CVPR 2026 Papers

Post-training of large-scale Vision-Language Models (VLMs) reveals a pronounced generalization gap: models fine-tuned with Reinforcement Learning (RL) consistently achieve superior out-of-distribution (OOD) performance compared to those trained with Supervised Fine-Tuning (SFT). This paper posits a data-centric explanation for this phenomenon, contending that RL's generalization advantage arises from an implicit data filtering mechanism that inherently prioritizes medium-difficulty training samples. To test this hypothesis, we systematically evaluate the OOD generalization of SFT models across training datasets of varying difficulty levels. Our results confirm that data difficulty is a critical factor, revealing that training on hard samples significantly degrades OOD performance. Motivated by this finding, we introduce Difficulty-Curated SFT (DC-SFT), a straightforward method that explicitly filters the training set based on sample difficulty. Experiments show that DC-SFT not only substantially enhances OOD generalization over standard SFT, but also surpasses the performance of RL-based training, all while providing greater stability and computational efficiency. This work offers a data-centric account of the OOD generalization gap in VLMs and establishes a more efficient pathway to achieving robust generalization. Code is available at https://github.com/byyx666/DC-SFT.

Fine-tuning large language models (LLMs) using reinforcement learning (RL) objectives has gained traction, especially in scenarios where labeled data is limited. Building on its success in the language domain, recent efforts have extended RL-based fine-tuning to multimodal tasks. Visual-RFT, for instance, applied Group Relative Policy Optimization (GRPO) to fine-tune multimodal LLMs (MLLMs) across various visual perception benchmarks, achieving notable improvements over standard supervised fine-tuning (SFT). However, its scope was limited by a narrow evaluation of RL adaptation strategies. In this work, we expand the landscape by introducing new RL-based baselines on the same benchmarks and conducting a deeper analysis of GRPO's training dynamics. We identify key limitations--such as reduced generation diversity, constrained policy exploration, and suboptimal reward formulation and aggregation. To address these, we propose DEVA: a framework that enhances Diversity via a flow-based training objective, encourages broader policy Exploration through global entropic regularization, and leverages alignment Volume as a non-verifiable reward combined with harmonic Aggregation. Applied to GRPO and other RL methods, DEVA delivers consistent gains in both quantitative (+5 to +13 points) and qualitative metrics. We further provide visualizations, ablations, and analyses to unpack the contributions of each component in our framework.

Multimodal large language models (MLLMs) have recently demonstrated remarkable reasoning and perceptual abilities for anomaly detection. However, most approaches remain confined to image-level anomaly detection and textual reasoning, while pixel-level localization still relies on external vision modules and dense annotations. In this work, we activate the intrinsic reasoning potential of MLLMs to perform anomaly detection, pixel-level localization, and interpretable reasoning solely from image-level supervision, without any auxiliary components or pixel-wise labels. Specifically, we propose Reasoning-Driven Anomaly Localization (ReAL), which extracts anomaly-related tokens from the autoregressive reasoning process and aggregates their attention responses to produce pixel-level anomaly maps. We further introduce a Consistency-Guided Reasoning Optimization (CGRO) module that leverages reinforcement learning to align reasoning tokens with visual attentions, resulting in more coherent reasoning and accurate anomaly localization. Extensive experiments on four public benchmarks demonstrate that our method significantly improves anomaly detection, localization, and interpretability. Remarkably, despite relying solely on image-level supervision, our approach achieves performance competitive with MLLM-based methods trained under dense pixel-level supervision. Code is available at https://github.com/YizhouJin313/ReADL.

Large Vision-Language Models (LVLMs) often omit or misrepresent critical visual content in generated image captions. Minimizing such information loss will force LVLMs to focus on image details to generate precise descriptions. However, measuring information loss during modality conversion is inherently challenging due to the modal gap between visual content and text output. In this paper, we argue that the quality of an image caption is positively correlated with the similarity between images retrieved via text search using that caption. Based on this insight, we further propose Cross-modal Identity Mapping (CIM), a reinforcement learning framework that enhances image captioning without requiring additional annotations. Specifically, the method quantitatively evaluates the information loss from two perspectives: Gallery Representation Consistency and Query-gallery Image Relevance. Supervised under these metrics, LVLM minimizes information loss and aims to achieve identity mapping from images to captions. The experimental results demonstrate the superior performance of our method in image captioning, even when compared with Supervised Fine-Tuning. Particularly, on the COCO-LN500 benchmark, CIM achieves a 20% improvement in relation reasoning on Qwen2.5-VL-7B. The codes are released on GitHub.

