RL for VLMs

Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a prevailing paradigm for enhancing reasoning in Multimodal Large Language Models (MLLMs). However, relying solely on outcome supervision risks reward hacking, where models learn spurious reasoning patterns to satisfy final answer checks. While recent rubric-based approaches offer fine-grained supervision signals, they suffer from high computational costs of instance-level generation and inefficient training dynamics caused by treating all rubrics as equally learnable. In this paper, we propose Stratified Rubric-based Curriculum Learning (RuCL), a novel framework that reformulates curriculum learning by shifting the focus from data selection to reward design. RuCL generates generalized rubrics for broad applicability and stratifies them based on the model's competence. By dynamically adjusting rubric weights during training, RuCL guides the model from mastering foundational perception to tackling advanced logical reasoning. Extensive experiments on various visual reasoning benchmarks show that RuCL yields a remarkable +7.83% average improvement over the Qwen2.5-VL-7B model, achieving a state-of-the-art accuracy of 60.06%.

Vision-Language-Action (VLA) models are advancing autonomous driving by replacing modular pipelines with unified end-to-end architectures. However, current VLAs face two expensive requirements: (1) massive dataset collection, and (2) dense reasoning annotations. In this work, we address both challenges with NORD (No Reasoning for Driving). Compared to existing VLAs, NORD achieves competitive performance while being fine-tuned on <60% of the data and no reasoning annotations, resulting in 3x fewer tokens. We identify that standard Group Relative Policy Optimization (GRPO) fails to yield significant improvements when applied to policies trained on such small, reasoning-free datasets. We show that this limitation stems from difficulty bias, which disproportionately penalizes reward signals from scenarios that produce high-variance rollouts within GRPO. NORD overcomes this by incorporating Dr. GRPO, a recent algorithm designed to mitigate difficulty bias in LLMs. As a result, NORD achieves competitive performance on Waymo and NAVSIM with a fraction of the training data and no reasoning overhead, enabling more efficient autonomous systems. Website: https://nord-vla-ai.github.io/

Reinforcement learning for agentic multimodal models often suffers from interaction collapse, where models learn to reduce tool usage and multi-turn reasoning, limiting the benefits of agentic behavior. We introduce PyVision-RL, a reinforcement learning framework for open-weight multimodal models that stabilizes training and sustains interaction. Our approach combines an oversampling-filtering-ranking rollout strategy with an accumulative tool reward to prevent collapse and encourage multi-turn tool use. Using a unified training pipeline, we develop PyVision-Image and PyVision-Video for image and video understanding. For video reasoning, PyVision-Video employs on-demand context construction, selectively sampling task-relevant frames during reasoning to significantly reduce visual token usage. Experiments show strong performance and improved efficiency, demonstrating that sustained interaction and on-demand visual processing are critical for scalable multimodal agents.

Reinforcement Learning with verifiable rewards (RLVR) has emerged as a primary learning paradigm for enhancing the reasoning capabilities of multi-modal large language models (MLLMs). However, during RL training, the enormous state space of MLLM and sparse rewards often leads to entropy collapse, policy degradation, or over-exploitation of suboptimal behaviors. This necessitates an exploration strategy that maintains productive stochasticity while avoiding the drawbacks of uncontrolled random sampling, yielding inefficient exploration. In this paper, we propose CalibRL, a hybrid-policy RLVR framework that supports controllable exploration with expert guidance, enabled by two key mechanisms. First, a distribution-aware advantage weighting scales updates by group rareness to calibrate the distribution, therefore preserving exploration. Meanwhile, the asymmetric activation function (LeakyReLU) leverages the expert knowledge as a calibration baseline to moderate overconfident updates while preserving their corrective direction. CalibRL increases policy entropy in a guided manner and clarifies the target distribution by estimating the on-policy distribution through online sampling. Updates are driven by these informative behaviors, avoiding convergence to erroneous patterns. Importantly, these designs help alleviate the distributional mismatch between the model's policy and expert trajectories, thereby achieving a more stable balance between exploration and exploitation. Extensive experiments across eight benchmarks, including both in-domain and out-of-domain settings, demonstrate consistent improvements, validating the effectiveness of our controllable hybrid-policy RLVR training. Code is available at https://github.com/zhh6425/CalibRL.

