RL for Language Model Reasoning

Reinforcement learning (RL) has emerged as a promising paradigm for training reasoning-oriented models by leveraging rule-based reward signals. However, RL training typically tends to improve single-sample success rates (i.e., Pass@1) while offering limited exploration of diverse reasoning trajectories, which is crucial for multi-sample performance (i.e., Pass@k). Our preliminary analysis reveals that this limitation stems from a fundamental squeezing effect, whereby probability mass is excessively concentrated on a narrow subset of high-reward trajectories, restricting genuine exploration and constraining attainable performance under RL training. To address this issue, in this work, we propose Steering Probability Squeezing (SPS), a training paradigm that interleaves conventional RL with inverse reinforcement learning (IRL). SPS treats on-policy rollouts as demonstrations and employs IRL to explicitly reshape the induced trajectory distribution, thereby enhancing exploration without introducing external supervision. Experiments on five commonly used reasoning benchmarks demonstrate that SPS can enable better exploration and improve Pass@k. Beyond algorithmic contributions, we provide an analysis of RL learning dynamics and identify an empirical upper bound on Pass@k, shedding light on intrinsic exploration limits in RL-based reasoning models. Our findings suggest that alternating between RL and IRL offers an effective pathway toward extending the exploration capacity of reasoning-oriented large language models.

Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a promising approach to improve the reasoning abilities of Large Language Models (LLMs). Among RLVR algorithms, Group Relative Policy Optimization (GRPO) and its variants have demonstrated strong performance and high training efficiency. However, GRPO-style objectives exhibit two issues on high accuracy prompts including mastered prompts (rollout accuracy =1) and majority-correct prompts (rollout accuracy in (0.5,1)). For mastered prompts, group-relative advantages vanish, yielding no training signal and unconstrained policy drift that can cause forgetting. For majority-correct prompts, the induced query weight shrinks as accuracy increases, weakening consolidation from partial correctness to mastery. To alleviate this, we propose Mastery-Consolidated Policy Optimization (MCPO), which introduces (i) a hinge-KL regularizer applied exclusively to mastered prompts to bound harmful policy drift between successive gradient steps, and (ii) a weighting mechanism that prioritizes majority-correct prompts to better allocate optimization effort. Extensive experiments across three mathematical benchmarks demonstrate that MCPO consistently improves pass@1 performance. Counter-intuitively, rather than restricting exploration, MCPO boosts pass@k metrics, indicating that mastery consolidation further catalyzes solution diversity.

Large reasoning models that use long chain-of-thought excel at problem-solving yet waste compute on redundant checks. Curbing this overthinking is hard: training-time length penalties can cripple ability, while inference-time early-exit adds system overhead. To bridge this gap, we propose Step-GRPO, a novel post-training framework that internalizes dynamic early-exit capabilities directly into the model. Step-GRPO shifts the optimization objective from raw tokens to semantic steps by utilizing linguistic markers to structure reasoning. We introduce a Dynamic Truncated Rollout mechanism that exposes the model to concise high-confidence trajectories during exploration, synergized with a Step-Aware Relative Reward that dynamically penalizes redundancy based on group-level baselines. Extensive experiments across three model sizes on diverse benchmarks demonstrate that Step-GRPO achieves a superior accuracy-efficiency trade-off. On Qwen3-8B, our method reduces token consumption by 32.0% compared to the vanilla model while avoiding the accuracy degradation observed in traditional length-penalty methods.

Although most of the automated theorem-proving approaches depend on formal proof systems, informal theorem proving can align better with large language models' (LLMs) strength in natural language processing. In this work, we identify a primary bottleneck in informal theorem proving as a lack of insight, namely the difficulty of recognizing the core techniques required to solve complex problems. To address this, we propose $\texttt{DeepInsight}$, a unified training framework designed to cultivate this essential reasoning skill and enable LLMs to perform insightful reasoning. Our framework consists of three components: (1) $\texttt{DeepInsightTheorem}$, a hierarchical dataset that structures informal proofs by explicitly extracting core techniques and proof sketches alongside the final proof; (2) a Progressive Multi-Stage SFT strategy that mimics the human learning process, teaching the model proof writing, planning, and insight identification; and (3) $\texttt{InsightPO}$, a policy optimization method that assigns structured rewards over this insight hierarchy. Our experiments on challenging mathematical benchmarks demonstrate that this insight-aware generation strategy significantly outperforms baselines. These results demonstrate that teaching models to identify and apply core techniques can substantially improve their mathematical reasoning.

