Generative RL 进展

Reinforcement learning enables large language model (LLM) agents to learn multi-step behaviors through interaction with their environments. However, rewards in many interactive tasks reflect only the final outcome, providing limited guidance on which intermediate decisions advance the task. Successful training trajectories contain intermediate states that can provide supervision for subsequent interactions. We introduce Trajectory-to-Step Policy Optimization (T2SPO), a method that uses past interaction trajectories to provide step-level feedback for policy learning. T2SPO derives remaining-distance targets from successful trajectories and pairs them with representations of the states visited along the way. Conditioned on these examples, a pretrained TabPFN regressor estimates the remaining distance to success at each state of a new rollout. Changes in this distance estimate across consecutive states yield auxiliary credit for agent steps alongside task-level supervision. As training proceeds, newly completed trajectories refresh the estimator's context, incorporating new experience without updating its parameters. Experiments with 1.5B and 7B language models on ALFWorld and WebShop show that T2SPO consistently improves overall task success over GRPO.

Agentic reinforcement learning (RL) trains a large language model (LLM) to act over long, multi-step interactions. However, a single localized error can cause task failure, while trajectory-level rewards provide limited guidance for assigning credit to individual decisions. To address this limitation, we introduce Segment-level Hindsight Advantage Reweighting for Policy Optimization (SHARPO), a credit-assignment mechanism that refines Group Relative Policy Optimization (GRPO) at the level of environment-facing segments. Inspired by the existing on-policy self-distillation (OPSD) method, SHARPO computes teacher-student log-probability gaps within each segment and uses the resulting signal to compute a bounded multiplier on the GRPO advantage. This multiplier is shared by all tokens within the segment, allowing credit to vary across different segments. With Qwen2.5-7B-Instruct, SHARPO outperforms existing baselines on the ALFWorld and WebShop benchmarks, including GRPO, SDAR, RLSD, and StepOPSD.

Preference-based alignment methods typically optimize against a single preference model, and can therefore be brittle when pairwise preferences are uncertain: noisy, heterogeneous, or shift after deployment. To address these issues, we propose Robust Nash Alignment, a game-theoretic framework for alignment to uncertain pairwise preferences. Our formulation has a major learner seeking a policy with a large worst-case win rate against both an adversarial competitor and any preference kernel lying in an ambiguity set around a nominal preference. When the ambiguity set captures the uncertainty in preferences, the resulting robust objective of the game directly yields a certified lower bound on worst-case performance. However, we note this problem is computationally challenging to optimize, and to address this, we introduce a four-player primal-dual proxy game involving the leader policy, follower policy, adversarial kernel, and dual variable, and develop a single-loop optimistic mirror descent-ascent algorithm for it. We show that the proxy always lower-bounds the truncated hard-constrained objective, quantify the proxy-to-hard gap, and characterize an exactness condition under which the proxy recovers the robust objective. We then prove an \(\mathcal{O}(1/\sqrt{T})\) average-iteration convergence for the proxy-game duality gap, which implies a near-optimal robust policy for the original robust objective. Experiments on controlled tabular games and LLM alignment with uncertain preference further validate the convergence theory and show improved performance over nominal baselines.

Autonomous red team agents increasingly stress-test AI-enabled cyber defenses by planning strategy and executing multistage attacks. Reinforcement learning (RL) and large language models (LLMs) offer complementary mechanisms for the planning and execution such agents require, and prior work has combined them in hybrid hierarchies. Yet a given architecture is typically developed and evaluated within a single environment, leaving open whether an observed advantage reflects a generally stronger decision mechanism or merely alignment with a particular setting. We address this gap with a controlled cross-environment comparison of two homogeneous hierarchical red team architectures: an RL planner with an RL executor (RL+RL) and an LLM planner with an LLM executor (LLM+LLM). We evaluate both against expert autonomous defenders in CybORG CAGE-4 and in Cyberwheel at two network scales, across 18 configurations under one unified disruption metric. We find a pronounced environment-dependent inversion. RL+RL wins the compact, densely rewarded CAGE-4 (78.5% disruption success versus 18.0% for the strongest LLM configuration) and the 100-host Cyberwheel network (81.0% versus 50.5%), while a pretrained cybersecurity LLM agent wins the larger, escalation-gated 1010-host Cyberwheel network (55.0% versus 0.0% for RL). A kill-chain analysis explains the inversion through architecture-specific bottlenecks that aggregate success rates conceal.In the 1010-host Cyberwheel network, RL discovers and compromises hosts but stalls at privilege escalation, whereas in CAGE-4, LLM agents obtain privileged access but rarely convert it into operational impact. These results indicate that conclusions drawn in a single environment may not generalize, and that hybrid planner-executor designs should be motivated by specific failure modes rather than the assumption that one architecture is universally preferable.