Long video understanding is challenging due to rich and complicated multimodal clues in long temporal range.Current methods adopt reasoning to improve the model's ability to analyze complex video clues in long videos via text-form reasoning.However,the existing literature suffers from the fact that the text-only reasoning under fixed video context may exacerbate hallucinations since detailed crucial clues are often ignored under limited video context length due to the temporal redundancy of long videos.To address this gap,we propose Video-TwG,a curriculum reinforced framework that employs a novel Think-with-Grounding paradigm,enabling video LLMs to actively decide when to perform on-demand grounding during interleaved text-video reasoning, selectively zooming into question-relevant clips only when necessary.Video-TwG can be trained end-to-end in a straightforward manner, without relying on complex auxiliary modules or heavily annotated reasoning tracesIn detail,we design a Two-stage Reinforced Curriculum Strategy, where the model first learns think-with-grounding behavior on a small short-video GQA dataset with grounding labels,and then scales to diverse general QA data with videos of diverse domains to encourage generalization. Further, to handle complex think-with-grounding reasoning for various kinds of data,we propose TwG-GRPO algorithm which features the fine-grained grounding reward, self-confirmed pseudo reward and accuracy-gated mechanism.Finally,we propose to construct a new TwG-51K dataset that facilitates training. Experiments on Video-MME, LongVideoBench, and MLVU show that Video-TwG consistently outperforms strong LVU baselines.Further ablation validates the necessity of our Two-stage Reinforced Curriculum Strategy and shows our TwG-GRPO better leverages diverse unlabeled data to improve grounding quality and reduce redundant groundings without sacrificing QA performance.

Video reasoning requires a fine-grained understanding of the temporal dependencies and event-level relations between objects and events in videos. Current Multimodal Large Language Models (MLLMs) are prone to severe temporal hallucinations in video reasoning. An underlying cause of these hallucinations is weak visual-temporal grounding and the lack of explicit structure for modelling event relations. Models often rely on auxiliary text, such as dense captions, rather than explicitly anchoring their reasoning in actual visual evidence. However, these textual representations are inherently unstructured and fail to provide explicit causal constraints needed to guide the model's reasoning. In this work, we propose GraphThinker, a reinforcement finetuning method that constructs a structured event representation of a video and enforces visual grounding to jointly reduce reasoning hallucinations. Specifically, we employ an MLLM to construct an Event-based Video Scene Graph (EVSG) that captures both intra- and inter-event relations, guiding a structured video reasoning process. Moreover, we address the weak grounding issue by introducing a novel visual attention reward during reinforcement finetuning that encourages the model to actively attend to reliable visual cues. On the RexTime dataset, GraphThinker achieves an over 4% improvement in IoU=0.3 for moment localisation. On the VidHalluc dataset, GraphThinker achieves a 9.8% improvement in reducing temporal sequence hallucination and a 7.6% gain in Binary QA in reducing action hallucination, compared to the state-of-the-art methods.

Fine-grained visual reasoning in multimodal large language models (MLLMs) is bottlenecked by single-pass global image encoding: key evidence often lies in tiny objects, cluttered regions, subtle markings, or dense charts. We present TikArt (Thinking Aperture), an aperture-guided agent that formulates multimodal reasoning as sequential evidence acquisition over regions of interest. TikArt follows a Think--Aperture--Observe (TAO) loop that interleaves language reasoning with two aperture actions: Zoom, which extracts rectangular crops, and Segment, which invokes an off-the-shelf segmenter to produce object-centric mask-based views for irregular targets. A mandatory Observation step after every aperture action writes local evidence back into text, yielding interpretable aperture trajectories and persistent linguistic memory. Built on Qwen3-VL-8B, TikArt is trained with GRPO-style reinforcement learning under a two-stage curriculum. To stabilize long-horizon tool-integrated learning, we introduce Relative Uncertainty Reduction (RUR), a dense reward computed by a frozen evaluator that favors evidence-building trajectories and mitigates degenerate tool use. Experiments on high-resolution reasoning, general multimodal understanding, and both referring and reasoning-oriented segmentation show consistent gains over the backbone, demonstrating that aperture-guided observation improves fine-grained visual reasoning and transfers naturally to pixel-level grounding.