Achieving strong policy generalization to unseen environments remains a core challenge in visual reinforcement learning, and segmenting task-relevant regions to mitigate the influence of irrelevant visual cues has emerged as a promising direction. However, existing methods rely solely on the current observation, lack temporal information, and fail to exploit preceding observations, leaving learned policies susceptible to task-irrelevant background variations and ultimately degrading policy performance. In this paper, we propose temporal segmentation for task-relevant mask in visual reinforcement learning, named TSTM, which extracts task-relevant regions from sequential observations by exploiting temporal information, thereby producing more reliable masks and improving policy generalization. TSTM introduces a temporal segmentation network with an encoder-temporal-decoder architecture, where a convolutional LSTM module captures temporal dependencies across observations. To reduce inference overhead, we further develop a lightweight student network as an efficient substitute for the teacher network. The resulting task-relevant masks are encoded by a CNN-based encoder, and invariant representation learning is employed to improve robustness by enforcing consistency between representations of the original and augmented observation sequences. With these task-relevant representations, we train an actor-critic agent to learn a policy with strong generalization capability. Experimental results demonstrate that TSTM achieves superior generalization performance over existing state-of-the-art methods on most visual RL tasks.

Generative models are often deployed to make decisions on behalf of users, such as vision-language models (VLMs) identifying which person in a room is a doctor to help visually impaired individuals. Yet, VLM decisions are influenced by the perceived demographic attributes of people in the input, which can lead to biased outcomes like failing to identify women as doctors. Moreover, when reducing bias leads to performance loss, users may have varying needs for balancing bias mitigation with overall model capabilities, highlighting the demand for methods that enable controllable bias reduction during inference. Activation steering is a popular approach for inference-time controllability that has shown potential in inducing safer behavior in large language models (LLMs). However, we observe that current steering methods struggle to correct biases, where equiprobable outcomes across demographic groups are required. To address this, we propose Direct Steering Optimization (DSO) which uses reinforcement learning to find linear transformations for steering activations, tailored to mitigate bias while maintaining control over model performance. We demonstrate that DSO achieves state-of-the-art trade-off between fairness and capabilities on both VLMs and LLMs, while offering practitioners inference-time control over the trade-off. Overall, our work highlights the benefit of designing steering strategies that are directly optimized to control model behavior, providing more effective bias intervention than methods that rely on pre-defined heuristics for controllability.

The recent success of reinforcement learning (RL) in large reasoning models has inspired the growing adoption of RL for post-training Multimodal Large Language Models (MLLMs) to enhance their visual reasoning capabilities. Although many studies have reported improved performance, it remains unclear whether RL training truly enables models to learn from visual information. In this work, we propose the Hallucination-as-Cue Framework, an analytical framework designed to investigate the effects of RL-based post-training on multimodal reasoning models from the perspective of model hallucination. Specifically, we introduce hallucination-inductive, modality-specific corruptions that remove or replace essential information required to derive correct answers, thereby forcing the model to reason by hallucination. By applying these corruptions during both training and evaluation, our framework provides a unique perspective for diagnosing RL training dynamics and understanding the intrinsic properties of datasets. Through extensive experiments and analyses across multiple multimodal reasoning benchmarks, we reveal that the role of model hallucination for RL-training is more significant than previously recognized. For instance, we find that RL post-training under purely hallucination-inductive settings can still significantly improve models' reasoning performance, and in some cases even outperform standard training. These findings challenge prevailing assumptions about MLLM reasoning training and motivate the development of more modality-aware RL-based training designs.