Frontier models have demonstrated exceptional capabilities following the integration of task-reward-based reinforcement learning (RL) into their training pipelines, enabling systems to evolve from pure reasoning models into sophisticated agents. However, debate persists regarding whether RL genuinely instills new skills within a base model or merely sharpens its existing distribution to elicit latent capabilities. To address this dichotomy, we present an explicit comparison between distribution sharpening and task-reward-based learning, utilizing RL as a tool to implement both paradigms. Our analysis reveals the inherent limitations of distribution sharpening, demonstrating from first principles how and why the optima can be unfavorable and the approach fundamentally unstable. Furthermore, our experiments using Llama-3.2-3B-Instruct, Qwen2.5-3B-Instruct and Qwen3-4B-Instruct-2507 on math datasets confirm that sharpening yields limited gains, whereas incorporating task-based reward signal can greatly help achieve robust performance improvements and stable learning.

Large language models (LLMs) increasingly rely on chain-of-thought (CoT) reasoning to solve complex tasks, yet a reasoning trace may merely accompany the final answer rather than causally contribute to it or faithfully reflect how it was reached. We introduce differentiable attention manipulation to train salient reasoning with reinforcement learning -- method which we call AtManRL. By optimizing an additive attention mask that identifies the CoT tokens needed to recover the model's answer, we derive a saliency reward and combine it with an outcome reward in GRPO with decoupled advantages. Because both saliency and faithfulness depend strongly on trace length, we evaluate against a sweep of length-penalty baselines that traces the trade-off a pure length penalty can reach, rather than against a single baseline. Across three models (Qwen2.5-3B, Granite-3.3-2B, and Phi-4-mini) and four benchmarks (GSM8K, MMLU, ProofWriter, and QuestBench), AtManRL sits above this trade-off curve by 3.9-4.0pp in saliency, measured as answer changes under mistake injection, and by 5.7-7.7pp in faithfulness, measured as hint verbalization under hint injection, while matching the accuracy of baselines of similar length. These results suggest that training explicitly for saliency is a promising route towards chain-of-thought reasoning that is more faithful to the model's actual decision process.

Verifiers have been demonstrated to enhance LLM reasoning via test-time scaling (TTS). Yet, they face significant challenges in complex domains. Error propagation from incorrect intermediate reasoning can lead to false positives for seemingly plausible solutions, while lacking external grounding makes verifiers unreliable on computation or knowledge-intensive tasks. To address these challenges, we propose Agentic Verifier, a framework that transforms reward modeling into a multi-turn, tool-augmented deliberative process. We introduce complementary forward and backward agents: one traces solutions from premises to conclusions, while the other re-checks conclusions against their underlying premises. This bidirectional process enables a comprehensive, reliable, and interpretable assessment of solutions. To facilitate practical deployment, we propose AgentV-RL. Through proactive exploration and reinforcement learning, the verifier autonomously interleaves tool-use with internal reasoning. Extensive experiments show that Agentic Verifier yields consistent performance gains under both parallel and sequential TTS. Notably, our 4B variant surpasses state-of-the-art ORMs by 25.2%, positioning it as a promising paradigm for agentic reward modeling.

Reinforcement learning has become a powerful approach for enhancing large language model reasoning, but faces a fundamental dilemma: training on easy problems can cause overfitting and pass@k degradation, while training on hard problems often results in sparse rewards. Recent question augmentation methods address this by prepending partial solutions as hints. However, uniform hint provision may introduce redundant information while missing critical reasoning bottlenecks, and excessive hints can reduce reasoning diversity, causing pass@k degradation. We propose PieceHint, a hint injection framework that strategically identifies and provides critical reasoning steps during training. By scoring the importance of different reasoning steps, selectively allocating hints based on problem difficulty, and progressively withdrawing scaffolding, PieceHint enables models to transition from guided learning to independent reasoning. Experiments on six mathematical reasoning benchmarks show that our 1.5B model achieves comparable average performance to 32B baselines while preserving pass@k diversity across all $k$ values.

