CVPR 2026 Papers

Vision Language Models (VLMs) demonstrate strong qualitative visual understanding, but struggle with metrically precise spatial reasoning required for embodied applications. The agentic paradigm promises that VLMs can use a wide variety of tools that could augment these capabilities, such as depth estimators, segmentation models, and pose estimators. Yet it remains an open challenge how to realize this vision without solely relying on handcrafted prompting strategies or enforcing fixed, predefined tool pipelines that limit VLMs' ability to discover optimal tool-use patterns. Reinforcement Learning could overcome this gap, but has so far been limited to reasoning with a single visual tool due to the large search space in multi-tool reasoning. We introduce Double Interactive Reinforcement Learning (DIRL), a two-phase training framework where VLMs learn to coordinate multiple tools through interactive exploration and feedback. In the teaching phase, we combine demonstrations from a single tool specialist trained via interactive RL with traces from a frontier model using all tools. In the exploration phase, the model further refines multi-tool coordination through continued RL. Our model, SpaceTools, with tool-augmented spatial reasoning ability, achieves state-of-the-art performance on spatial understanding benchmarks (RoboSpatial-Home, BLINK, Internal Benchmark) and demonstrates reliable real-world manipulation using a 7-DOF robot as a tool. DIRL provides substantial improvements over the vanilla SFT (+12% on RoboSpatial) and RL (+16% on RoboSpatial) baselines. Project page: https://spacetools.github.io/.

Humanoid robotics has strong potential to transform daily service and caregiving applications. Although recent advances in general motion tracking within physics engines (GMT) have enabled virtual characters and humanoid robots to reproduce a broad range of human motions, these behaviors are primarily limited to contact-less social interactions or isolated movements. Assistive scenarios, by contrast, require continuous awareness of a human partner and rapid adaptation to their evolving posture and dynamics. In this paper, we formulate the imitation of closely interacting, force-exchanging human-human motion sequences as a multi-agent reinforcement learning problem. We jointly train partner-aware policies for both the supporter (assistant) agent and the recipient agent in a physics simulator to track assistive motion references. To make this problem tractable, we introduce a partner policies initialization scheme that transfers priors from single-human motion-tracking controllers, greatly improving exploration. We further propose dynamic reference retargeting and contact-promoting reward, which adapt the assistant's reference motion to the recipient's real-time pose and encourage physically meaningful support. We show that \model is the first method capable of successfully tracking assistive interaction motions on established benchmarks, demonstrating the benefits of a multi-agent RL formulation for physically grounded and socially aware humanoid control.

Unified interpreting and synthesizing human behaviors within 3D environments is vital for advancing spatial intelligence and humanoid robotics. Despite recent advancements (e.g., HSI-GPT), two fundamental capabilities expected of a unified model--understanding and generation--still lag behind specialist models. This is primarily due to 1) single-granularity codebook overemphasizes low-frequency motion details while neglecting motion semantics, 2) limited decoding capacity of the motion detokenizer which restricts the fidelity of human-scene interactions, 3) only relying on supervised fine-tuning (SFT) failing to capture high-level motion semantics and logical reasoning with an end-to-end mapping. To this issue, we develop HSI-GPT2--a reasoning-enhanced, dual-granularity motion-representational large Scene-Motion-Language model, powered by reinforcement learning (RL) with Chain-of-Thought (CoT) reasoning. First, HSI-GPT2 introduces a Dual-granularity Motion Tokenizer, DMoTok, which jointly preserves both fine-grained motion details and text-aligned motion semantics for various HSI-related tasks. Further, a motion diffusion decoder functions as a motion detokenizer, translating deep semantics and detailed features of LLMs into physically grounded human motions. Finally, we curate a Motion Chain-of-Thought (MoCoT) data engine and extend a Group Relative Policy Optimization (GRPO) paradigm to execute long-horizon and compositionallyrich commands. Results on standard HSI benchmarks confirm the clear superiority of HSI-GPT2 in enhancing interaction quality, semantic alignment, behavioral diversity, and generalization to unseen 3D scenes.