Reinforcement learning with verifiable rewards (RLVR) has improved the reasoning capabilities of large language models (LLMs), yet their predictions remain sensitive to task-irrelevant prompt features. We investigate this sensitivity through semifactual prompt interventions that preserve the underlying problem and its answer. Our analysis reveals substantial variation in token-level sensitivity and shows that suppressing high-drift token candidates during decoding improves reasoning accuracy without updating model weights. These findings highlight a limitation of Group Relative Policy Optimization (GRPO), which assigns the same outcome-derived advantage to every response token and may reinforce potential spurious dependence alongside useful reasoning. Motivated by this observation, we introduce Semifactual Credit-Augmented Policy Optimization (SCAPO), a causally inspired variant of GRPO that incorporates semifactual stability into token-level credit assignment. SCAPO measures token probability drift for fixed responses under semifactual interventions and uses normalized stability scores to reduce advantages for relatively unstable tokens during early training, while granting no additional credit for stability alone. On Qwen3-4B-Base and Qwen3-1.7B-Base, SCAPO improves AIME 2024-2026 accuracy over GRPO by 5.63 and 4.17 percentage points, respectively. At both model scales, SCAPO achieves the best results on most evaluated mathematics benchmarks and all evaluated out-of-distribution benchmarks among the compared methods. These results suggest that semifactual stability provides an effective training signal for improving reasoning and generalization through finer-grained credit assignment in RLVR. The code is available at https://github.com/DtYXs/SCAPO.

Inverse Reinforcement Learning (IRL) aims to recover a reward function that explains expert demonstrations. Existing IRL methods typically rely on a bi-level optimization procedure that alternates between reward learning and policy optimization, leading to substantial computational burden and training instability. In this work, we introduce a different route that eliminates policy optimization entirely by leveraging diffusion policies. Our key insight is that a diffusion policy encodes the action-gradient structure of the optimal soft Q function, enabling reward learning to be cast as a sequence of value recovery problems, thereby allowing us to bypass reward-policy loops inherent in prior IRL methods. Specifically, our method proceeds in three stages: (I) recovering the optimal soft Q function via action-gradient matching and estimating the corresponding soft value function (LogSumExp of Q values) in a way inspired by Gumbel regression; (II) calibrating these soft values by inferring a state-dependent offset; (III) extracting the reward by enforcing Bellman consistency. This leads to Loop-Free Inverse Reinforcement Learning (LFIRL), a fully offline algorithm that operates in a simple, loop-free, and sequential manner. LFIRL is simple to implement and significantly improves training efficiency while maintaining strong reward recovery performance. Empirically, across Maze, Franka Kitchen, Adroit Hand Pen, and Push-T benchmarks, LFIRL achieves 2-3x speedup over the fastest baselines, while matching or surpassing state-of-the-art methods in reward recovery quality.

Standard language model RL algorithms credit every token of a long rollout with the same advantage determined by the terminal reward. Actor-critic methods can provide finer-grained credit assignment, but learned critics are generally considered too inaccurate to trust when training LLMs with RL. In recent works, even when a critic is present, it is used only for baseline estimation, so every trajectory must be rolled out to its terminal reward. We introduce Actor-Critic with Action Chunking (AC2) that removes the need to roll every trajectory to completion. AC2 instead assigns credit to action chunks: short continuations of prefixes of past trajectories. A learned critic scores the state reached at the end of each action chunk, allowing the policy to update without observing a terminal reward. We make critic-based credit assignment reliable through three design choices. First, we introduce local readiness which uses critic-based updates on a problem only when the critic is sufficiently accurate on that particular problem. Second, when available, we provide the critic with a reference solution from a previous successful rollout. Third, we assign credit over action chunks of 10k tokens rather than individual tokens, giving the critic a more meaningful portion of the trajectory to evaluate. We train Qwen3-4B on FineProofs-RL using AC2 and evaluate on IMO-ProofBench. AC2 exceeds GRPO's peak validation score of 18.5% using 2.5x fewer decoding FLOPs. This gain comes from two sources, (1) AC2 requires 25% fewer training steps to reach this score, and (2) each step generates fewer tokens because the policy does not need to continue every trajectory to completion. Conceptually, we demonstrate that we can remove the need to roll out every trajectory to completion, opening up a large previously unexplored design space for LLM RL algorithms.

