RL for VLMs

To evaluate whether vision-language models can reason about geological histories, it is necessary to construct observations for which the underlying process history is known. Furthermore, reasoning over geological histories is not just a question of recognizing visual patterns, but also of understanding temporal and structural relationships that may be only indirectly visible or highly ambiguous. When ground-truth event histories are not uniquely identifiable or are unavailable, it remains an open challenge to teach models capable of visual reasoning to produce valid geological reconstructions that are consistent with both observed evidence and geological principles. We therefore investigate whether defining a verifiable geological reasoning task can improve geological event reconstruction across observation domains through reinforcement learning with verifiable rewards (RLVR). To this end, we present Geo-Strat-RL, a synthetic environment that generates stratigraphic observations and compact visible-evidence event histories. The environment combines a geological generator with an executable verifier that scores chronology, event identity, deposition, and structural relationships. We show that RLVR improves geological reconstruction in vision-language models (VLMs), increasing geological content scores on held out stratigraphic diagrams. We further evaluate the same held-out geological histories in a synthetic seismic observation domain by converting the generated scenes into acoustic-impedance-derived amplitude sections. In this controlled paired-renderer setting, we present evidence that geological reasoning learned from stratigraphic diagram-domain RLVR training transfers to synthetic seismic representations without seismic-specific training examples, supporting the hypothesis that RLVR can teach reusable geological reasoning concepts across related observation formats.

Video large language models (LLMs) are often constrained by computation and memory budgets, leading them to use reduced frame rates and spatial resolutions, which may cause them to miss critical information for question answering (QA). A practical and efficient solution is a two-stage paradigm: first perform coarse video understanding to localize relevant segments, and then re-watch these segments at higher temporal or spatial fidelity. In this paper, we present video-SALMONN-R$^3$, the first end-to-end video-LLM that enables re-watch through reinforcement learning without relying on chain-of-thought (CoT) cold-start. This design removes the need for costly CoT data annotations and avoids CoT-based supervised fine-tuning (SFT), which can otherwise degrade the pretrained video understanding abilities. To address the mismatch between the reasoning-first behavior induced by re-watch and the answer-first tendency of pretrained video-LLMs, we propose a re-answer strategy, in which the model first produces a direct answer in the first watch and then refines it after re-watching. Finally, to improve question adherence during re-watching, we propose a re-ask mechanism that re-injects the query when revisiting localized segments. Experimental results show that video-SALMONN-R$^3$ consistently outperforms both the base model and the QA-SFT baseline, while surpassing prior re-watch-based approaches with significantly lower computational cost. Code, models, and data will be publicly released upon acceptance.

While Vision-Language Models (VLMs) show great promise in volumetric medical report generation, they frequently suffer from visual hallucinations and a lack of grounding in 3D CT data. Current Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) strategies typically optimize text fidelity alone, essentially rewarding correct diagnoses derived from language priors rather than genuine visual perception. To address this, we propose cross-view aligned Evidence-driven Multimodal Reinforcement Learning (Evidence-MRL, noted as E-MRL), a reliable RL reasoning framework that formulates the generation process as a Markov Decision Process of "diagnosis-localization-verification". Unlike standard approaches, our model is explicitly trained to identify a "key evidence slice" alongside the global diagnostic report, grounding its findings in verifiable visual evidence. Crucially, we introduce a novel cross-view consistency reward, which validates the semantic alignment between the golden-standard report and a local visual re-query of the selected key slice, providing additional rewards for correctly-localized reasoning. Experiments on large-scale 3D CT tumor datasets demonstrate that E-MRL significantly reduces hallucinations and improves diagnostic accuracy compared to SFT and RL baselines, offering a clinically interpretable solution for visually-grounded and tumor analysis.

Following the paradigm shift initiated by OpenAI o3, interleaved reasoning with code to enhance multimodal large language models (MLLMs) has become a pivotal research frontier. The existing literature focuses primarily on tool-use within vision-perception tasks. However, such approaches typically rely on predefined heuristics for visual manipulation and are inherently incapable of addressing numerical computation problems due to their exclusive focus on visual operations. This paper empowers MLLMs with adaptive interleaved reasoning capabilities through extended reinforcement learning training on code-augmented complex numerical computation tasks. To this end, we propose a comprehensive three-component solution consisting of: a two-stage cold-start data construction pipeline, data filtering strategies for RL dataset curation, and an adaptive tool-invocation strategy leveraging a group-constrained reward function for interleaved reasoning trajectories. Extensive experiments demonstrate that after Reinforcement Learning training with the group-constrained reward function, performance improves by an average of 6.1 percentage points (pp) on evaluation benchmarks. Specifically, the accuracy for interleaved reasoning samples increases by 9.9 pp, and the overall success rate of tool-use exceeds 95%. Our data and code are available at: https://github.com/CongHan0808/AIR.git.