As reinforcement Learning with Verifiable Rewards (RLVR) has become the dominant paradigm for scaling reasoning capabilities in LLMs, a new failure mode emerges: LLMs gaming verifiers. We study this phenomenon on inductive reasoning tasks, where models must induce and output logical rules. We find that RLVR-trained models systematically abandon rule induction. Instead of learning generalizable patterns (e.g., ``trains carrying red cars go east''), they enumerate instance-level labels, producing outputs that pass verifiers without capturing the relational patterns required by the task. We show that this behavior is not a failure of understanding but a form of reward hacking: imperfect verifiers that check only extensional correctness admit false positives. To detect such shortcuts, we introduce Isomorphic Perturbation Testing (IPT), which evaluates a single model output under both extensional and isomorphic verification, where the latter enforces invariance under logically isomorphic tasks. While genuine rule induction remains invariant, shortcut strategies fail. We find that shortcut behavior is specific to RLVR-trained reasoning models (e.g., GPT-5, Olmo3) and absent in non-RLVR models (e.g., GPT-4o, GPT-4.5, Ministral). Moreover, shortcut prevalence increases with task complexity and inference-time compute. In controlled training experiments, extensional verification directly induces shortcut strategies, while isomorphic verification eliminates them. These results show that RLVR can incentivize reward hacking not only through overt manipulation but also by exploiting what the verifier fails to enforce.

Reinforcement learning has emerged as an effective paradigm for training large language models to perform search-augmented reasoning. However, existing approaches rely on trajectory-level rewards that cannot distinguish precise search queries from vague or redundant ones within a rollout group, and collapse to a near-zero gradient signal whenever every sampled trajectory fails. In this paper, we propose IG-Search, a reinforcement learning framework that introduces a step-level reward based on Information Gain (IG). For each search step, IG measures how much the retrieved documents improve the model's confidence in the gold answer relative to a counterfactual baseline of random documents, thereby reflecting the effectiveness of the underlying search query. This signal is fed back to the corresponding search-query tokens via per-token advantage modulation in GRPO, enabling fine-grained, step-level credit assignment within a rollout. Unlike prior step-level methods that require either externally annotated intermediate supervision or shared environment states across trajectories, IG-Search derives its signals from the policy's own generation probabilities, requiring no intermediate annotations beyond standard question-answer pairs. Experiments on seven single-hop and multi-hop QA benchmarks demonstrate that IG-Search achieves an average EM of 0.430 with Qwen2.5-3B, outperforming the strongest trajectory-level baseline (MR-Search) by 1.6 points and the step-level method GiGPO by 0.9 points on average across benchmarks, with particularly pronounced gains on multi-hop reasoning tasks. Despite introducing a dense step-level signal, IG-Search adds only ~6.4% to per-step training wall-clock time over the trajectory-level baseline and leaves inference latency unchanged, while still providing a meaningful gradient signal even when every sampled trajectory answers incorrectly.

Reinforcement Learning (RL) has emerged as a critical driver for enhancing the reasoning capabilities of Large Language Models (LLMs). While recent advancements have focused on reward engineering or data synthesis, few studies exploit the model's intrinsic representation characteristics to guide the training process. In this paper, we first observe the presence of high-magnitude activations within the query and key vectors when processing long contexts. Drawing inspiration from model quantization -- which establishes the criticality of such high-magnitude activations -- and the insight that long-context reasoning inherently exhibits a sparse structure, we hypothesize that these weights serve as the pivotal drivers for effective model optimization. Based on this insight, we propose LongAct, a strategy that shifts from uniform to saliency-guided sparse updates. By selectively updating only the weights associated with these significant activations, LongAct achieves an approximate 8% improvement on LongBench v2 and enhances generalization on the RULER benchmark. Furthermore, our method exhibits remarkable universality, consistently boosting performance across diverse RL algorithms such as GRPO and DAPO. Extensive ablation studies suggest that focusing on these salient features is key to unlocking long-context potential.

