improvement of +40.4 over the baseline and +37.0% over Fleming-VL, underscoring its robust spatial reasoning and localization performance. Evaluations across radiology, ophthalmology, pathology, and emergency care confirm MedMO's broad cross-modality generalization and reliable spatial reasoning.

Energy-based predictive world models provide a powerful approach for multi-step visual planning by reasoning over latent energy landscapes rather than generating pixels. However, existing approaches face two major challenges: (i) their latent representations are typically learned in Euclidean space, neglecting the underlying geometric and hierarchical structure among states, and (ii) they struggle with long-horizon prediction, which leads to rapid degradation across extended rollouts. To address these challenges, we introduce GeoWorld, a geometric world model that preserves geometric structure and hierarchical relations through a Hyperbolic JEPA, which maps latent representations from Euclidean space onto hyperbolic manifolds. We further introduce Geometric Reinforcement Learning for energy-based optimization, enabling stable multi-step planning in hyperbolic latent space. Extensive experiments on CrossTask and COIN demonstrate around 3% SR improvement in 3-step planning and 2% SR improvement in 4-step planning compared to the state-of-the-art V-JEPA-2.

Learning latent actions from large-scale videos is crucial for the pre-training of scalable embodied foundation models, yet existing methods often struggle with action-irrelevant distractors. Although incorporating action supervision can alleviate these distractions, its effectiveness is restricted by the scarcity of available action labels. Optical flow represents pixel-level motion between consecutive frames, naturally suppressing background elements and emphasizing moving objects. Motivated by this, we propose robust Latent Action learning with Optical Flow constraints (LAOF), a pseudo-supervised framework that leverages the agent's optical flow as an action-driven signal to learn latent action representations robust to distractors. Experimental results show that the latent representations learned by LAOF outperform existing methods on downstream imitation learning and reinforcement learning tasks. This superior performance arises from optical flow constraints, which substantially stabilize training and improve the quality of latent representations under extremely label-scarce conditions, while remaining effective as the proportion of action labels increases to 10%. Importantly, even without action supervision, LAOF matches or surpasses action-supervised methods trained with 1% of action labels.