Vision-Language-Action (VLA) models have established a powerful paradigm for generalist robotic manipulation by grounding control into the semantic reasoning of VLMs. Prevailing architectures typically model actions continuously via diffusion or flow processes, or discretely through either autoregressive generation or parallel decoding. Recently, Discrete Diffusion VLAs (dVLAs) have emerged as a distinct alternative, unifying vision, language, and action into a single discrete token space via masked generative modeling. While combining iterative refinement with unified representations, its training has thus far been restricted to Supervised Fine-Tuning (SFT), leaving the potential of Reinforcement Learning (RL) for further policy refinement largely unexplored. A fundamental challenge in RL for dVLAs is that the marginal probability of the final action generated by dVLAs remains intractable. To solve this problem, we propose \textbf{dVLA-RL}, shifting the learning objective from the marginal action probability to the joint probability of the sampled generation path. Specifically, by modeling the denoising process as a Markov Decision Process (MDP), we mathematically formulate this path probability as a product of step-wise transitions. This trajectory-level objective provides a unified formulation that natively accommodates variable denoising steps. Leveraging this intrinsic fexibility, we introduce a unified step scheduling approach for complex multi-task learning, tailoring denoising steps to specific task complexities to maximize both success rates and computational effciency. Extensive evaluations demonstrate that our approach achieves a success rate of \textbf{99.7%} on LIBERO. Furthermore, it establishes strong VLA-based results on RoboTwin 2.0 by delivering a \textbf{30.6%} improvement over the SFT baseline, remaining competitive with strong World-Action Model baselines.

Multi-view 3D Visual Question Answering (MV3D-VQA) requires integrating partial observations into a coherent 3D scene representation and selecting informative viewpoints for multi-step spatial reasoning. However, current multimodal LLMs are typically trained with sparse, answer-level supervision, which often yields inconsistent cross-view reasoning and brittle view selection. We present DR-MV3D (Dense Reward for MV3D-VQA), a map-grounded learning framework that provides dense, verifiable rewards to supervise the reasoning process. Our approach decomposes MV3D-VQA into (i) allocentric global map construction, (ii) question-conditioned view-trajectory planning, and (iii) egocentric grounding for answer prediction. To make intermediate steps learnable without manual annotations, we introduce two rewards: a global consistency reward that aligns the predicted map with geometry-consistent pseudo targets from frozen 3D vision foundation models (e.g., VGGT + SAM3), and a local trajectory reward that supervises ordered viewpoint selection. We optimize the full pipeline with trajectory-level policy optimization (GRPO). Experiments on MindCube, VSI-Bench, and BLINK (MV) show that DR-MV3D consistently improves over strong multi-image baselines, supporting the effectiveness of process-level dense supervision for multi-view 3D reasoning.

Scaling reinforcement learning for visual mathematical reasoning requires more than generating harder questions: as data volume grows, the reward labels themselves must remain reliable. Yet existing data pipelines scale supervision while trusting the labeller, and policy-side methods assume the underlying answers are already correct. We instead treat scaling as a verifiable data-construction problem and decouple two axes before any policy update: prompt difficulty, expanded by route-specific evolution operators, and answer reliability, enforced by offline hypothesis-test falsification. We instantiate this as VeriEvol, an iterative framework with two extensible components: a type-aware evolution module that rewrites low-difficulty image-question seeds into harder, image-grounded prompts; and HTV-Agent, a verifier that accepts an answer only after multi-source counter-evidence has failed to refute it. The resulting verified data scales in volume, extends by adding evolution routes or verifier channels, and plugs directly into existing GRPO-style RL recipes. On a five-benchmark visual-math suite, scaling evolved SFT data from 10K to 250K samples raises the mean accuracy from 35.42 to 54.73; then, with backbone, SFT initialization, and GRPO recipe held fixed, VeriEvol adds a cumulative +3.88 over an un-evolved RL baseline, of which +1.82 comes from evolved prompts and +2.06 from the HTV-Agent verifier. We release the prompts, data, models, code, and the full verifier trace of every sample, so that downstream work can scale and audit the pipeline rather than only inspect its outputs.