Recent advances in Multimodal Large Language Models (MLLMs) have enabled automated generation of structured layouts from natural language descriptions. Existing methods typically follow a code-only paradigm that generates code to represent layouts, which are then rendered by graphic engines to produce final images. However, they are blind to the rendered visual outcome, making it difficult to guarantee readability and aesthetics. In this paper, we identify visual feedback as a critical factor in layout generation and propose Visual Feedback Layout Model (VFLM), a self-improving framework that leverages visual feedback iterative refinement. VFLM is capable of performing adaptive reflective generation, which leverages visual information to reflect on previous issues and iteratively generates outputs until satisfactory quality is achieved. It is achieved through reinforcement learning with a visually grounded reward model that incorporates OCR accuracy. By rewarding only the final generated outcome, we can effectively stimulate the model's iterative and reflective generative capabilities. Experiments across multiple benchmarks show that VFLM consistently outperforms advanced MLLMs, existing layout models, and code-only baselines, establishing visual feedback as critical for design-oriented MLLMs. Our code and data are available at https://github.com/FolSpark/VFLM.

Reinforcement learning with verifiable rewards (RLVR) has shown promise for GUI automation, enabling agents to learn from binary task completion signals. However, when task difficulty exceeds model capacity, on-policy exploration fails to discover correct actions, creating zero-advantage traps that eliminate learning signals. While incorporating off-policy expert demonstrations seems intuitive, it causes persistent high-entropy states due to distribution mismatch, disrupting effective learning. We propose GUI-SAGE, a self-explanation framework that generates in-distribution reasoning trajectories for GUI automation. By conditioning on ground-truth actions, our method produces in-distribution guidance that avoids the confusion caused by out-of-distribution expert demonstrations. We further introduce Entropy-Modulated Credit Assignment, which recalibrates learning weights by jointly considering prediction confidence and reward signals, enabling amplified updates for confident correct actions and attenuated updates for uncertain explorations. Extensive experiments on AndroidControl and GUI-Odyssey demonstrate that GUI-SAGE-3B achieves competitive performance with 81.1% success rate, substantially outperforming existing methods. Our analysis validates that self-explanations maintain stable learning dynamics while expert demonstrations cause entropy collapse, and that entropy modulation provides the largest improvements on in-distribution samples.

Deformable image registration (DIR) remains a fundamental yet challenging problem in medical image analysis, largely due to the prohibitively high-dimensional deformation space of dense displacement fields and the scarcity of voxel-level supervision. Existing reinforcement learning frameworks often project this space into coarse, low-dimensional representations, limiting their ability to capture spatially variant deformations. We propose MorphSeek, a fine-grained representation-level policy optimization paradigm that reformulates DIR as a spatially continuous optimization process in the latent feature space. MorphSeek introduces a stochastic Gaussian policy head atop the encoder to model a distribution over latent features, facilitating efficient exploration and coarse-to-fine refinement. The framework integrates unsupervised warm-up with weakly supervised fine-tuning through Group Relative Policy Optimization, where multi-trajectory sampling stabilizes training and improves label efficiency. Across three 3D registration benchmarks (OASIS brain MRI, LiTS liver CT, and Abdomen MR-CT), MorphSeek achieves consistent Dice improvements over competitive baselines while maintaining high label efficiency with minimal parameter cost and low step-level latency overhead. Beyond optimizer specifics, MorphSeek advances a representation-level policy learning paradigm that achieves spatially coherent and data-efficient deformation optimization, offering a principled, backbone-agnostic, and optimizer-agnostic solution for scalable visual alignment in high-dimensional settings.

Learning methods using synthetic data have attracted attention as an effective approach for increasing the diversity of training data while reducing collection costs, thereby improving the robustness of model discrimination. However, many existing methods improve robustness only indirectly through the diversification of training samples and do not explicitly teach the model which regions in the input space truly contribute to discrimination; consequently, the model may learn spurious correlations caused by synthesis biases and artifacts. Motivated by this limitation, this paper proposes a learning framework that uses provenance information obtained during the training data synthesis process, indicating whether each region in the input space originates from the target object, as an auxiliary supervisory signal to promote the acquisition of representations focused on target regions. Specifically, input gradients are decomposed based on information about target and non-target regions during synthesis, and input gradient guidance is introduced to suppress gradients over non-target regions. This suppresses the model's reliance on non-target regions and directly promotes the learning of discriminative representations for target regions. Experiments demonstrate the effectiveness and generality of the proposed method across multiple tasks and modalities, including weakly supervised object localization, spatio-temporal action localization, and image classification.