Group Relative Policy Optimization (GRPO) has emerged as a memory-efficient reinforcement fine-tuning (RFT) technique for reasoning-intensive tasks. How- ever, GRPO training dynamics on small language models (SLMs) remain poorly understood, limiting its reliable adoption and reproducibility in open and resource- constrained environments. In this work, we present a systematic study of GRPO fine-tuning for SLMs ranging from 1.5B to 7B parameters under a practical single- node 8xA100 compute budget. Our study spans multiple model families and reasoning domains, including mathematics, coding, and multiple-choice question answering (MCQ) in science. Across these settings, we analyze how group size affects policy convergence, training stability, and downstream benchmark per- formance. We further characterize tensor-level update dynamics during GRPO training and investigate whether the choice of LoRA target modules and layers can improve the performance of GRPO-tuned models. While our initial GRPO-tuned models outperform their base counterparts on approximately 80% of mathematical benchmark evaluations, they demonstrate limited capability on MCQ and code reasoning tasks. Guided by our mechanistic evaluations, we refined our LoRA and reward-shaping configurations to improve performance in latter domains. These findings provide practical guidance for GRPO training for SLMs.

Legal language models require reward signals that capture not only answer correctness but also the multidimensional quality of legal responses. Existing reward methods, however, often rely on coarse-grained holistic judgments, providing limited domain specificity and interpretability. We introduce LexReward, a taxonomy-driven framework for legal reward modeling. LexReward characterizes legal response quality along three complementary dimensions: Style, covering lexical and syntactic quality; Element, assessing legal subjects, facts, statutes, and decisions; and Chain, evaluating the order, completeness, correctness, and non-redundancy of legal reasoning. For each dimension, we develop rubrics that specify evaluation criteria and quality levels. The resulting rewards are used to construct pairwise preference data for Direct Preference Optimization (DPO) and reward-model training. Experiments show that the rubric-based rewards reliably distinguish legal responses of different quality and that DPO training on the preference data improves performance across all three dimensions. The learned reward models, LexRM, also support effective downstream optimization: each dimension-specific reward model improves policy performance in its corresponding dimension through reinforcement learning, without requiring reference answers at reward time. Dimension-wise analyses further support the effectiveness of the proposed taxonomy and reward construction.

Reinforcement learning with a verifiable reward (RLVR) offers a scalable approach to training language-model agents, yet sparse outcome rewards can leave early training with little signal for policy improvement. We identify an On-Policy Acceleration Phenomenon: in our main comparisons, RLVR initialized with on-policy distillation reaches high performance earlier in training and achieves both higher average performance during subsequent RLVR and higher final performance than the alternative baselines. Motivated by this observation, we study On-Policy Warmup (OPW), a teacher-guided stage in which the student trains with teacher supervision on its own interaction trajectories before transitioning to RLVR. Unlike imitation on fixed teacher-generated trajectories, OPW targets states induced by the student's own decisions, including imperfect actions and recovery situations. We provide a theoretical explanation by connecting on-policy reverse-KL distillation to trajectory-level distribution matching. Under a competent teacher and sufficiently small population distillation loss, this connection yields a lower bound on initial verifier success and a corresponding bound on reward-discovery complexity. For group-relative RLVR, we further characterize when increased success probability produces more reward-informative groups. Together, our findings support on-policy distillation as an effective warmup for agentic RLVR and identify initial reward discovery as a mechanism that can contribute to the observed acceleration.

Black-box distillation is a practical route for transferring capabilities from API-accessible large language models that expose only text outputs into smaller student models. Recent on-policy adversarial methods such as GAD improve over SeqKD by forming an adversarial loop between a critic and a student, where the critic provides rewards for GRPO-based student policy optimization over the student's sampled responses. However, GRPO computes advantages from the within-group relative rewards of student samples for the same prompt, whereas the critic is trained primarily to distinguish teacher responses from student responses. This objective mismatch can produce reward groups with collapsed scale or fragile margins, leading to brittle grouped optimization signals. We propose Groupwise Reward Geometry Conditioning (GRGC), a two-stage framework that improves advantage construction by shaping student-side reward groups during both critic training and policy optimization. To improve critic-side conditioning, Gaussian groupwise Optimal Transport calibration regularizes the critic during training to produce reward groups with non-collapsed spread and smooth rank-wise gaps by matching sorted prompt-wise rewards to group-centered Gaussian quantiles. Building on this conditioned reward geometry, policy-side group power modulation reshapes the prompt-wise reward groups before they are converted into advantages, preserving the critic-induced ordering while increasing optimization-relevant margin separability. Extensive experiments across diverse teachers, student model families and scales, and training datasets demonstrate the effectiveness of GRGC on both in-distribution and out-of-distribution evaluations, while introducing negligible overhead over GAD. The code is available at https://github.com/2018cx/GRGC.