Large Vision-Language Models (LVLMs) have demonstrated remarkable capabilities in multimodal reasoning. However, prevailing reinforcement learning (RL) paradigms lack explicit counterfactual enhancement and causal learning mechanisms. This fundamental deficiency results in severe grounding failures, manifesting as a tendency to ignore visual evidence in favor of language priors or exhibiting hallucination drift during long chain-of-thought reasoning. To address this root cause, we propose CounterFactual Policy Optimization (CFPO), a novel framework that enforces causal consistency between visual perception and textual reasoning. CFPO introduces a cross-modal counterfactual enhancement mechanism, which regularizes the policy by maximizing the discrepancy between the model's predictions and those from a counterfactual state where critical visual cues are suppressed. This approach seamlessly integrates with standard algorithms like GRPO and DAPO without requiring external reward models or additional supervision. Extensive experiments demonstrate that CFPO significantly improves reasoning fidelity, achieving consistent gains of 3.17%-6.25% over standard RL baselines and 1.32%-2.13% over the state-of-the-art perception-aware method (PAPO). Code is available at https://github.com/Raven-July/CFPO.

Reinforcement learning with verifiable rewards (RLVR) is increasingly applied to large vision-language models (LVLMs), yet outcome-only optimization can drive a model to stop attending to the video and instead exploit linguistic priors -- a failure we call a visual shortcut. While the existence of such perception bypass is by now documented, how it forms, whether it can be undone, and when intervention still helps remain open. We treat the strength of a grounding penalty, lambda, as a control knob and characterize the formation-reversal dynamics of visual shortcuts along the training time axis. On a held-out, out-of-distribution diagnostic set, we find: (i) a sharp onset -- shortcut reliance emerges abruptly over a narrow window of optimization steps and is robust across random seeds; (ii) a monotone dose-response -- increasing lambda progressively suppresses the shortcut, and at an intermediate dose the trajectory first forms and then reverses the shortcut, exposing a hysteresis-like asymmetry between acquiring and removing it; and (iii) a critical intervention window -- applying the penalty before onset arrests shortcut formation, whereas the same penalty applied after consolidation is markedly less effective. Together these results recast visual-shortcut collapse not as a binary defect but as a controllable, time-dependent, and asymmetric process, with direct implications for when and how strongly to regularize multimodal RLVR.

In multimodal video reasoning, reinforcement learning-based methods typically rely on simplistic and inflexible reasoning-length control strategies that fail to adapt to the model's evolving competence. This mismatch may suppress necessary exploration at early stages, while encouraging redundant reasoning and inefficient decoding once the model becomes more competent. In this paper, we propose CARE, a competence-aware reward shaping framework for adaptive reasoning length optimization in multimodal reasoning. Specifically, CARE maintains a smoothed competence estimate via an exponential moving average of pass rates, and uses it to route training into progressive stages that shift the reward preference from exploration-oriented long-form reasoning to efficiency-oriented concise reasoning. To avoid conflating verbosity with intrinsic task complexity, CARE further normalizes reasoning effort with batch-level statistics, and introduces a posterior amplifier to strengthen reward signals for unexpectedly strong performance on historically difficult samples. The proposed mechanism is seamlessly integrated into the GRPO training pipeline and incurs no additional inference-time overhead. Extensive experiments on multiple video reasoning and general video understanding benchmarks demonstrate that CARE consistently improves reasoning accuracy, stabilizes reinforcement learning, and significantly enhances token efficiency. Moreover, CARE exhibits a characteristic inverted-U trajectory of reasoning length during training, and yields shorter yet more informative reasoning traces at convergence, indicating effective adaptive allocation of reasoning budget. We provide the source code for our proposed CARE framework and experiments at https://github.com/1Pansy/Video-CARE.

Passive models for long video understanding typically rely on a "watch-it-all" paradigm, processing frames uniformly regardless of query difficulty, causing computational cost to grow with video duration. Although interactive frameworks have emerged, they often rely on global pre-scanning, and their context cost still scales with video length. We propose OmniAgent, the first native omni-modal agent that formulates video understanding as a POMDP-based iterative Observation-Thought-Action cycle. OmniAgent executes on-demand actions to selectively distill audio-visual cues into a persistent textual memory, effectively decoupling reasoning complexity from raw video duration. To operationalize this, we introduce (1) Agentic Supervised Fine-Tuning to bootstrap native active perception via best-of-N trajectory synthesis with dual-stage quality control, and (2) Agentic Reinforcement Learning with TAURA (Turn-aware Adaptive Uncertainty Rescaled Advantage), which leverages turn-level entropy to steer credit assignment toward pivotal discovery turns. Crucially, OmniAgent exhibits positive test-time scaling, where performance improves as the number of reasoning turns increases, validating the efficacy of active perception. Empirical results across ten benchmarks (e.g., VideoMME, LVBench) demonstrate that OmniAgent achieves state-of-the-art performance among open-source models. Notably, on LVBench, our 7B agent outperforms the 10$\times$ larger Qwen2.5-VL-72B (50.5% vs. 47.3%).