Reinforcement learning (RL) promises to unlock capabilities beyond imitation learning for Vision--Language--Action (VLA) models, but its requirement for massive real-world interaction prevents direct deployment on physical robots. Recent work attempts to use learned world models as simulators for policy optimization, yet closed-loop imagined rollouts inevitably suffer from hallucination and long-horizon error accumulation. Such errors not only degrade visual fidelity, but also mislead policy optimization by providing unreliable learning signals. We propose WoVR, a reliable world-model-based RL framework for post-training VLA policies. Instead of assuming a faithful world model, WoVR explicitly regulates how RL interacts with imperfect imagined dynamics. It improves rollout stability through a controllable action-conditioned video world model, reshapes imagined interaction to reduce effective error depth via Keyframe-Initialized Rollouts, and maintains policy--simulator alignment through World Model-Policy co-evolution. Extensive experiments demonstrate that WoVR enables stable long-horizon imagined rollouts and effective policy optimization, achieving superior LIBERO performance and consistent real-world gains across multiple robotic platforms. These results show that world models can serve as practical simulators for RL when hallucination is explicitly controlled. Additional visualization results are available at https://wovr-corl.github.io.

Simulation offers a scalable and low-cost way to enrich vision-language-action (VLA) training, reducing reliance on expensive real-robot demonstrations. However, most sim-real co-training methods rely on supervised fine-tuning (SFT), which treats simulation as a static source of demonstrations and does not exploit large-scale closed-loop interaction. Consequently, real-world gains and generalization are often limited. In this paper, we propose an RL-based sim-real Co-training (RL-Co) framework that leverages interactive simulation while preserving real-world capabilities. Our method follows a generic two-stage design: we first warm-start the policy with SFT on a mixture of real and simulated demonstrations, then fine-tune it with reinforcement learning in simulation while adding an auxiliary supervised loss on real-world data to anchor the policy and mitigate catastrophic forgetting. We evaluate our framework on four real-world tabletop manipulation tasks using two representative VLA architectures, OpenVLA and $π_{0.5}$, and observe consistent improvements over real-only fine-tuning and SFT-based co-training, including +24% real-world success on OpenVLA and +20% on $π_{0.5}$. Beyond higher success rates, RL co-training yields stronger generalization to unseen task variations and substantially improved real-world data efficiency, providing a practical and scalable pathway for leveraging simulation to enhance real-robot deployment.

This paper develops LongNav-R1, an end-to-end multi-turn reinforcement learning (RL) framework designed to optimize Visual-Language-Action (VLA) models for long-horizon navigation. Unlike existing single-turn paradigm, LongNav-R1 reformulates the navigation decision process as a continuous multi-turn conversation between the VLA policy and the embodied environment. This multi-turn RL framework offers two distinct advantages: i) it enables the agent to reason about the causal effects of historical interactions and sequential future outcomes; and ii) it allows the model to learn directly from online interactions, fostering diverse trajectory generation and avoiding the behavioral rigidity often imposed by human demonstrations. Furthermore, we introduce Horizon-Adaptive Policy Optimization. This mechanism explicitly accounts for varying horizon lengths during advantage estimation, facilitating accurate temporal credit assignment over extended sequences. Consequently, the agent develops diverse navigation behaviors and resists collapse during long-horizon tasks. Experiments on object navigation benchmarks validate the framework's efficacy: With 4,000 rollout trajectories, LongNav-R1 boosts the Qwen3-VL-2B success rate from 64.3% to 73.0%. These results demonstrate superior sample efficiency and significantly outperform state-of-the-art methods. The model's generalizability and robustness are further validated by its zero-shot performance in long-horizon real-world navigation settings. All source code is open-sourced at https://github.com/UMich-CURLY/LongNav-R1.