Preference distillation typically treats a teacher response as preferred and the student's own response as rejected. This assumes that self-generated failures are the most informative negatives and that rejects must come from a model at least as large as the student, making generation costly at scale. We find neither assumption holds: across students from 7B to 72B, smaller frozen models generate rejects with less inference compute yet train stronger students than self-generated rejects, before and after sequence-level knowledge distillation, on code generation and mathematical reasoning. To explain this result, we derive a finite-horizon utility bound for Direct Preference Optimization in a linearized feature model. The bound characterizes favorable reject distributions and motivates three interventions. First, mixing rejects from smaller and student-scale models improves performance as the smaller model's share increases. Second, reassigning rejects to other prompts and shuffling their code tokens still outperform length-matched gibberish, showing that task structure contributes to reject utility. Third, selecting candidates with lower likelihood under the reference policy improves net transfer when higher-likelihood candidates provide less useful contrast. Lower-likelihood selections outperform higher-likelihood ones for every source. These results suggest that effective rejects preserve task structure while limiting coupling to the reference policy, and that smaller frozen models can provide them at low cost.

Online reinforcement learning has been extended to flow matching for diffusion model (DM) image generation. However, this paradigm faces three limitations: (1) Window selection. Existing methods manually set the stochastic differential equation (SDE) sampling window, i.e., the denoising steps where exploration noise is injected. We instead determine it from each model's denoising trajectory. (2) Reward saturation. Current methods rely on scoring models trained on human annotations; we find that such scores are extremely high and nearly indistinguishable on the latest SOTA open-source DMs, making advantage estimation largely ineffective. (3) Sample inefficiency. A single scalar reward collapses different failure modes into almost identical scores, leaving minimal gradient guidance for targeted improvement. To address these issues, we propose CAST (Causal Advantage-Structured Training), an RL fine-tuning method for pretrained DMs, which (1) identifies the denoising step at which each model fixes the objects and their spatial arrangement in the image and uses that timing to set the SDE window, (2) decomposes each prompt via Causal Scene Graphs (CSG) into verifiable-atoms, i.e., minimal semantic units such as an object, count, attribute, or spatial relation that can each be checked independently, and rewards each atom separately, and (3) projects the signed atom-level advantages into pixel space through teacher-forced attention and uses them to spatially weight the SDE policy objective. We fine-tune two of the strongest open-source DMs, FLUX.2-dev and Qwen-Image-2512, with CAST, and evaluate them on GenEval 2, a compositional benchmark, and on Qwen-Image-Bench for overall quality. Within almost the same training budget, CAST's improvement over the base model on the most challenging GenEval 2 prompts is up to 3.07x that of Flow-GRPO, while overall generation quality also improves.

Immersive displays can enable rich and diverse virtual experiences. Manually authoring every possible experience to realize this potential, however, is prohibitively expensive, difficult to scale, and impractical. Generative AI models could remove this bottleneck, but today's models are built for conventional displays and cannot generate the high-resolution, stereoscopic $360^\circ$ content required for immersive viewing. Further, temporal and stereo inconsistencies that may be tolerable on conventional displays can become highly disruptive when viewed through an immersive headset. Here, we address this gap with a zero-shot generative pipeline that extends existing video diffusion models into 4K stereoscopic $360^\circ$ videos. Inspired from binocular vision and depth perception, we develop an epipolar-aware $360^\circ$ image matching metric that captures the temporal and stereo geometric inconsistencies across views. We then use this metric as a preference signal for direct preference optimization with limited training data. Our work enables $360^\circ$ stereo video generation and provides a scalable path for bringing generative content to immersive displays, allowing diverse mixed reality experiences on demand.

Adapting a pretrained generative model to an arbitrary preference expressed as a utility function underlies reward alignment, guided design, and constraint satisfaction, enabling diverse applications. Existing fine-tuning methods trade off generality against computational cost: they either restrict the family class of supported preferences to keep optimization simple or preserve generality at the expense of efficiency. We introduce Fenchel Tilt Flow Control (FTFC), which decouples utility optimization from generative-model fitting. FTFC first optimizes for a target distribution by jointly fitting an effective reward and density-ratio weights on pretrained samples. Method combines the utility's variational structure with Fenchel duality, supporting general $f$-divergence penalties that determine how rewards are transformed into an distribution-correction weights. These weights are then frozen and used to modify a diffusion or flow model in a single stage of importance-weighted denoising or flow matching, without differentiating through sampling trajectories. We establish exact duality for concave utilities under suitable conditions and show that weighted fitting reproduces the optimal target distribution for a given utility. Across image and molecule generation benchmarks, FTFC improves over baselines on diverse preference functions, while also being up to $20\times$ more efficient. roposed method enables adaptation beyond expected-reward maximization without complex optimization, while preserving robustness for more general class of the utility functions compared to baselines.