Reinforcement learning has improved the reasoning ability of large language models, but applying outcome-only rewards to video multimodal large language models (Video-MLLMs) provides limited guidance on which visual evidence should support the answer. Inspired by multisensory integration, where consistent cues can enhance the salience and reliability of perceptual estimates, we introduce Consensus Frame GRPO (CF-GRPO), a temporal-annotation-free process-level reward framework for evidence-aware video reasoning. CF-GRPO constructs a consensus frame prior from intrinsic video cues, including temporal coverage, scene-transition cues, and query-conditioned visual relevance. It then computes a model-side frame-use score from visual and response representations and optimizes their agreement through the Consensus Frame Reward (CFR). With salience-aware sparse aggregation and distribution sharpening, CFR provides a high-contrast reward signal without requiring human temporal annotations. Experiments show that VideoCFR achieves competitive performance across complex video reasoning benchmarks and improves several metrics over representative Video-MLLM and RL baselines, while the consensus prior provides an interpretable view of the evidence frames emphasized during training. The implementation is available at https://github.com/1Pansy/VideoCFR.

Knowledge distillation transfers a teacher's competence to a small student but is brittle in the small-student regime: forcing the student to imitate logits from a much larger teacher hurts generalization on benchmark families beyond the training corpus. Reinforcement learning (RL) avoids logit imitation by training on the student's own rollouts. However, on questions where every rollout fails - yielding zero advantage and being silently discarded - injecting a stronger teacher's response into the policy gradient breaks the on-policy assumption and induces drift. We introduce Zone of Proximal Policy Optimization (ZPPO), inspired by Vygotsky's zone of proximal development. ZPPO keeps the teacher inside the prompt rather than the policy gradient. On hard questions, where the student's mean rollout accuracy is below half, ZPPO constructs two reformulated prompts. A Binary Candidate-included Question (BCQ) pairs one correct teacher response with one incorrect student response as anonymized candidates the student uses as references. A Negative Candidate-included Question (NCQ) aggregates the student's wrong rollouts into a single prompt to surface their shared failure modes. A prompt replay buffer recirculates each hard question until it either graduates - the student's mean rollout accuracy on it reaches half or more - or is FIFO-evicted under finite capacity, amplifying BCQ and NCQ inside the student's current zone of proximal development. We post-train Qwen3.5 students at four scales (0.8B-9B) as vision-language models with a 27B teacher and evaluate them on a 31-benchmark suite (16 VLM, 10 LLM, 5 Video); ZPPO outperforms off/on-policy distillation and GRPO, with the largest gains at the smallest scale.

Current Vision-Language-Action (VLA) models face a trade-off between efficient action generation and explicit deliberation. Directly decoding actions from vision-language backbone representations enables low-latency control, whereas textual reasoning, pixel-level subgoals, or world-model evaluation of decoded actions can improve planning but incur substantial latency and computational cost. We propose PearlVLA, a VLA framework that progressively refines a VLM-derived latent plan using feedback from the predicted consequence of each intermediate plan. PearlVLA uses a frozen latent world model (LaWM) pretrained on action-free video. At each refinement round, the current plan produces a continuous latent action code, and the LaWM predicts the corresponding latent visual subgoal. A future-guided plan refiner uses this subgoal to update the plan, so each revision reshapes the next LaWM query. After K rounds, the refined plan is passed once to the host policy's action interface to produce an action chunk. We further introduce Causal Refinement-Grouped Process-Reward RL to optimize latent refinement by comparing rewards from the longer-horizon imagined futures of plan edits made at the same refinement state. Experiments on the LIBERO and RoboCasa benchmarks show that PearlVLA performs competitively against strong existing methods.

Multimodal large language models (MLLMs) integrate strong text reasoning with visual inputs, yet their responses can be inconsistent with the underlying images, indicating ineffective utilization of visual evidence during inference. The prevailing training paradigm relies on large-scale caption-based pretraining for general alignment, followed by supervised fine-tuning and reinforcement learning to enable instruction following and complex reasoning. However, such pretraining provides only weak visual grounding: short, coarse captions bias models toward salient objects while neglecting fine-grained visual evidence. In this paper, we introduce Visual Evidence Pre-Alignment (VEPA), an intermediate stage between pretraining and post-training that explores a novel sufficiency-driven objective with Group Relative Policy Optimization (GRPO) to optimize question-conditioned visual evidence descriptions. Extensive experiments across diverse benchmarks show that our VEPA consistently enhances performance on visually demanding evaluations and complements standard supervised post-training. Further analyses show that the income stems from strengthened, transferable visual grounding, rather than from additional task-specific training.