Vision-language models (VLMs) excel at multimodal understanding, yet their text-only decoding forces them to verbalize visual reasoning, limiting performance on tasks that demand visual imagination. Recent attempts train VLMs to render explicit images, but the heavy image-generation pre-training often hinders the reasoning ability. Inspired by the way humans reason with mental imagery--the internal construction and manipulation of visual cues--we investigate whether VLMs can reason through interleaved multimodal trajectories without producing explicit images. To this end, we present a Machine Mental Imagery framework, dubbed as "Mirage", which augments VLM decoding with latent visual tokens alongside ordinary text. Concretely, whenever the model chooses to "think visually", it recasts its hidden states as next tokens, thereby continuing a multimodal trajectory without generating pixel-level images. Begin by supervising the latent tokens through distillation from ground-truth image embeddings, we then switch to text-only supervision to make the latent trajectory align tightly with the task objective. A subsequent reinforcement learning stage further enhances the multimodal reasoning capability. Experiments on diverse benchmarks demonstrate that \Model unlocks stronger multimodal reasoning without explicit image generation.

Reward models are critical for aligning vision-language systems with human preferences, yet current approaches suffer from hallucination, weak visual grounding, and an inability to use tools for verification, limiting their reliability on complex multimodal reasoning tasks.We present **ARM-Thinker**, an **A**gentic multimodal **R**eward **M**odel that autonomously invokes external tools (e.g., image cropping, doc page retrieval) to ground judgments in verifiable evidence, replacing static, non-interactive reward scoring.This enables the model to verify fine-grained visual details, cross-reference multi-page evidence, and validate reasoning claims, which are capabilities absent in existing reward models.We train ARM-Thinker with multi-stage reinforcement learning, jointly optimizing tool-calling decisions and judgment accuracy.To evaluate agentic reward modeling, we introduce **ARMBench-VL**, comprising three benchmarks that assess fine-grained visual grounding (image-level tools), multi-page document understanding (retrieval tools), and instruction following (text-level verification).ARM-Thinker achieves +16.2% average improvement on reward modeling benchmarks, +9.6% on tool-use tasks, and outperforms baselines on multimodal math and logical reasoning benchmarks.Our results demonstrate that agentic capabilities significantly enhance both accuracy and interpretability of reward models.

Inspired by the success of Reinforcement Learning with Human Feedback (RLHF) in image generation, recent work has adapted reward-based learning to image super-resolution (ISR) by using Image Quality Assessment (IQA) models as rewards. However, existing IQA models typically output only a single global score and are insensitive to local, fine-grained distortions, allowing perceptually undesirable artifacts to obtain spuriously high rewards and leading to reward hacking. To address this issue, we propose FinPercep-RM, a fine-grained perceptual reward model built on an encoder-decoder architecture that predicts both a global quality score and a Perceptual Degradation Map for spatially localizing and quantifying local defects. We further introduce FGR-30k, a dataset containing diverse and subtle distortions produced by real-world super-resolution models, to train the reward model. While FinPercep-RM provides stronger supervision, its increased complexity also makes generator policy learning unstable. We therefore develop a Co-evolutionary Curriculum Learning (CCL) strategy, in which the reward model and the ISR model evolve synchronously: the reward signal progressively increases in complexity, while the ISR model starts with simple global supervision for fast convergence and gradually transitions to fine-grained rewards. This easy-to-hard design stabilizes training and suppresses reward hacking. Extensive experiments across multiple ISR models demonstrate improvements in both global quality and local realism. Code will be available at https://github.com/lyd-2022/FinPercep-RM.

Extreme low-light Raw image restoration remains challenging due to overwhelming noise and severe detail loss.In this paper, we exploit the potential of the dual-exposure setting for this severely ill-posed problem.Existing methods suffer from unreliable cross-exposure alignment, resulting in degraded detail recovery and compromised color fidelity. To address these challenges, we propose RawMetaDiff, a novel generative diffusion framework that restores a high-fidelity Raw image from a short-exposure input, conditioned on a potentially misaligned long-exposure reference under the guidance of Raw metadata.At its core, we propsed two complementary mechanisms: the Meta-Assistant Color Transfer (MACT) enforces color consistency by aligning global color statistics along the channel dimension,while the Meta-Normed Cross Attention (MNCA) leverages Raw metadata to establish robust cross-exposure spatial correspondences and inject shadow details.To support robust diffusion training, we first collect a 1K real-world, dual-exposure Raw dataset, namely DERaw, and then design a realistic degradation model to synthesize data that closely approximates real-world conditions.Extensive experiments on both synthetic and real-world datasets demonstrate that RawMetaDiff significantly outperforms existing methods, justifying an effective new solution for extreme low-light Raw image restoration from the generative perspective.