Masked diffusion language models (dLLMs) have shown strong potential for faster inference through parallel token generation when combined with confidence-based samplers. However, recent work has shown that such methods can defer unmasking high-entropy fork positions at which multiple plausible continuations exist. This results in reduced generation diversity, as shown by worse pass@k scaling, and limits gains obtainable from RL post-training. To avoid this flexibility trap, prior work advocated for autoregressive (AR) sampling. Here, we show that discarding confidence-based sampling is unnecessary and, once inference cost is taken into account, wasteful. We first propose Fork-dLLM, a simple hybrid sampler that uses AR-style ordering only at uncertain fallback steps while retaining parallel generation otherwise. We then extend the same principle to post-training with ForkGRPO, which uses Fork-dLLM rollouts and applies the GRPO objective only at fallback steps, preserving exact policy-likelihood ratios while substantially reducing rollout and optimization cost. In our experiments, Fork-dLLM matches the strong pass@k scaling of AR sampling while being 2-3x more efficient, and ForkGRPO achieves downstream performance comparable to or better than AR-based GRPO baselines at a substantially lower training cost.

Training LLM agents with reinforcement learning (RL) is bottlenecked by environments, which must provide verifiable rewards, support long-horizon interaction, and scale cheaply. Existing approaches rely on costly human-curated data or on LLM-generated environments that risk hallucinations and benchmark contamination. We show that LLMs can instead be trained into capable search agents using synthetic environments generated entirely by rules, whose generation requires no LLM and has zero marginal cost. We build PhantomEnvironments, multi-turn RL environments from fictional worlds, where agents must search a corpus of templated articles to answer multi-hop questions. Despite sharing no facts with the real world, these strikingly simple environments yield agents that transfer to real-world multi-hop search benchmarks, often outperforming real-world training data on newer benchmarks. Trained agents generalize to unseen fictional universes, and Qwen models learn to scale their search budget roughly linearly with question difficulty, suggesting emergent search scaling from environment interaction alone. Ablating environment complexity reveals that hop count drives transfer more than constraints or comparisons: even the simplest rule-generated environments are a surprisingly effective, free resource for training generalizable LLM agents.

LLM agents rely on retrieval tools to access external knowledge, yet visual agentic search remains severely bottlenecked by standard single-step retrievers. In current pipelines, the agent must issue text queries for every intermediate step, struggling when visual clues are difficult to describe or when the retriever fails to surface necessary intermediate evidence within its top results. We hypothesize that offloading multi-step navigation across the entire embedding space directly to the retrieval tool resolves this performance bottleneck. To study this systematically, we introduce VHOP, a flexible data generation framework and benchmark with five core difficulty levels testing both visual matching and search planning. Using this framework, we develop VHOP-Router, an end-to-end training pipeline---combining supervised fine-tuning, online imitation learning, and reinforcement learning---that transforms a standard embedding model into an autoregressive multi-step retriever. Operating directly in the visual latent space, VHOP-Router retrieves linked image chains in a single tool call without requiring the agent to formulate intermediate text queries. Experiments show VHOP-Router boosts retrieval performance from under 5% to 76.3%. In agentic search, it improves task success rates by 52.7% and reduces the average token length by 61% from 1886 to 728, whereas upgrading the agent yields only a 3.7% gain. Compared to a strong baseline where the agent retrieves the top 50 results per step, VHOP-Router maintains superior performance while reducing in-context images by $23\times$ and cutting the cumulative API payload by $35\times$. The models also generalize robustly to unseen difficulty levels and realistic test sets. Ultimately, VHOP and VHOP-Router provide an efficient and effective solution for visual agentic search that leaves native LLM capabilities entirely intact.