Spatial VLMs have made substantial progress in geometric perception, yet complex spatial reasoning requiring multi-step inference over depth, distance, and scene relations remains challenging. Moreover, different spatial queries call for fundamentally different strategies: some are best addressed through purely linguistic, step-by-step deduction, while others require explicit 3D grounding before quantitative inference. We present Dual-Path Spatial Reasoning via Reinforcement Learning for Spatial VLMs (SR-REAL), a unified framework that equips a spatial VLM with two complementary reasoning paths: Language-Only Reasoning (LOR), which performs step-by-step linguistic deduction, and Detect-Then-Reason (DTR), which detects 3D geometric cues (e.g., centers or bounding boxes) via region tokens before explicit geometric inference. SR-REAL begins with a cold-start supervised fine-tuning stage that constructs LOR and DTR chain-of-thought supervision and exposes a region-to-3D interface, followed by RL that optimizes the policy model with accuracy and format rewards; for DTR, a discrete center-based detection reward further refines geometric alignment. Across diverse spatial benchmarks, SR-REAL significantly outperforms spatial VLM baselines: (i) a single RL-trained model supports both reasoning paths, with DTR excelling in region-aware tasks through precise 3D localization and LOR enhancing general spatial reasoning; (ii) jointly training both paths fosters mutual reinforcement; (iii) high-quality, blended cold-start data is crucial for stable RL optimization; and (iv) the model generalizes across datasets and domains without per-task tuning, demonstrating positive transfer between LOR and DTR.

Large language models (LLMs) often fail when answering requires identifying a small but decisive piece of evidence within a long or complex context, such as a single line in a tool trace or a subtle detail in an image. We propose ContextRL, a context-aware reinforcement learning (RL) method that improves long-horizon reasoning and multimodal performance through an \emph{indirect} auxiliary objective. Instead of supervising only the final answer, ContextRL presents the model with a query, an answer, and two highly similar contexts, and rewards it for selecting the context that supports the query--answer pair, thereby encouraging fine-grained grounding. We construct contrastive context data in two domains: for coding agents, trajectories serve as contexts, yielding 1k pairs built via condition filtering; for multimodal reasoning, images serve as contexts, yielding 7K pairs built via generative editing and similarity search. ContextRL achieves average gains of +2.2% over standard GRPO on 5 long-horizon benchmarks, and +1.8% across 12 diverse visual question answering benchmarks. To disentangle the effect of the proposed objective from that of additional data, we compare against data-augmentation baselines that repurpose the same contrastive contexts as standard query--context--answer examples. These baselines provide little to no improvement, showing that the gains arise from the proposed context-selection objective rather than from the contrastive data alone.

When pretrained VLA policies are fine-tuned through online RL, each rollout episode produces only a single binary outcome (success or failure), yet the actor update requires per-transition supervision. Existing approaches commonly reduce this sparse outcome to a single scalar reward or advantage signal, which conflates distinct forms of transition-level feedback and provides limited guidance once basic task success becomes achievable. First, a single scalar signal conflates the two objectives of viability and efficiency; once basic success is achieved, the binary label provides no gradient to distinguish efficient completions from slow ones. Second, real-world rollouts mix autonomous and intervention segments; naively assigning episode outcomes across these boundaries introduces incorrect credit assignment. To address these issues, we propose Hierarchical Advantage-Weighted Behavior Cloning (HABC), which trains separate critic heads for these two objectives on different data subsets and combines their outputs with a state-adaptive balance. A state-adaptive gate $g_t$ merges their one-step advantages, prioritizing viability when success is uncertain and shifting to efficiency only when viability is high, and converts the result into per-transition weights on the actor loss. Intervention-aware credit assignment further restricts outcome labels to segments executed by the current policy, preventing supervision from leaking across intervention boundaries. In real-robot experiments on three contact-rich bimanual tasks, HABC raises success from supervised fine-tuning (SFT) baselines of 36%, 44%, and 12% to 92%, 88%, and 38%.

Visual thinking should not only sound right; it should show its evidence. While recent vision-language models (VLMs) can produce natural-language reasoning traces, these traces often leave the supporting image regions implicit, making them hard to verify and difficult to supervise. We introduce visually grounded thinking, a reasoning process in which models interleave natural-language thoughts with explicit point or box groundings of the visual evidence used at each step. This lets the model express intermediate reasoning in language while grounding key objects in the image regions they refer to. To train this behavior, we construct a scalable synthesis pipeline that distills correct visual reasoning traces, extracts the visual objects required by the traces, grounds them with a SAM3-based agent, and derives aligned point and box supervision from the resulting masks. We further propose grounding-aware reinforcement learning, which combines answer correctness rewards with dense grounding rewards that score whether generated object references match the correct image evidence. Across two counting benchmarks and four spatial reasoning benchmarks, adding visually grounded thinking to Gemma3-4B-IT consistently improves performance over the original model and the non-grounded thinking baseline. On spatial reasoning, the visually grounded thinking 4B models match, and in some cases surpass, Gemma3-27B-IT from the same model family. Our analysis shows that point grounding is well suited to counting, while box grounding benefits most from explicit grounding rewards on spatial tasks. Overall, our results show that VLMs think better when their intermediate thoughts are tied to the image regions that make them true.