While chain-of-thought (CoT) reasoning has substantially improved multimodal large language models (MLLMs) on complex reasoning tasks, existing approaches largely rely on long textual reasoning trajectories and provide limited mechanisms for learning stable visual attention policies. Our analysis shows that current MLLMs exhibit weak visual focus: early-stage visual misalignment is rarely corrected during subsequent reasoning, leading to error propagation and failed inferences. We argue that this limitation stems from inadequate credit assignment for visual attention during training. To address this issue, we propose SAYO, a visual reasoning model trained with a reinforcement learning (RL) framework that introduces a region-level visual attention-based reward. This reward explicitly aligns optimization signals with visually grounded reasoning steps, enabling the model to learn more reliable attention behaviors. Extensive experiments across multiple multimodal benchmarks demonstrate that SAYO consistently improves performance on diverse reasoning and perception tasks. Project is open at https://github.com/Cratileo/Sayo.

Real-world visual question answering (VQA) may be context-dependent: an image-question pair could be under-specified, such that the correct answer depends on external information that is not observable in the image. In such cases, directly answering can lead to confident but incorrect predictions. We propose Clarify-or-Answer (CoA), an ask-or-answer agent that separately models the decision to ask or answer, and what to ask if needed. CoA first determines whether clarification is necessary; if so, it generates one or more clarification questions to iteratively enrich the context until ambiguity is resolved. To validate this framework, we introduce CONTEXTCLARIFY, a dataset of ambiguous VQA questions along with a contrast set of non-ambiguous questions. We further introduce GRPO-CR(Clarification Reasoning), a reinforcement learning approach that optimizes clarification question generation with multiple reward signals encouraging well-formed, focused, non-trivial questions that resolve ambiguity. Across three VLMs and three datasets, CoA achieves consistent improvements at both the module and system levels, with a single clarification turn providing substantial gains and recursive multi-turn execution offering modest additional improvements in select settings.

While Vision-Language Models (VLMs) excel at 2D visual understanding, they remain constrained by 2D-centric paradigm that severely limits genuine 3D spatial reasoning. To bridge this gap, we introduce Think3D, a novel framework that equips VLM agents with interactive, 3D chain-of-thought reasoning capabilities. By integrating a suite of 3D manipulation tools, Think3D transforms perception into active spatial exploration, mirroring human geometric reasoning. Think3D consistently improves proprietary models, including GPT-4.1 and Gemini 2.5 Pro, across BLINK Multi-view, MindCube-1K, and VSI-Bench-Tiny. We further propose Think3D-RL to teach smaller open-weight models how to manipulate 3D space effectively. Using only final-answer rewards, without process supervision or handcrafted exploration trajectories, Think3D-RL enables Qwen3-VL-4B to autonomously learn effective 3D exploration strategies. After training, the model exhibits tool-use patterns similar to those of stronger proprietary models, while shifting the effect of 3D tool use on MindCube-1K from a performance drop to a substantial improvement. These results show that active exploration in 3D space provides an effective and general paradigm for improving spatial reasoning in multimodal agents. Code, models, and data are available at https://github.com/zhangzaibin/spagent.

Vision-language models achieve strong performance across a wide range of multimodal understanding and reasoning tasks, yet their multi-step reasoning remains unstable. Repeated sampling over the same input often produces divergent reasoning trajectories and inconsistent final predictions. To address this, we introduce two complementary approaches inspired by test-time scaling: (1) CASHEW, an inference-time framework that stabilizes reasoning by iteratively aggregating multiple candidate trajectories into higher-quality reasoning traces, with visual verification of object presence providing grounding cues during aggregation, and (2) CASHEW-RL, a learned variant that internalizes this aggregation behavior within a single model. CASHEW-RL is trained using Group Sequence Policy Optimization (GSPO) with a composite reward that encourages correct answers grounded in minimal yet sufficient visual evidence, while adaptively allocating reasoning effort based on task difficulty. This training objective enables robust self-aggregation at inference. Extensive experiments on 13 image understanding, video understanding, and video reasoning benchmarks show significant performance improvements, with CASHEW/CASHEW-RL achieving average gains of +4.1/+5.3 percentage points on Qwen3-VL-4B and +2.6/+3.0 percentage points on Qwen3-VL-8B, respectively.