Pre-training on large-scale videos to improve reinforcement learning efficiency is promising yet remains challenging. Existing methods typically treat the agent as an indivisible entity, modeling motion patterns globally. Such global modeling is tightly coupled with the morphology, hindering transfer across domains. In contrast, despite the vast disparity in global motions, the local components exhibit similar motion patterns across different agents. Building on this insight, we propose a novel Deconstruct-Recompose Paradigm (DRP) for learning transferable local motion representations. Specifically, in the Deconstruct phase, we identify multiple local points and track their frame-wise motions, defining each as an Atomic Action. We introduce a Dual-Attention Encoder (DAE) to learn local motion representations from these Atomic Actions, capturing their spatiotemporal relationships. In the Recompose phase, we compose local motion representations with a learnable Motion Aggregation Token '[MAT]' via latent dynamics model learning. Additionally, an adapter bridges local motion and downstream action-specific dynamics to accelerate policy learning. Extensive experiments demonstrate that our method effectively transfers to diverse robotic control and manipulation tasks, significantly improving sample efficiency and performance.

In this paper, we present a method to reconstruct physically plausible human-object interactions (HOI) from monocular videos. While existing kinematic-based approaches produce visually plausible motion, they often result in physical artifacts such as interpenetration and object floating. To overcome these issues, we introduce a physics-guided reconstruction framework that begins with a kinematic estimate and then refines it through a reinforcement learning (RL) policy trained to reproduce the interaction in a physics simulator. Because kinematic estimates are typically noisy, naive RL training can fail. Therefore, we propose an adaptive sampling strategy with a dual self-updating mechanism that automatically identifies the frames with the most informative and reliable kinematic reconstruction. Our process progressively improves reconstruction quality and yields physically consistent HOI sequences. We demonstrate our approach on two standard benchmarks and achieve clear improvements in physical plausibility metrics over state-of-the-art methods.

We present a physics-based method for simulating full-body agents that recover balance by stepping or applying contact forces after being perturbed in dense crowds. While traditional 2D crowd simulations focus on navigation and social interactions in moderately dense settings, interactions in highly dense environments are predominantly physical, leading to push propagation, falls, and potential hazards. Existing models cannot capture how forces are transmitted through the body at the limb level. To address this, we use physics-based anthropomorphic simulations combined with a two-stage deep reinforcement learning framework. In the first stage, a policy is pre-trained using reference motion data and general balance rewards, enabling agents to handle a wide range of perturbations. In the second stage, an adaptive phase refines the policy to allow socially aware interactions, using hand contacts for stabilization guided by an online heuristic targeting neighbors' shoulders based on mechanical efficiency and collision risk. Ablation studies validate the training framework and reward components, and simulations reproduce trends observed in empirical studies of push propagation. Our method scales to large populations, offering new opportunities to study safety and collective behavior in dense crowds.

Multi-task learning (MTL) involves the simultaneous optimization of multiple task-specific losses, often leading to gradient conflicts and scale imbalances that result in negative transfer. While existing multi-task optimization methods attempt to mitigate these challenges, they either lack the stochasticity needed to escape poor local minima or fail to explicitly address conflicts at the gradient level. In this work, we propose TaskForce, a novel multi-task optimization framework incorporating cooperative multi-agent reinforcement learning (MARL), where agents learn to find an effective joint optimization strategy based on their respective task gradients and losses. To keep the optimization process compact yet informative, agents observe a summary of the training dynamics that consists of the gradient Gram matrix---capturing both gradient magnitudes and pairwise alignments---and task loss values. Each agent then predicts the balancing parameters that determine the weight of their contribution to the final gradient update. Crucially, we design a hybrid reward function that incorporates both gradient-based signals and loss improvement dynamics, enabling agents to effectively resolve gradient conflicts and avoid poor convergence by considering both direct gradient information and the resulting impact on loss reduction. TaskForce achieves consistent improvements over state-of-the-art MTL baselines on NYU-v2, Cityscapes, and QM9, demonstrating the promise of cooperative MARL in complex multi-task scenarios.