Reinforcement learning for multimodal large language models (MLLMs) is often hindered by severe reward sparsity in complex reasoning tasks. This challenge is particularly pronounced in human-centered scenarios involving states, emotions, intentions, and behaviors, where heterogeneous multimodal signals and subjective human factors make high-quality chain-of-thought (CoT) annotations expensive and difficult to obtain. Although many multimodal datasets provide expert-annotated ground-truth labels, directly using these labels for supervised fine-tuning may encourage shortcut learning in multimodal perception and provides limited transparency for safety-critical human--AI interaction. To address these limitations, we propose OmniOPSD, a Rationale-Privileged On-Policy Self-Distillation framework that uses frontier-generated rationales as teacher-side privileged evidence rather than student imitation targets. OmniOPSD uses frontier-generated evidence-aware rationales only as training-time privileged evidence context for a local teacher. The student samples its own rollout from the original multimodal input, while the rationale-privileged teacher scores the same tokens and provides dense token-level supervision. Thus, the student learns on its own trajectory distribution without directly imitating frontier-model completions, and inference requires no labels, rationales, CoT annotations, or closed-source model access. Experiments on MER-UniBench show that OmniOPSD achieves state-of-the-art performance with an average score of $84.19$, and ablations further support the value of rationale-privileged teacher guidance.

Vision-Language Models (VLMs) are AI systems that process both images and text, yet they often struggle with compositional visual reasoning questions that require chaining multiple steps together, such as identifying objects, counting them, and comparing the results. Existing approaches improve this reasoning by training models on human-written step-by-step explanations, but creating these annotations is expensive and difficult to scale. We propose a self-questioning framework that trains a VLM to break visual questions into smaller sub-questions and answer each one before producing a final response, using a reinforcement learning algorithm called Group Relative Policy Optimization (GRPO). The model is never shown examples of how to decompose questions, it discovers this behavior on its own, guided by a reward signal that scores whether the output contains sub-questions and whether the final answer is correct. We apply this framework to a 3-billion-parameter model, training on both synthetic scenes of geometric shapes (CLEVR) and real-world photographs (A-OKVQA). On A-OKVQA, both self-questioning and standard reinforcement learning substantially improve accuracy over the untrained model (52.2% and 51.6% vs. 46.8%). We introduce the first self-questioning VLM by rewarding not only the final answer like standard RL but additionally for generating intermediate sub-questions, enabling it to discover compositional decomposition strategies. These results suggest that teaching AI systems to ask themselves intermediate questions is a promising strategy for complex visual reasoning, particularly when the difficulty of a question warrants explicit step-by-step decomposition.

Hateful and propagandistic memes exploit the interplay between images and text to convey harmful intent that neither modality reveals alone. Although thinking-based multimodal large language models (MLLMs) have advanced vision-language understanding, their application to meme content moderation remains underexplored. We propose a reinforcement learning-based post-training method that improves classification performance and reference-based explanation quality in thinking-based MLLMs via task-specific rewards and Group Relative Policy Optimization (GRPO). Concretely, we (i) conduct a systematic empirical study of off-the-shelf MLLMs for hateful and propagandistic meme understanding across English and Arabic benchmarks, (ii) extend existing meme datasets with weakly supervised chain-of-thought (CoT) rationales via distillation and multi-LLM fine-grained propaganda annotations, (iii) introduce a GRPO-based objective with thinking-length regularization that jointly optimizes classification accuracy and explanation quality, and (iv) investigate self-supervised GRPO on unlabeled memes using consensus-based pseudo-labels. Experiments on the Hateful Memes and ArMeme benchmarks show that our approach improves over previously reported results on FHM accuracy (up to +2.1%, from 79.9% to 82.0%) and on ArMeme macro-F1 (up to +7.6 points, from 0.536 to 0.612 with explanations; +6.1 compared to the original ArMeme benchmark), while also generating natural-language explanations. On ArMeme, sequence-classification baselines remain stronger in terms of raw accuracy, whereas our approach provides more balanced per-class performance along with explanations. We publicly release our code, data extensions, and evaluation resources.