Recent advances in reinforcement learning (RL) have improved the reasoning capabilities of large language models (LLMs) and vision-language models (VLMs). However, the widely used Group Relative Policy Optimization (GRPO) consistently suffers from entropy collapse, causing the policy to converge prematurely and lose diversity. Existing exploration methods introduce additional bias or variance during exploration, making it difficult to maintain optimization stability. We propose Unified Entropy Control for Reinforcement Learning (UEC-RL), a framework that provides targeted mechanisms for exploration and stabilization. UEC-RL activates more exploration on difficult prompts to search for potential and valuable reasoning trajectories. In parallel, a stabilizer prevents entropy from growing uncontrollably, thereby keeping training stable as the model consolidates reliable behaviors. Together, these components expand the search space when needed while maintaining robust optimization throughout training. Experiments on both LLM and VLM reasoning tasks show consistent gains over RL baselines on both Pass@1 and Pass@$k$. On Geometry3K, UEC-RL achieves a 37.9% relative improvement over GRPO, indicating that it sustains effective exploration without compromising convergence and underscoring UEC-RL as a key for scaling RL-based reasoning in large models. Our code is available at https://github.com/597358816/UEC-RL.

While reinforcement learning with verifiable rewards (RLVR) significantly enhances LLM reasoning by optimizing the conditional distribution P(y|x), its potential is fundamentally bounded by the base model's existing output distribution. Optimizing the marginal distribution P(y) in the Pre-train Space addresses this bottleneck by encoding reasoning ability and preserving broad exploration capacity. Yet, conventional pre-training relies on static corpora for passive learning, leading to a distribution shift that hinders targeted reasoning enhancement. In this paper, we introduce PreRL (Pre-train Space RL), which applies reward-driven online updates directly to P(y). We theoretically and empirically validate the strong gradient alignment between log P(y) and log P(y|x), establishing PreRL as a viable surrogate for standard RL. Furthermore, we uncover a critical mechanism: Negative Sample Reinforcement (NSR) within PreRL serves as an exceptionally effective driver for reasoning. NSR-PreRL rapidly prunes incorrect reasoning spaces while stimulating endogenous reflective behaviors, increasing transition and reflection thoughts by 14.89x and 6.54x, respectively. Leveraging these insights, we propose Dual Space RL (DSRL), a Policy Reincarnation strategy that initializes models with NSR-PreRL to expand the reasoning horizon before transitioning to standard RL for fine-grained optimization. Extensive experiments demonstrate that DSRL consistently outperforms strong baselines, proving that pre-train space pruning effectively steers the policy toward a refined correct reasoning subspace.

Reinforcement Learning with Verifiable Rewards (RLVR) has catalyzed significant advances in the reasoning capabilities of Large Language Models (LLMs). However, effectively managing the exploration and exploitation trade-off remains a critical challenge. In this paper, we fully analyze the exploration and exploitation dilemma of extremely hard and easy samples during the training and propose a new fine-grained trade-off mechanism. Concretely, we introduce a perplexity space disentangling strategy that divides the sample space into distinct exploration (high perplexity) and exploitation (low perplexity) subspaces, thereby mining fine-grained samples requiring exploration-exploitation trade-off. Subsequently, we propose a bidirectional reward allocation mechanism with a minimum impact on verification rewards to implement perplexity-guided exploration and exploitation, enabling more stable policy optimization. Finally, we have evaluated our method on two mainstream tasks: mathematical reasoning and function calling, and experimental results demonstrate the superiority of the proposed method, confirming its effectiveness in enhancing LLM performance by fine-grained exploration-exploitation trade-off.

Reasoning capabilities in large language models (LLMs) have generally advanced significantly. However, it is still challenging for existing reasoning-based LLMs to perform effective decision-making abilities in multi-agent environments, due to the absence of explicit foresight modeling. To this end, strategic reasoning, the most fundamental capability to anticipate the counterpart's behaviors and foresee its possible future actions, has been introduced to alleviate the above issues. Strategic reasoning is fundamental to effective decision-making in multi-agent environments, yet existing reasoning enhancement methods for LLMs do not explicitly capture its foresight nature. In this work, we introduce Foresight Policy Optimization (FoPO) to enhance strategic reasoning in LLMs, which integrates opponent modeling principles into policy optimization, thereby enabling explicit consideration of both self-interest and counterpart influence. Specifically, we construct two curated datasets, namely Cooperative RSA and Competitive Taboo, equipped with well-designed rules and moderate difficulty to facilitate a systematic investigation of FoPO in a self-play framework. Our experiments demonstrate that FoPO significantly enhances strategic reasoning across LLMs of varying sizes and origins. Moreover, models trained with FoPO exhibit strong generalization to out-of-domain strategic scenarios, substantially outperforming standard LLM reasoning optimization baselines.