While Multimodal Large Language Models (MLLMs) excel at single-image understanding, they exhibit significantly degraded performance in multi-image reasoning scenarios. Multi-image reasoning presents fundamental challenges including complex inter-relationships between images and scattered critical information across image sets. Inspired by human cognitive processes, we propose a Cognition-Inspired Meta-Action Framework (CINEMA), which decomposes multi-image reasoning into five structured meta-actions: Global, Focus, Hint, Think, and Answer, explicitly modeling the sequential cognitive steps humans naturally employ. For cold-start training, we introduce a Retrieval-Based Tree Sampling strategy that generates high-quality meta-action trajectories to bootstrap the model with reasoning patterns. During reinforcement learning, we adopt a two-stage paradigm: an exploration phase with Diversity-Preserving Strategy to avoid entropy collapse, followed by an annealed exploitation phase with DAPO to gradually strengthen exploitation. To train our model, we construct a dataset of 56k cold-start and 58k reinforcement learning instances spanning multi-image, multi-frame, and single-image tasks. We conduct extensive evaluations on multi-image reasoning benchmarks, video understanding benchmarks, and single-image benchmarks, achieving competitive state-of-the-art performance on several key benchmarks. Our model surpasses GPT-4o on the MUIR and MVMath benchmarks and notably outperforms specialized video reasoning models on video understanding benchmarks, demonstrating the effectiveness and generalizability of our human cognition-inspired reasoning framework.

Multimodal Large Language Models (MLLMs) have substantially advanced Radiology Report Generation (RRG), yet aligning them through reinforcement learning (RL) remains challenging due to heterogeneous medical supervision. Vanilla Group Relative Policy Optimization (GRPO) assigns uniform credit across the entire generation, leading to segment interference, token dilution, and rationale-diagnosis decoupling, which exacerbates clinical hallucinations. We propose HERO (Hierarchical rationalE Reasoning Optimization), a factorized policy optimization framework that aligns heterogeneous supervision with three optimization granularities. HERO separately optimizes reasoning, diagnosis, and rationale faithfulness through complementary segment-, token-, and completion-level optimization with a heterogeneous reward formulation covering diagnostic accuracy, reasoning quality, and think-answer consistency. Experiments on MIMIC-CXR and IU-Xray show that HERO outperforms strong supervised and reinforcement learning baselines, achieving state-of-the-art clinical efficacy while producing more rationale-supported and think-answer-consistent reports, thereby substantially mitigating clinical hallucinations.

Generating humorous memes is a challenging multimodal task that moves beyond direct image-to-caption supervision. It requires a nuanced reasoning over visual content, contextual cues, and subjective humor. To bridge this gap between visual perception and humorous punchline creation, we propose HUMOR}, a novel framework that guides VLMs through hierarchical reasoning and aligns them with group-wise human preferences. First, HUMOR employs a hierarchical, multi-path Chain-of-Thought (CoT): the model begins by identifying a template-level intent, then explores diverse reasoning paths under different contexts, and finally anchors onto a high-quality, context-specific path. This CoT supervision, which traces back from ground-truth captions, enhances reasoning diversity. We further analyze that this multi-path exploration with anchoring maintains a high expected humor quality, under the practical condition that high-quality paths retain significant probability mass. Second, to capture subjective humor, we train a pairwise reward model that operates within groups of memes sharing the same template. Following established theory, this approach ensures a consistent and robust proxy for human preference, even with subjective and noisy labels. The reward model then enables a group-wise reinforcement learning optimization, guaranteeing providing a theoretical guarantee for monotonic improvement within the trust region. Extensive experiments show that HUMOR empowers various VLMs with superior reasoning diversity, more reliable preference alignment, and higher overall meme quality. Beyond memes, our work presents a general training paradigm for open-ended, human-aligned multimodal generation, where success is guided by comparative judgment within coherent output group.