Reasoning capabilities of multimodal large language models (MLLMs) have improved considerably in recent years. Existing approaches typically rely on explicit chain-of-thought or continuous latent-space trajectories to enhance multi-step reasoning. However, these methods generally assume that an input admits a single latent interpretation and unfold reasoning along a fixed path or under a uniform computation budget. In real-world multimodal settings, visual observations are often subject to occlusion, blur, viewpoint variation, or semantic ambiguity, giving rise to multiple plausible interpretations. A uniform reasoning strategy not only limits the model's ability to explore multiple hypotheses but also incurs high memory usage and rollout cost. We present DLWM (Diverse Latent World Models), a multimodal reasoning framework that combines latent-space reasoning with reinforcement learning. First, we construct a set of diverse latent world hypotheses in continuous latent space, each capturing a different plausible interpretation of the visual input, and unfold latent reasoning independently on each hypothesis. An orthogonality-based diversity regularizer explicitly prevents hypothesis collapse. Second, we formulate the latent reasoning process as a resource-constrained sequential decision problem and introduce a resource-aware reinforcement learning policy that adaptively allocates computation across hypotheses, dynamically deciding whether to expand, terminate, or merge reasoning paths, thereby substantially reducing memory footprint and improving rollout efficiency. Experiments on multiple multimodal reasoning benchmarks demonstrate that DLWM outperforms existing methods by 2-5 points in accuracy while reducing memory usage by 24%.

Existing Vision-Language-Action (VLA) models predominantly rely on explicit Chain-of-Thought (CoT) reasoning to bridge perception and action. While effective, this paradigm suffers from high computational costs and error propagation in multi-step tasks. In this paper, we propose Adaptive Variable Alignment VLA (AVA-VLA), a novel Latent Reasoning VLA framework that models reasoning as a sequence of unobservable latent variables, bypassing the need for explicit text generation. However, latent trajectories are inherently susceptible to noise interference and misalignment with downstream objectives. To address this, we introduce a Reinforcement Learning-based Denoising mechanism that treats latent state generation as a sequential decision process, optimizing reasoning trajectories via task-level rewards. Furthermore, we incorporate an Early-Exit Strategy that adaptively terminates reasoning based on state confidence, enabling a dynamic trade-off between depth and efficiency. Extensive experiments on embodied decision benchmarks demonstrate that AVA-VLA achieves a 6x inference speedup over explicit CoT methods while attaining a 98.3% average success rate on LIBERO, improving both efficiency and long-horizon stability over full-reasoning baselines.

Reinforcement learning with verifiable rewards (RLVR) has successfully elicited the reasoning capabilities of large language models, motivating its extension to multimodal scenarios. Existing methods primarily focus on improving the visual coverage of reasoning traces and mitigating visual hallucinations, but underestimate the semantic inconsistency between the reasoning process and the final answer. In this paper, we delve into thinking-answer inconsistency in RLVR for large vision-language models (LVLMs), showing thorough analyses of rollouts collected throughout Group Relative Policy Optimization (GRPO) training process and post-RLVR evaluation outputs that this issue persists during training and remains present during inference. Motivated by the analysis, we propose Consistency-Oriented Reasoning Alignment (CORA), which introduces thinking-answer semantic consistency into RLVR through a lightweight plug-and-play consistency reward model, and further incorporates Hybrid Reward Advantage Splitting (HRAS) to stably coordinate task and consistency optimization. Extensive experiments across representative multimodal reasoning benchmarks and mainstream LVLMs show that CORA improves task performance while effectively mitigating thinking-answer inconsistency, leading to more faithful reasoning traces.

Letting a vision-language model (VLM) think longer at test time has driven much recent progress. A natural way to bring this to spatial grounding is visual self-correction: the model predicts a bounding box, sees it rendered on the image, and refines it over several steps. We build a faithful instance of this idea, Iterative Visual Thinking (IVT), with a two-phase recipe: a supervised warm-up in which the base model's own predictions serve as realistic errors that a teacher VLM turns into corrective reasoning traces (yielding training data without human annotation), followed by GRPO with a simple IoU reward. Measured the way such systems are usually reported, it works: the trained model surpasses the single-shot base by +2.4pp [email protected]. We show this gain is a measurement mirage. The reported number silently keeps, per sample, the trajectory step closest to the ground-truth box: an oracle that needs the very answer it predicts. Re-scored under deployable, label-free stopping rules the improvement vanishes, and the best policy is not to iterate at all: stopping at step 0 matches the base and beats every shippable rule. The cause is a verification failure, since the model can generate a better box somewhere in its trajectory but cannot identify it. Self-verification confidence correlates only weakly with correctness (r about 0.22), and a counterfactual overlay shows the loop reacts to the presence of a rendered box rather than its correctness. We distill the lesson into an honest-trajectory evaluation protocol: accuracy under fixed label-free policies plus an explicit oracle-shippable gap.