LLM agents often admit multiple high-quality solutions to the same task, differing in reasoning structure, tool-use pattern, or interaction trajectory. Yet existing notions of diversity in LLM post-training are mostly implicit, arising from general stochasticity and regularization mechanisms rather than explicitly targeting task-relevant behavioral variation. While such implicit diversity can be useful, it does not directly specify which forms of behavioral variation should be encouraged for a given task. In this work, we study explicit trajectory diversity in RL-based post-training for LLMs. Our key idea is to define diversity through user-specified, task-specific trajectory descriptors, which map each sampled trajectory to an interpretable behavioral representation, and then measure diversity as a set-level functional over the resulting descriptor matrix. Building on this formulation, we introduce Trajectory-guided Joint Policy Optimization(TJPO), a single-policy framework that optimizes explicit diversity over sampled trajectory groups, avoiding the need for population-based policy training, and instantiate it within group-based policy optimization through trajectory-level learning signals. This design makes the diversity objective both interpretable and controllable. Experiments on Sokoban and ALFWorld show that TJPO improves task-specific trajectory diversity while maintaining competitive task performance. Descriptor and trajectory analyses show that the learned variation follows the specified behavioral dimensions and includes distinct successful strategies. Extra experiment results suggest that explicitly shaping trajectory diversity can help LLM agents satisfy user requirements and remain effective when task conditions change.

Graph agents extend large language models (LLMs) with the ability to actively explore and reason over knowledge graphs through multi-step interactions with graph tools. However, training capable graph agents typically requires large collections of question-answer pairs and reasoning trajectories, whose manual construction is costly and difficult to scale. Moreover, employing proprietary LLMs to generate such supervision further risks exposing sensitive graph data to external services. Therefore, we propose GraphCert to bootstrap agentic graph reasoning with certified evidence rubrics during post-training. Specifically, the Bootstrapped Graph Quizzer guided by generation controls produces graph-grounded QA pairs and marks supporting evidence, which undergo execution certification and semantic curation. The accepted evidence is then canonicalized into certified evidence rubrics that later reward Graph Solver evidence alignment alongside answer correctness during GRPO training. Experiments on five graph reasoning domains in GRBENCH demonstrate that GraphCert consistently outperforms substantially larger LLM agents and post-training method. Furthermore, our analysis demonstrates that the learned policy transfers robustly across heterogeneous graph domains, suggesting that GraphCert acquires reusable graph-reasoning capabilities rather than domain-specific patterns. These results establish executable self-certification as an effective approach to self-training compact graph reasoning agents. Our code will be made publicly available.

Chain-of-thought (CoT) traces often serve as a proxy for how Large Language Models (LLMs) arrive at their answers. However, growing evidence shows that models' CoT often fails to reflect their internal computations and can be changed without affecting their final answers. In this work, we measure and improve the alignment between the reasoning described in an LLM's CoT and what it computes internally. We propose CoT-Interpretability Alignment (CIA), a metric that measures the agreement between a model's CoT traces and its internal reasoning strategies as detected by interpretability tools. We evaluate CIA on three tasks (two-hop question answering, hint intervention, and integer multiplication) across three LLMs, finding that LLMs exhibit limited alignment across all tasks (44.8-75.9%). We then experiment with improving CIA via post-training, setting both the task accuracy and parametric faithfulness signals as a reward. Experiments show that we can substantially improve CoT parametric faithfulness while maintaining or improving the task accuracy. We provide rich analysis, such as their generalization patterns. Our work provides both a framework for auditing CoT parametric faithfulness and a pathway toward making models' explicit reasoning more trustworthy. Code and data are available at https://github.com/yihuaihong/CIA-minimal-repro.

Reinforcement Learning with Verifiable Rewards (RLVR) has been extended to Large Vision-Language Models (LVLMs), and perception-aware methods further encourage policies to rely on visual evidence. Yet relying on the image does not guarantee that visual claims are supported by it. Before RL training, 27.81% of the correctly answered responses of Qwen2.5-VL-7B on four multimodal reasoning benchmarks contain at least one direct visual claim that the image does not support. Since outcome-level RL rewards each response as a whole, these claims inherit the positive credit of the correct answer. We introduce a fixed-rollout counterfactual diagnostic that re-scores the same response under an intervened image to separate Evidence-Function Sensitivity (EFS), how strongly the model's predictions change, from claim persistence, whether the model keeps supporting the same claim rather than retracting it. The diagnostic reveals Sensitivity-Persistence Decoupling (SPD): under DAPO and VPPO, EFS increases and claims become more retractable overall, yet unsupported claims become significantly more persistent, whereas GRPO raises EFS without this deterioration. We therefore propose Persistence-Aware Credit Gating (PACG), which attenuates positive credit for unusually persistent visual claims and leaves all other credit unchanged. It requires no supported/unsupported labels and adds no inference cost. On Qwen2.5-VL-7B, PACG raises the nine-benchmark average over three seeds from 58.1% to 59.9% with DAPO and from 59.8% to 60.9% with VPPO, while making unsupported claims more retractable. The gains extend to a larger model, a newer backbone, and the accuracy of HallusionBench also improves consistently. These results suggest that visual sensitivity and claim retractability are complementary dimensions of multimodal credit assignment.