Vision--language models (VLMs) have recently shown promise for assisting clinical reasoning in dermatological diagnosis. However, their trustworthiness and clinical utility remain limited by three key challenges: heterogeneous datasets with inconsistent diagnostic labels and concept annotations, the lack of grounded diagnostic rationales for reliable reasoning supervision, and limited scalability when transferring knowledge from small, densely annotated datasets to large collections with sparse labels. To address these challenges, we propose Skin-R1, a dermatology-oriented VLM that integrates textbook-grounded clinical reasoning supervision with reinforcement learning (RL) to improve the accuracy and robustness of diagnostic prediction. First, we construct a textbook-based reasoning generator that synthesizes hierarchy-aware and differential-diagnosis (DDx) diagnostic trajectories derived from authoritative dermatology knowledge. Second, these trajectories are used for supervised fine-tuning (SFT), establishing a clinically grounded reasoning foundation for the model. Finally, we introduce an RL training framework that incorporates the hierarchical structure of dermatological diseases into the reward design, enabling the model to generalize grounded diagnostic reasoning to large-scale datasets with sparse annotations. Extensive experiments across multiple dermatology benchmarks demonstrate that Skin-R1 consistently improves diagnostic accuracy and robustness compared to state-of-the-art Med-VLM baselines. Ablation studies further highlight the critical role of grounded reasoning supervision introduced during the SFT stage.

Multimodal large language models (MLLMs) have achieved remarkable progress in vision-language tasks, but continue to struggle with spatial reasoning. Existing spatial MLLMs rely on large-scale datasets, explicit 3D inputs, architecture-specific modifications, or sparse Reinforcement Learning (RL) methods that provide insufficient guidance for spatially-grounded reasoning. We introduce SpatialThinker. To our knowledge, it is the first MLLM unifying Scene Graph Generation (SGG) and visual reasoning in a single pass via online RL. The model simulates human-like spatial perception by constructing a mental scene graph of task-relevant objects and relations, and reasoning toward an answer via dense spatial rewards. Our contributions are threefold: (1) SGG-grounded reasoning: integrating SGG directly within the reasoning chain rather than as a disjoint preprocessing step; (2) STVQA-7K: a high-quality spatial VQA training dataset via a scalable synthesis pipeline; and (3) a dense spatial reward design that enforces structured grounding during RL and generalizes to improve broad visual perception. SpatialThinker-7B achieves 3.6$\times$ larger gains over SFT and $1.7\times$ better in- and out-of-distribution generalization than sparse RL. Trained on only 7K samples, SpatialThinker-7B matches GPT-5 and outperforms GPT-4o, while SpatialThinker-30B surpasses both GPT-5 and Claude 4 Sonnet on average across 14 spatial and real-world benchmarks, demonstrating that structured spatial grounding with reward-aligned reasoning enables robust spatial understanding with limited data.

Vision-Language-Action (VLA) models such as OpenVLA, Octo, and $π_0$ have shown strong generalization by leveraging large-scale demonstrations, yet their performance is still fundamentally constrained by the quality and coverage of supervised data. Reinforcement learning (RL) provides a promising path for improving and fine-tuning VLAs through online interaction. However, conventional policy gradient methods are computationally infeasible in the context of flow-matching based models due to the intractability of the importance sampling process, which requires explicit computation of policy ratios. To overcome this limitation, we propose Flow Policy Optimization (FPO) algorithm, which reformulates importance sampling by leveraging per-sample changes in the conditional flow-matching objective. Furthermore, FPO achieves stable and scalable online reinforcement fine-tuning of the $π_0$ model by integrating structure-aware credit assignment to enhance gradient efficiency, clipped surrogate objectives to stabilize optimization, multi-step latent exploration to encourage diverse policy updates, and a Q-ensemble mechanism to provide robust value estimation. We evaluate FPO on the LIBERO benchmark and the ALOHA simulation task against supervised, preference-aligned, diffusion-based, autoregressive online RL, and $π_0$-FAST baselines, observing consistent improvements over the imitation prior and strong alternatives with stable learning under sparse rewards. In addition, ablation studies and analyses of the latent space dynamics further highlight the contributions of individual components within FPO, validating the effectiveness of the proposed computational modules and the stable convergence of the conditional flow-matching objective during online RL.