To enhance the perception and reasoning capabilities of multimodal large language models in complex visual scenes, recent research has introduced agent-based workflows. In these works, MLLMs autonomously utilize image cropping tool to analyze regions of interest for question answering. While existing training strategies, such as those employing supervised fine-tuning and reinforcement learning, have made significant progress, our empirical analysis reveals a key limitation. We demonstrate the model's strong reliance on global input and its weak dependence on the details within the cropped region. To address this issue, we propose a novel two-stage reinforcement learning framework that does not require trajectory supervision. In the first stage, we introduce the "Information Gap" mechanism by adjusting the granularity of the global image. This mechanism trains the model to answer questions by focusing on cropped key regions, driven by the information gain these regions provide. The second stage further enhances cropping precision by incorporating a grounding loss, using a small number of bounding box annotations. Experiments show that our method significantly enhances the model's attention to cropped regions, enabling it to achieve state-of-the-art performance on high-resolution visual question-answering benchmarks. Our method provides a more efficient approach for perceiving and reasoning fine-grained details in MLLMs. Code is available at: https://github.com/XuanPu-Z/LFPC.

In this paper, we propose BuildingGPT, a novel auto-regressive model for building wireframe reconstruction from point clouds with reinforcement learning.Unlike prior works based on detection or diffusion models, BuildingGPT reformulates the building wireframe reconstruction task into a sequence prediction problem.Based on the hierarchical building wireframe tokenization, the wireframe sequences are organized in a structurally- and semantically-aware order for the next-token prediction.The point cloud encoder first transforms the input point cloud into a fixed-length latent code that serves as the starting of the sequence.Then, BuildingGPT auto-regressively predicts tokens conditioned on the latent code and previously generated tokens.With token sequence predicted, the building wireframe is obtained through detokenization.To enhance the model performance, we adopt a two-stage training paradigm including the pre-training and post-training.After the auto-regressive pre-training, Direct Preference Optimization (DPO) is employed as a post-training strategy to align reconstruction results with human preferences.Extensive experiments on the large-scale MunichWF dataset show that BuildingGPT outperforms existing state-of-the-art methods.We commit to release the code and dataset.

Recent advances in visual generation have increasingly explored the integration of reasoning capabilities. They incorporate textual reasoning, i.e., think, either before (as pre-planning) or after (as post-refinement) the generation process, yet they lack on-the-fly multimodal interaction during the generation itself. In this preliminary study, we introduce Thinking-while-Generating (TwiG), the first interleaved framework that enables co-evolving textual reasoning throughout the visual generation process. As visual content is progressively generating, textual reasoning is interleaved to both guide upcoming local regions and reflect on previously synthesized ones. This dynamic interplay produces more context-aware and semantically rich visual outputs. To unveil the potential of this framework, we investigate three candidate strategies, zero-shot prompting, supervised fine-tuning (SFT) on our curated TwiG-50K dataset, and reinforcement learning (RL) via a customized TwiG-GRPO strategy, each offering unique insights into the dynamics of interleaved reasoning. We hope this work inspires further research into interleaving textual reasoning for enhanced visual generation. Code is released at: https://github.com/ZiyuGuo99/Thinking-while-Generating.

This paper does not introduce a novel method but instead establishes a straightforward, incremental, yet essential baseline for video temporal grounding (VTG), a core capability in video understanding. While multimodal large language models (MLLMs) excel at various video understanding tasks, the recipes for optimizing them for VTG remain under-explored. In this paper, we present TimeLens, a systematic investigation into building MLLMs with strong VTG ability, along two primary dimensions: data quality and algorithmic design. We first expose critical quality issues in existing VTG benchmarks and introduce TimeLens-Bench, comprising meticulously re-annotated versions of three popular benchmarks with strict quality criteria. Our analysis reveals dramatic model re-rankings compared to legacy benchmarks, confirming the unreliability of prior evaluation standards. We also address noisy training data through an automated re-annotation pipeline, yielding TimeLens-100K, a large-scale, high-quality training dataset. Building on our data foundation, we conduct in-depth explorations of algorithmic design principles, yielding a series of meaningful insights and effective yet efficient practices. These include interleaved textual encoding for time representation, a thinking-free reinforcement learning with verifiable rewards (RLVR) approach as the training paradigm, and carefully designed recipes for RLVR training. These efforts culminate in TimeLens models, a family of MLLMs with state-of-the-art VTG performance among open-source models and even surpass proprietary models such as GPT-5 and Gemini-2.5-Flash. All codes, data, and models will be released to facilitate future research.