Reinforcement learning with verifiable rewards (RLVR) has become a prominent approach for improving language-model performance on reasoning tasks using automatically checked answers. Yet convention-matched evaluation cannot reveal whether reinforcing one reporting convention reduces adherence to a different request that the initial policy already follows. To test this, we train matched policies under two reporting conventions and evaluate each policy under both current requests, using the same initial policy as a shared reference. We complement this crossed design with controlled interventions and independent human calibration. On GSM8K, boxed-format RLVR reduces the fraction of Qwen2.5-7B responses containing the requested hash-format payload by 35.33--74.37 percentage points relative to a 95.45% initial baseline in four of five training seeds; the fifth improves by 2.50 points. In the four deteriorating runs, almost every response that omits the requested payload instead retains the trained boxed convention, and the same four seeds deteriorate under two fixed paraphrases. Changing only the final-answer marker in supervised targets reverses which reporting convention the model prefers across three seeds, providing controlled evidence that this preference is learnable. Across three settings with independent human calibration, gains under a convention-sensitive scorer exceed the corresponding gains in committed-answer correctness, i.e., the correctness of the answer the model actually commits to. Together, these results separate three distinct post-training outcomes: learned reporting preference, current-request adherence, and committed-answer correctness. They show that convention-matched accuracy alone does not fully characterize post-training behavior and motivate evaluating current-request adherence alongside convention-matched task accuracy.

Process reward models (PRMs) have become a key component for LLMs, as their step-level feedback supports both post-training and test-time reasoning. However, training strong PRMs remains costly: human step annotation is difficult to scale, while Monte Carlo estimation is computationally expensive and can drift from the intrinsic correctness of steps. To get effective PRMs at low cost, we introduce TIPS (Thinking-Induced Process Supervision), an outcome-only reinforcement learning (RL) framework for training generative PRMs. In TIPS, the model generates a chain-of-thought (CoT) followed by step-level labels and an outcome label. The reward depends solely on whether the predicted outcome matches the ground truth, and the resulting group-relative advantage is used to optimize the entire generated response. Intuitively, when checking intermediate steps helps determine the outcome, more accurate checks can lead to better outcome judgments and higher rewards. Outcome-only RL can therefore reinforce step-level verification without explicit process supervision. We validate the effectiveness of TIPS across math and agent benchmarks and four backbone families. Notably, TIPS-Qwen3-4B-Thinking-2507 reaches 85.2 F1 on ProcessBench with only 3.2K outcome-labeled trajectories, surpassing all evaluated trained PRMs and strong prompt-only judges such as GPT-5.4-Instruct and Claude-4.7-Opus, while still trailing o1-mini. Code and data are available at https://github.com/RUCBM/TIPS.

Self-rewarding reinforcement learning (RL) enables large language models (LLMs) to self-evolve without human labels. Existing ensemble-based methods construct reward references from rollout groups and assign rewards accordingly. However, a response's reward representation also depends on its randomly sampled group context, i.e., the other responses in its group. Using only one group-context realization may miss desired reward signals and provide unreliable guidance for policy optimization. To address this issue, we propose Group-Marginalized Advantage Estimation (GMAE), which aggregates reward realizations across possible contexts into a response-level distribution and estimates expected advantages. Experiments across eight benchmarks and four base models demonstrate strong performance and cross-domain generalization. GMAE also exhibits stable learning, low extra cost, and good applicability across training datasets and RL backbones.

Reinforcement Learning has significantly advanced the complex reasoning capabilities of MLLMs. However, prevailing RL algorithms suffer a severe failure in emotion reasoning tasks. These methods heavily rely on deterministic hard-label supervision and point-wise isolated evaluation, creating a fundamental gap with the inherently subjective and continuously distributed nature of human emotions. Furthermore, unlike explicit physical objects, emotional states are deeply implicit within visual cues. This abstract nature exacerbates visual hallucinations in MLLMs, leading to plausible yet ungrounded emotional evidence. To address these limitations, we propose Diversity-Aware Subjective Policy Optimization (DSPO), a reinforcement learning framework that jointly promotes subjective affective coverage and visual grounding. First, we construct a context-grounded emotional distribution prior in the VAD space by combining the lexical prior of the annotated emotion with image-specific contextual information. Based on this prior, we introduce a Distribution-Aligned Emotional Diversity Reward (DEDR), which measures the leave-one-out marginal contribution of each candidate emotion within a rollout. DEDR rewards candidates whose inclusion brings the predicted affective set closer to the context-grounded prior, thereby preserving plausible subjective interpretations without encouraging unconstrained dispersion. We further develop Counterfactual Visual Intervention Gating (CVIG), which masks the visual region highlighted in the reasoning process and uses the resulting candidate-wise probability changes to reduce the weights of interpretations unsupported by visual evidence. Extensive experiments demonstrate that DSPO achieves state-of-the-art performance across multiple public benchmarks, especially on the cross-domain performance, i.e., improving +10.8% on average cross-domain accuracy than EMO-R3.