RL-based post-training has been widely adopted to enable interleaved visual and textual reasoning in unified multimodal models capable of both text and image generation. However, most existing approaches are built upon autoregressive (AR) unified models, which require full image regeneration during visual reasoning. In this work, we demonstrate that multimodal discrete diffusion models are effective alternatives to AR models for reinforcement learning in interleaved reasoning, owing to their ability to perform efficient visual rollouts via localized visual editing rather than full image-token regeneration. This reduces rollout computation during GRPO by 26.9% compared to AR baselines, with minimal performance drop. Despite the improved efficiency, we find that joint reward assignment, which employs a shared reward signal across modalities, introduces cross-modal interference between unrelated image and text token sequences during RL updates. To address this issue, we propose factorized reward assignment, a strategy that assigns rewards independently to text and vision segments. With factorized reward assignment, our RL approach achieves an 11.2% improvement over joint reward assignment and a 38.04% improvement over the base model.

Interleaved thinking, where a unified multimodal model alternates between textual reasoning and visual generation, has shown promise on spatial and physical tasks. However, in complex long-chain scenarios, we identify a fundamental failure mode: generated images diverge from the textual context while subsequent text ignores the visual evidence, causing the two modalities to alternate without genuinely informing each other. We term this Modal Isolation and attribute it to compounding information loss at modality boundaries. We decompose each reasoning cycle into atomic operations and define modality transition loss, quantifying cross-modal hallucination (text-to-image) and visual utilization deficit (image-to-text) at each boundary. We propose MoTiF (Modality Tiransition Fidelity), a two-stage training framework that directly optimizes these transitions: Reflective SFT trains the model to detect and recover from erroneous visual outputs; Flow-GRPO improves image generation fidelity via reinforcement learning. All training signals in MoTiF derive from transition-level fidelity rather than end-task accuracy. Across four visual puzzle benchmarks, this transition-level supervision substantially improves both cross-modal coherence and final task accuracy. The results demonstrate that effective interleaved reasoning requires explicit structural supervision at modality boundaries, not merely scaling or end-task optimization.

While recent vision-language models (VLMs) demonstrate strong multimodal understanding, they remain limited in spatial reasoning tasks that require active evidence acquisition and multi-step visual interaction. This limitation suggests that relying solely on implicit visual representations from vision encoders is insufficient for recovering fine-grained spatial evidence. We introduce PERception-Interaction-reason Agent (PERIA), a tool-augmented visual agent for spatial reasoning tasks across map reasoning, visual probing, and vision reconstruction. PERIA uses two lightweight tool families: vision perception tools for exposing textual, symbolic, and spatial evidence, and vision interaction tools for manipulating visual context, tracing paths, and verifying spatial relations. To train PERIA, we develop a unified recipe that combines supervised tool-use trajectory synthesis, composite rewards, and Observation-Relaxed Group-in-Group Policy Optimization (OR-GIGPO) for effective multi-tool behavior. Experiments on 13 benchmarks from 8 datasets show that PERIA-8B improves over the Qwen3-8B backbone by 10.0% on in-distribution benchmarks and 4.4% on out-of-distribution benchmarks, while outperforming previous state-of-the-art baselines of similar size by 7.0%-14.8%. It also achieves performance comparable to much larger models such as Qwen3-VL-235B-A22B-Thinking and GPT-5, demonstrating the effectiveness of PERIA in enhancing spatial reasoning capabilities.

Spatial reasoning remains a challenge for Multimodal Large Language Models (MLLMs), as it requires reliable multi-hop inference over both intermediate states and state transitions. Current studies often leave intermediate states unverified and treat state transitions as implicit processes, which limits reliability in multi-hop spatial reasoning. To address this, we propose State-aware Visualization-of-Thought (SVoT), a reinforcement learning framework that generates interleaved, verifiable intermediate states and visualizations. SVoT integrates transition reasoning chains into the generation processes, enabling the model to verify action preconditions and effects through interleaved textual and visual reasoning. We train SVoT via Group Relative Policy Optimization (GRPO), instantiating verification through reward design and evaluating the efficacy of different fine-grained rewards. As existing benchmarks reduce state transitions to single-variable updates, substantially simplifying the problems, we establish five domains by extending classical environments and introducing two novel domains, Pacman and Gather, that require multi-object interactions and numerical reasoning. These domains support systematic evaluation of multi-hop spatial reasoning with quantitative verification of generated intermediate states and transition reasoning. SVoT with transition-aware supervision achieves state-of-the-art performance across the introduced domains, yielding up to a 65% absolute accuracy gain on out-of-distribution test sets.