Flow-GRPO successfully applies reinforcement learning to flow models, but uses uniform credit assignment across all steps. This ignores the temporal structure of diffusion generation: early steps determine composition and content (low-frequency structure), while late steps resolve details and textures (high-frequency details). Moreover, assigning uniform credit based solely on the final image can inadvertently reward suboptimal intermediate steps, especially when errors are corrected later in the diffusion trajectory. We propose Stepwise-Flow-GRPO, which assigns credit based on each step's reward improvement. By leveraging Tweedie's formula to obtain intermediate reward estimates and introducing gain-based advantages, our method achieves superior sample efficiency and faster convergence. We also introduce a DDIM-inspired SDE that improves reward quality while preserving stochasticity for policy gradients.

Offline Multi-Agent Reinforcement Learning (MARL) is critical for coordinating multiple agents in costly and unsafe environments, yet existing methods struggle with high sensitivity to reward functions and weak generalization to new goals, limiting its practical impact. Inspired by single-agent Offline Goal-Conditioned RL (OGCRL), we propose the first goal-conditioned offline MARL framework, extending OGCRL to multi-agent settings under both fully decentralized and centralized training with decentralized execution (CTDE) paradigms. To systematically evaluate this setting, we introduce MangoBench, the first fully cooperative multi-goal benchmark for MARL, covering 3 environments, 4 agent types, and 47 tasks, designed to assess joint-control locomotion, synchronous and asynchronous bimanual manipulation, and robustness to high-dimensional inputs. Extensive experiments demonstrate that our baselines achieve strong multi-goal generalization under sparse rewards, yet no method dominates all tasks, revealing both the intrinsic complexity and the unexplored potential of goal-conditioned offline MARL.

Thinking with images has emerged as an effective paradigm for advancing visual reasoning, extending beyond text-only chains of thought by injecting visual evidence into intermediate reasoning steps. However, existing methods fall short of human-like abstract visual thinking, as their flexibility is fundamentally limited by external tools. In this work, we introduce Monet, a training framework that enables multimodal large language models (MLLMs) to reason directly within the latent visual space by generating continuous embeddings that function as intermediate visual thoughts. We identify two core challenges in training MLLMs for latent visual reasoning--high computational cost in latent-vision alignment and insufficient supervision over latent embeddings, and address them with a three-stage distillation-based supervised fine-tuning (SFT) pipeline. We further reveal a limitation of applying GRPO to latent reasoning: it primarily enhances text-based reasoning rather than latent reasoning. To overcome this, we propose VLPO (Visual-latent Policy Optimization), a reinforcement learning method that explicitly incorporates latent embeddings into policy gradient updates.To support SFT, we construct Monet-SFT-125K, a high-quality text-image interleaved CoT dataset containing 125K real-world, chart, OCR, and geometry CoTs. Our model, Monet-7B, shows consistent gains across real-world perception and reasoning benchmarks and exhibits strong out-of-distribution generalization on challenging abstract visual reasoning tasks. We also empirically analyze the role of each training component and discuss our early unsuccessful attempts, providing insights for future developments in visual latent reasoning.

Unified multimodal models significantly improve visual generation by combining vision-language models (VLMs) with diffusion models. However, existing methods struggle to fully balance sufficient interaction and flexible implementation due to vast representation difference. Considering abundant and hierarchical information in VLM's layers from low-level details to high-level semantics, we propose ParaUni. It extracts features from variants VLM's layers in a Parallel way for comprehensive information interaction and retains a flexible separation architecture to enhance generation in Unified multimodal model. Concretely, visual features from all VLM's layers are fed in parallel into a Layer Integration Module (LIM), which efficiently integrates fine-grained details and semantic abstractions and provides the fused representation as a condition to the diffusion model. To further enhance performance, we reveal that these hierarchical layers respond unequally to different rewards in Reinforcement Learning (RL). Crucially, we design a Layer-wise Dynamic Adjustment Mechanism (LDAM) to facilitate multiple reward improvements that aligns the hierarchical properties of these layers using RL. Extensive experiments show ParaUni leverages complementary multi-layer features to substantially improve generation quality and shows strong potential for multiple reward advances during RL stages.