Recent advances in generative planning have made trajectory inpainting a promising approach to offline goal-conditioned reinforcement learning. However, these methods typically specify the planning horizon before generating plan content, even though the appropriate horizon depends on the route itself. A horizon that is too short can force infeasible transitions, whereas one that is too long can introduce redundant motion. We introduce HorizonFlow, a hierarchical planner that treats plan length as an output of generation rather than a prescribed input. Its subgoal route planner guides its action-prefix controller through a sequence of latent subgoals. Both components combine insertion-based generation with flow matching to jointly generate continuous plan content and length, using the partially generated plan to guide token insertion. HorizonFlow reuses the resulting length information to select candidates and steer generation toward shorter plans without a separate learned value model. Across Maze2D, Multi2D, and OGBench navigation and visual manipulation benchmarks, HorizonFlow achieves the highest average performance among the compared methods.

Recently, Group Relative Policy Optimization (GRPO) and its variants have been developed for policy optimization and demonstrated notable performance gains. However, these methods usually incur substantial computational overhead due to per-question multi-rollout sampling and repeated per-token probability evaluation across rollouts. Furthermore, low-information or highly homogeneous trajectories can degrade downstream learning signal efficiency, hindering model optimization and limiting final performance. To address these issues, we propose FastRL, a novel plug-and-play reinforcement learning framework that simultaneously improves training efficiency and the effectiveness of policy learning. Specifically, 1) We introduce an advantage-aware pruning strategy to selectively preserve high-advantage trajectories while maximizing inter-trajectory gradient diversity. 2) Then, we design an adaptive rollout sampling mechanism to dynamically adjust the sampling scale across different training stages based on historical pruning distributions, balancing exploration adequacy and computational efficiency. Experiments demonstrate that FastRL can be seamlessly integrated into GRPO, DAPO, and GSPO variants, achieving an average 2.07$\times$ training speedup on Geometry3K and GeoQA8K-R1V, along with an approximately 1.64% improvement in average accuracy on visual reasoning benchmarks. Source codes will be available at https://github.com/Nicozwy/FastRL.

Group-relative policy optimization relies on reward-derived advantages and sequence-level likelihood weights, both of which can be sensitive to localized outliers. Extreme rewards can collapse the contrast among clean responses after group normalization, while token-level log-ratio perturbations can alter sequence weights and clipping decisions. We introduce RoVR-GSPO, a dual-channel robust optimizer that addresses these failure modes separately. Its reward channel combines robust reference estimation with bounded residual credit, while its ratio channel uses differentiable SoftRoVR aggregation to construct robust sequence weights. We provide stability and efficiency analyses for both channels. Experiments on mathematical reasoning, long-context summarization, and tool-call annotation show consistent improvements over GSPO, while controlled perturbation studies demonstrate stronger robustness to reward contamination and token-ratio anomalies.

Streaming video assistance requires models to answer asynchronous questions from an observed prefix under a fixed context budget. Existing approaches model response timing or compress history, but an online state formed before future questions are known can omit visual details before later questions reveal their relevance; the retained state alone cannot recover them. We introduce Watch-Think-Interact (WTI), a closed-loop framework for multi-question streaming video reasoning. WTI maintains compact natural-language memory entries tagged with source-video time ranges; these entries support direct reasoning when sufficient and otherwise anchor selective recall of finer visual evidence. For each question, WTI answers when current context and memory suffice, continues watching when required evidence has not appeared, or recalls a relevant past interval and decides again after incorporating the returned chunks, without replaying the full observed history. To train this behavior, we construct WTI-82K, comprising 82,335 timed questions across 4,812 causally aligned trajectories, and develop Stream-GDPO to optimize complete multi-question streaming rollouts using trajectory-level feedback for response timing, source-video recall, and memory updates. WTI achieves state-of-the-art aggregate performance among the compared open-source streaming baselines, reaching 83.3% on StreamingBench and 73.6% weighted overall accuracy on OVO-Bench.