Supervised fine-tuning (SFT) followed by Group Relative Policy Optimization (GRPO) is a common post-training recipe. We conduct a controlled ablation over SFT-GRPO data overlap, evaluating Qwen3-8B (thinking disabled) post-trained for Lean 4 autoformalization under six conditions that differ solely in training recipe: a base model, SFT-only, GRPO-only, and three SFT+GRPO configurations where 0 percent, 30 percent, or 100 percent of the GRPO prompts coincide with the SFT corpus. Keeping SFT and GRPO data disjoint consistently outperforms full overlap at zero additional compute cost. Evaluating on Gaokao-Formal and PutnamBench under both compile pass at k and semantic pass at k assessed by an LLM judge, we find that lower overlap is monotonically associated with higher compilation and semantic accuracy. At 0 percent overlap, GRPO yields a 10.4 percentage point semantic gain over SFT alone on Gaokao, while at 100 percent overlap both metrics remain flat, rendering the GRPO stage effectively redundant. We further show that dual-metric evaluation reveals compile semantic gaps exceeding 30 percentage points for the highest compiling models, a disparity invisible under compile-only benchmarking. To our knowledge, this is the first controlled investigation of SFT-GRPO data overlap as a post-training hyperparameter, demonstrating how model behavior varies based on the degree of data sharing between training stages.

While Aspect-based Sentiment Analysis (ABSA) systems have achieved high accuracy in identifying sentiment polarities, they often operate as "black boxes," lacking the explicit reasoning capabilities characteristic of human affective cognition. Humans do not merely categorize sentiment; they construct causal explanations for their judgments. To bridge this gap, we propose ABSA-R1, a large language model framework designed to mimic this ``reason-before-predict" cognitive process. By leveraging reinforcement learning (RL), ABSA-R1 learns to articulate the why behind the what, generating natural language justifications that ground its sentiment predictions. We introduce a Cognition-Aligned Reward Model (formerly sentiment-aware reward model) that enforces consistency between the generated reasoning path and the final emotional label. Furthermore, inspired by metacognitive monitoring, we implement a performance-driven rejection sampling strategy that selectively targets hard cases where the model's internal reasoning is uncertain or inconsistent. Experimental results on four benchmarks demonstrate that equipping models with this explicit reasoning capability not only enhances interpretability but also yields superior performance in sentiment classification and triplet extraction compared to non-reasoning baselines.

Process reward models (PRMs) provide fine-grained supervision for reasoning, but reliable PRMs often require step annotations or heavy verification pipelines, making them costly to scale and refresh during online RL. Implicit PRMs reduce this cost by training log-likelihood-ratio rewards from trajectory-level outcome labels. However, the log-ratio is constrained only as a sequence-level aggregate during training, while inference decomposes it into token- or step-level scores for partial prefixes. This train-inference mismatch leaves local credits weakly identified, so distribution-wide scoring can amplify misleading advantages. We propose Implicit Prefix-Value Reward Model (IPVRM), which directly learns the probability of eventual correctness for each prefix from outcome labels. Step signals are then obtained as temporal-difference (TD) differences between consecutive prefix values, aligning the training target with inference-time use. IPVRM markedly improves step-verification F1 on ProcessBench. To exploit these prefix values during policy optimization, we further introduce Distribution-Level RL (DistRL), which applies TD advantages to both sampled tokens and high-probability candidate tokens, providing dense counterfactual updates without additional rollouts. Experiments show that DistRL brings limited gains with unreliable implicit rewards, but consistently improves downstream reasoning when paired with IPVRM. The implementation of our method is available at https://github.com/gaoshiping/IPVRM .