Building reconstruction aims to extract compact wireframes from point clouds. Recent edge-based methods achieve impressive results but suffer from sparse supervision from one-to-one matching, which leaves most edge proposals under-optimized. In this paper, we present Group Relative Edge Optimization (GREO), the first attempt to incentivize dense supervision across edges proposals through reinforcement learning-style optimization in wireframe reconstruction. Specifically, GREO computes edge-level rewards based on geometric alignment quality and transforms them into target confidence distributions via group-wise normalization. In addition, we incorporate entropy regularization to maintain distributional stability and prevent confidence collapse. This joint optimization enables dense and discriminative supervision across all edge proposals through cross-entropy minimization. Experiments on the large-scale Building3D dataset demonstrate that our powerful and versatile GREO integrates seamlessly into existing edge-based methods as a plug-and-play training strategy, achieving state-of-the-art performance on both the Entry-level and Tallinn benchmarks while adding no inference overhead.

Predicting wildfire risk is a reasoning-intensive spatial problem that requires the integration of visual, climatic, and geographic factors to infer continuous risk maps. Existing methods lack the causal reasoning and multimodal understanding required for reliable generalization. We introduce FireScope-Bench, a large-scale dataset and benchmark that couples Sentinel-2 imagery and climate data with expert-defined risk rasters across the USA, and real wildfire events in Europe for cross-continental evaluation. Building on this dataset, we propose FireScope, a VLM-based reasoning-to-generation framework that learns from both reinforcement learning and visual supervision to predict risk rasters with complementary reasoning traces. When trained in the USA and tested in Europe, FireScope achieves substantial performance gains, while expert feedback and automated analysis confirm that its reasoning traces are faithful and semantically meaningful. Our findings demonstrate that reasoning can ground raster prediction models, improving both generalization and interpretability. To our knowledge, this is the first framework to (1) demonstrate that language-based reasoning can improve generalization in visual generation, (2) propose a high-resolution wildfire risk model that can be applied across continents, and (3) enable systematic studies of robust cross-continental generalization for multimodal fire risk models. We believe that FireScope-Bench has the potential to serve as a foundation for advancing reasoning-driven, interpretable and generalizable spatial modeling. Data and source code are available at github.com/insait-institute/FireScope.

Open-Vocabulary Human-Object Interaction (OV-HOI) detection aims to recognize novel HOI categories beyond the training set. Existing OV-HOI detection approaches typically leverage CLIP to extract global visual representations and perform cross-attention between learnable queries and global features to localize human-object pairs. However, such one-stage paradigms tend to overfit seen interactions, limiting their generalization to unseen categories, while the coarse spatial awareness of CLIP also hinders the localization of fine-grained interaction cues. To address these issues, we propose a novel Semantic-Diversified and Interaction-Focused framework (SD-IF), which integrates reinforcement-guided adaptive optimization to jointly enhance semantic generalization and spatial discrimination. Specifically, we introduce a Semantic Diversification (SD) module that applies reinforcement-driven stochastic semantic perturbations and dual-level semantic exploration, expanding the semantic coverage of queries while maintaining visual coherence and effectively encouraging exploration beyond the seen semantic clusters. Furthermore, we design an Interaction Focusing (IF) module that formulates an actor-critic optimization scheme to adaptively refine attention distributions based on detection features and interaction representations, guided by a hybrid reward combining spatial focusing and semantic consistency. Extensive experiments on two standard benchmarks demonstrate the effectiveness of our SD-IF.

High-quality 4D reconstruction enables photorealistic and immersive rendering of the dynamic real world. However, unlike static scenes that can be fully captured with a single camera, high-quality dynamic scenes typically require dense arrays of tens or even hundreds of synchronized cameras. Dependence on such costly lab setups severely limits practical scalability. The reliance on such costly lab setups severely limits practical scalability. To this end, we propose a sparse-camera dynamic reconstruction framework that exploits abundant yet inconsistent generative observations. Our key innovation is the Spatio-Temporal Distortion Field, which provides a unified mechanism for modeling inconsistencies in generative observations across both spatial and temporal dimensions. Building on this, we develop a complete pipeline that enables 4D reconstruction from sparse and uncalibrated camera inputs. We evaluate our method on multi-camera dynamic scene benchmarks, achieving spatio-temporally consistent high-fidelity renderings and significantly outperforming existing approaches.