Recent advances in large video-language models (VLMs) rely on extensive fine-tuning techniques that strengthen alignment between textual and visual comprehension. Leading pipelines typically pair supervised fine-tuning (SFT) with reinforcement learning from preference data to enhance video comprehension. However, as VLMs scale in parameter size, so does the cost of gathering enough human feedback. To make fine-tuning more cost-effective, recent frameworks explore reinforcement learning with AI feedback (RLAIF), which replace human preference with AI as a judge. Current RLAIF frameworks rely on a specialized reward model trained with video narratives to create calibrated scalar rewards -- an expensive and restrictive pipeline. We propose Oracle-RLAIF, a novel framework that replaces the trained reward model with a more general Oracle ranker which acts as a drop-in model ranking candidate model responses rather than scoring them. Alongside Oracle-RLAIF, we introduce $GRPO_{rank}$, a novel rank-based loss function based on Group Relative Policy Optimization (GRPO) that directly optimizes ordinal feedback with rank-aware advantages. Empirically, we demonstrate that Oracle-RLAIF consistently outperforms leading VLMs using existing fine-tuning methods when evaluated across various video comprehension benchmarks. Oracle-RLAIF paves the path to creating flexible and data-efficient frameworks for aligning large multi-modal video models with reinforcement learning from rank rather than score.

While reasoning technology like Chain of Thought (CoT) has been widely adopted in Vision Language Action (VLA) models, it demonstrates promising capabilities in end to end autonomous driving. However, recent efforts to integrate CoT reasoning often fall short in simple scenarios, introducing unnecessary computational overhead without improving decision quality. To address this, we propose AdaThinkDrive, a novel VLA framework with a dual mode reasoning mechanism inspired by fast and slow thinking. First, our framework is pretrained on large scale autonomous driving (AD) scenarios using both question answering (QA) and trajectory datasets to acquire world knowledge and driving commonsense. During supervised fine tuning (SFT), we introduce a two mode dataset, fast answering (w/o CoT) and slow thinking (with CoT), enabling the model to distinguish between scenarios that require reasoning. Furthermore, an Adaptive Think Reward strategy is proposed in conjunction with the Group Relative Policy Optimization (GRPO), which rewards the model for selectively applying CoT by comparing trajectory quality across different reasoning modes. Extensive experiments on the Navsim benchmark show that AdaThinkDrive achieves a PDMS of 90.3, surpassing the best vision only baseline by 1.7 points. Moreover, ablations show that AdaThinkDrive surpasses both the never Think and always Think baselines, improving PDMS by 2.0 and 1.4, respectively. It also reduces inference time by 14% compared to the always Think baseline, demonstrating its ability to balance accuracy and efficiency through adaptive reasoning.