Group Relative Policy Optimization (GRPO) has significantly advanced the reasoning ability of large language models (LLMs), particularly in their mathemat ical reasoning performance. However, GRPO and related entropy regularization methods still struggle with token-level sparse-rewards, which is an inherent chal lenge in chain-of-thought (CoT) reasoning. These approaches often rely on undifferen tiated token-level entropy regularization, which easily leads to entropy collapse or model degradation under sparse token rewards. In this work, we propose TEPO, a novel token-level framework that (1) leverages sequence-level likelihood to link group-level rewards with individual tokens via token-level aggregation, and (2) introduces a token-level KL-Divergence mask constraint that targets tokens with positive advantages and decreasing entropy to mitigate abrupt policy updates. Experiments demonstrate that TEPO not only achieves state-of-the-art performance on mathematical reasoning benchmarks but also markedly enhances training stability, reducing convergence time by 50% compared with GRPO/DAPO.

Group Relative Policy Optimization (GRPO) enhances LLM reasoning but often induces overconfidence, where incorrect responses yield lower perplexity than correct ones, degrading relative calibration as described by the Area Under the Curve (AUC). Existing approaches either yield limited improvements in calibration or sacrifice gains in reasoning accuracy. We first prove that this degradation in GRPO-style algorithms stems from their uncertainty-agnostic advantage estimation, which inevitably misaligns optimization gradients with calibration. This leads to improved accuracy at the expense of degraded calibration. We then propose Calibration-Aware Policy Optimization (CAPO). It adopts a logistic AUC surrogate loss that is theoretically consistent and admits regret bound, enabling uncertainty-aware advantage estimation. By further incorporating a noise masking mechanism, CAPO achieves stable learning dynamics that jointly optimize calibration and accuracy. Experiments on multiple mathematical reasoning benchmarks show that CAPO-1.5B significantly improves calibration by up to 15% while achieving accuracy comparable to or better than GRPO, and further boosts accuracy on downstream inference-time scaling tasks by up to 5%. Moreover, when allowed to abstain under low-confidence conditions, CAPO achieves a Pareto-optimal precision-coverage trade-off, highlighting its practical value for hallucination mitigation.

RLVR improves reasoning in large language models, but its effectiveness is often limited by severe reward sparsity on hard problems. Recent hint-based RL methods mitigate sparsity by injecting partial solutions or abstract templates, yet they typically scale guidance by adding more tokens, which introduce redundancy, inconsistency, and extra training overhead. We propose KnowRL (Knowledge-Guided Reinforcement Learning), an RL training framework that treats hint design as a minimal-sufficient guidance problem. During RL training, KnowRL decomposes guidance into atomic knowledge points (KPs) and uses Constrained Subset Search (CSS) to construct compact, interaction-aware subsets for training. We further identify a pruning interaction paradox -- removing one KP may help while removing multiple such KPs can hurt -- and explicitly optimize for robust subset curation under this dependency structure. We train KnowRL-Nemotron-1.5B from OpenMath-Nemotron-1.5B. Across eight reasoning benchmarks at the 1.5B scale, KnowRL-Nemotron-1.5B consistently outperforms strong RL and hinting baselines. Without KP hints at inference, KnowRL-Nemotron-1.5B reaches 70.08 average accuracy, already surpassing Nemotron-1.5B by +9.63 points; with selected KPs, performance improves to 74.16, establishing a new state of the art at this scale. The model, curated training data, and code are publicly available at https://github.com/Hasuer/KnowRL.

Reinforcement learning with verifiable rewards (RLVR) has become a central paradigm for improving reasoning and code generation in large language models, and GRPO-style training is widely adopted for its simplicity and effectiveness. However, an important design choice remains underexplored: how token-level policy gradient terms are aggregated within each sampled group. Standard GRPO uses sequence aggregation, while recent work has advocated token aggregation as a better alternative. We show that these two rules induce different optimization biases: token aggregation introduces sign-length coupling, while sequence aggregation implicitly downweights longer responses through sequence-level equal weighting. To address this tension, we propose Balanced Aggregation (BA), a simple drop-in replacement that computes token-level means separately within the positive and negative subsets and then combines them with sequence-count-based weights. Experiments with Qwen2.5-Math-7B and Qwen3-1.7B on DAPO-17k and Polaris, evaluated on six reasoning and coding benchmarks, show that BA consistently improves training stability and final performance over standard token and sequence aggregation. Our analysis further shows that the relative effectiveness of token and sequence aggregation is largely governed by response-length variation and the positive-negative length gap, highlighting aggregation as a critical design dimension in GRPO-style RLVR.