Multi-modal large language models (MLLMs) have advanced general-purpose video understanding but struggle with long, high-resolution videos---they process every pixel equally in their vision transformers (ViTs) or LLMs despite significant spatiotemporal redundancy. We introduce AutoGaze, a lightweight module that removes redundant patches before processed by a ViT or an MLLM. Trained with next-token prediction and reinforcement learning, AutoGaze autoregressively selects a minimal set of multi-scale patches that reconstructs the video within a user-specified error threshold, eliminating redundancy while preserving information. Empirically, AutoGaze reduces visual tokens by 4x-100x and accelerates ViTs and MLLMs by up to 19x, scaling MLLMs to 1K-frame 4K-resolution videos and achieving superior results on video benchmarks (e.g., 67.0% on VideoMME). Furthermore, we introduce HLVid: the first high-resolution, long-form video QA benchmark with 5-minute 4K-resolution videos, where an MLLM scaled with AutoGaze improves over the baseline by 10.1% and outperforms the previous best MLLM by 4.5%. Project page: https://autogaze.github.io/.

Human behaviors in real-world environments are inherently interactive, with an individual's motion shaped by surrounding agents and the scene. Such capabilities are essential for applications in virtual avatars, interactive animation, and human-robot collaboration. We target real-time human interaction-to-reaction generation, which generates the ego's future motion from dynamic multi-source cues, including others' actions, scene geometry, and optional high-level semantic inputs. This task is fundamentally challenging due to (i) limited and fragmented interaction data distributed across heterogeneous single-person, human-human, and human-scene domains, and (ii) the need to produce low-latency yet high-fidelity motion responses during continuous online interaction. To address these challenges, we propose ReMoGen (Reaction Motion Generation), a modular learning framework for real-time interaction-to-reaction generation. ReMoGen leverages a universal motion prior learned from large-scale single-person motion datasets and adapts it to target interaction domains through independently trained Meta-Interaction modules, enabling robust generalization under data-scarce and heterogeneous supervision. To support responsive online interaction, ReMoGen performs segment-level generation together with a lightweight Frame-wise Segment Refinement module that incorporates newly observed cues at the frame level, improving both responsiveness and temporal coherence without expensive full-sequence inference. Extensive experiments across human-human, human-scene, and mixed-modality interaction settings show that ReMoGen produces high-quality, coherent, and responsive reactions, while generalizing effectively across diverse interaction scenarios.

Multi-turn reinforcement learning (RL) for multi-modal agents built upon vision-language models (VLMs) is hampered by sparse rewards and long-horizon credit assignment. Recent methods densify the reward by querying a teacher that provides step-level feedback, e.g., Guided Thought Reinforcement (GTR) and On-Policy Distillation, but rely on costly, often privileged models as the teacher, limiting practicality and reproducibility. We introduce GTR-Turbo, a highly efficient upgrade to GTR that matches its performance without training on or querying an expensive teacher model. Specifically, GTR-Turbo merges the weights of checkpoints produced during ongoing RL training and then uses the resulting merged model as a "free" teacher to guide subsequent RL via supervised fine-tuning or soft logit distillation. This design removes dependence on privileged VLMs (e.g., GPT or Gemini), mitigates the "entropy collapse" observed in prior work, and maintains stable training. Across diverse visual agentic tasks, GTR-Turbo improves the accuracy of the baseline model by 10-30% while reducing wall-clock training time by 50% and compute cost by 60% relative to GTR.

We revisit the problem of estimating the fundamental matrix of a pair of perspective cameras, a cornerstone of geometric computer vision. As is well-known, linear solvers require at least 8 point correspondences, whereas nonlinear minimal solvers require just 7 in the uncalibrated case or 5 in the calibrated case. In this paper, we consider a special case of the 7-point problem where 5 of the points are configured to lie on two lines, which has previously been shown to have a unique solution. As a theoretical contribution, we offer a completely elementary analysis of how this uniqueness manifests in the standard 7-point algorithm. On a practical level, we provide the first linear solver for the minimal problem associated to this special configuration. Additionally, we evaluate a heuristic 5-point fundamental matrix solver based on the construction of virtual midpoints. When combined with early non-minimal fitting, the runtime and accuracy of our solver is competitive with the state-of-the-art (SOTA) on multiple benchmarks. Code is available at https://github.com/CIVA-Lab/v-umlaut.