Multimodal large language models (MLLMs) have recently demonstrated remarkable reasoning abilities under reinforcement learning (RL) paradigm. However, most existing multimodal medical reasoning models focus on basic reasoning, which refers to shallow inference based on visual feature matching. In contrast, real-world clinical diagnosis extends beyond basic reasoning, demanding complex reasoning that integrates heterogeneous clinical information (such as chief complaints and medical history) with multimodal medical imaging data. To bridge this gap, we introduce MM-Retinal-Reason, an ophthalmic multimodal dataset covering the full spectrum of perception and reasoning. Specifically, it is the first dataset in ophthalmology to encompass both basic and complex reasoning tasks with Chain-of-Thought (CoT) trajectories, aiming to enhance visual-centric reasoning and emulate realistic clinical decision-making. Building upon MM-Retinal-Reason, we propose OphthaReason, the first RL-enhanced ophthalmic multimodal reasoning model with step-by-step reasoning traces. To enable flexible adaptation to both basic and complex reasoning tasks, we further introduce Uncertainty-Aware Dynamic Thinking (UADT), which estimates sample-level uncertainty via entropy and dynamically modulates exploration depth through a shaped advantage mechanism. Comprehensive experiments demonstrate the effectiveness of our model on both basic and complex reasoning tasks, outperforming general-purpose MLLMs, medical MLLMs, RL-based medical MLLMs, and ophthalmic MLLMs by at least 15.47%. Project Page: link.

While Multimodal Large Language Models (MLLMs) demonstrate remarkable capabilities across diverse domains, their application to specialized anomaly detection (AD) remains constrained by domain adaptation challenges. Existing Group Relative Policy Optimization (GRPO) based approaches suffer from two critical limitations: inadequate training data utilization when models produce uniform responses, and insufficient supervision over reasoning processes that encourage immediate binary decisions without deliberative analysis. We propose a comprehensive framework addressing these limitations through two synergistic innovations. First, we introduce a multi-stage deliberative reasoning process that guides models from region identification to focused examination, generating diverse response patterns essential for GRPO optimization while enabling structured supervision over analytical workflows. Second, we develop a fine-grained reward mechanism incorporating classification accuracy and localization supervision, transforming binary feedback into continuous signals that distinguish genuine analytical insight from spurious correctness. Comprehensive evaluation across multiple industrial datasets demonstrates substantial performance improvements in adapting general vision-language models to specialized anomaly detection. Our method achieves superior accuracy with efficient adaptation of existing annotations, effectively bridging the gap between general-purpose MLLM capabilities and the fine-grained visual discrimination required for detecting subtle manufacturing defects and structural irregularities.

Process Reward Models (PRMs) provide fine-grained supervision over intermediate reasoning steps, yet most existing models lack long-range reasoning and deep thinking capabilities. Although prior work has introduced Chain-of-Thought (CoT) to reward models, the high annotation cost of CoT-PRM still limits its scalability. To address these challenges, we propose the Visual Reasoning Process Reward Model (VRPRM) and design a label-efficient two-stage training strategy. Experimental results demonstrate that with only 5.5K CoT-PRM supervised step labels and 40K non-CoT PRM reinforcement learning training examples, VRPRM achieves a macro F1 score of 69.18 on VisualProcessBench, outperforming VisualPRM-8B, a non-thinking PRM trained on 400K examples, by 7.00 points. In Best-of-N evaluation, VRPRM-RL brings absolute Overall improvements of 21.77, 23.42, and 23.12 points on InternVL2.5-8B, 26B, and 38B policies, respectively, with a maximum relative improvement of 59.0%.

Recent progress in large-scale reinforcement learning (RL) has notably enhanced the reasoning capabilities of large language models (LLMs), especially in mathematical domains. However, current multimodal LLMs (MLLMs) for mathematical reasoning often rely on one-to-one image-text pairs and single-solution supervision, overlooking the diversity of valid reasoning perspectives and internal reflections. In this work, we introduce MathV-DP, a novel dataset that captures multiple diverse solution trajectories for each image-question pair, fostering richer reasoning supervision. We further propose Qwen-VL-DP, a model built upon Qwen-VL, fine-tuned with supervised learning and enhanced via group relative policy optimization (GRPO), a rule-based RL approach that integrates correctness discrimination and diversity-aware reward functions. Our method emphasizes learning from varied reasoning perspectives and distinguishing between correct yet distinct solutions. Extensive experiments on the MathVista's minitest and Math-V benchmarks demonstrate that Qwen-VL-DP significantly outperforms prior base MLLMs in both accuracy and generative diversity, highlighting the importance of incorporating diverse perspectives and reflective reasoning in multimodal mathematical reasoning.