Large Language Models (LLMs) exhibit strong abilities in natural language understanding and generation, yet they struggle with knowledge-intensive reasoning. Structured Knowledge Graphs (KGs) provide an effective form of external knowledge representation and have been widely used to enhance performance in classical Knowledge Base Question Answering (KBQA) tasks. However, performing precise multi-hop reasoning over KGs for complex queries remains highly challenging. Most existing approaches decompose the reasoning process into a sequence of isolated steps executed through a fixed pipeline. While effective to some extent, such designs constrain reasoning flexibility and fragment the overall decision process, often leading to incoherence and the loss of critical intermediate information from earlier steps. In this paper, we introduce KG-Reasoner, an end-to-end framework that integrates multi-step reasoning into a unified "thinking" phase of a Reasoning LLM. Through Reinforcement Learning (RL), the LLM is trained to internalize the KG traversal process, enabling it to dynamically explore reasoning paths, and perform backtracking when necessary. Experiments on eight multi-hop and knowledge-intensive reasoning benchmarks demonstrate that KG-Reasoner achieves competitive or superior performance compared to the state-of-the-art methods. Codes are available at the repository: https://github.com/Wangshuaiia/KG-Reasoner.

We have witnessed remarkable advances in LLM reasoning capabilities with the advent of DeepSeek-R1. However, much of this progress has been fueled by the abundance of internet question-answer (QA) pairs, a major bottleneck going forward, since such data is limited in scale and concentrated mainly in domains like mathematics. In contrast, other sciences such as physics lack large-scale QA datasets to effectively train reasoning-capable models. In this work, we show that physics simulators can serve as a powerful alternative source of supervision for training LLMs for physical reasoning. We generate random scenes in physics engines, create synthetic question-answer pairs from simulated interactions, and train LLMs using reinforcement learning on this synthetic data. Our models exhibit zero-shot sim-to-real transfer to real-world physics benchmarks: for example, training solely on synthetic simulated data improves performance on IPhO (International Physics Olympiad) problems by 5-10 percentage points across model sizes. These results demonstrate that physics simulators can act as scalable data generators, enabling LLMs to acquire deep physical reasoning skills beyond the limitations of internet-scale QA data. Code available at: https://sim2reason.github.io/.

Vision-language models (VLMs) have shown impressive capabilities in perceptual tasks, yet they degrade in complex multi-hop reasoning under multiplayer game settings with imperfect and deceptive information. In this paper, we study a representative multiplayer task, Murder Mystery Games, which require inferring hidden truths based on partial clues provided by roles with different intentions. To address this challenge, we propose a collaborative multi-agent framework for evaluating and synthesizing high-quality, role-driven multiplayer game scripts, enabling fine-grained interaction patterns tailored to character identities (i.e., murderer vs. innocent). Our system generates rich multimodal contexts, including character backstories, visual and textual clues, and multi-hop reasoning chains, through coordinated agent interactions. We design a two-stage agent-monitored training strategy to enhance the reasoning ability of VLMs: (1) chain-of-thought based fine-tuning on curated and synthetic datasets that model uncertainty and deception; (2) GRPO-based reinforcement learning with agent-monitored reward shaping, encouraging the model to develop character-specific reasoning behaviors and effective multimodal multi-hop inference. Extensive experiments demonstrate that our method significantly boosts the performance of VLMs in narrative reasoning, hidden fact extraction, and deception-resilient understanding. Our contributions offer a scalable solution for training and evaluating VLMs under uncertain, adversarial, and socially complex conditions, laying the groundwork for future benchmarks in multimodal multi-hop reasoning under imperfect information.

While large language models hold promise for complex medical applications, their development is hindered by the scarcity of high-quality reasoning data. To address this issue, existing approaches typically distill chain-of-thought reasoning traces from large proprietary models via supervised fine-tuning, then conduct reinforcement learning (RL). These methods exhibit limited improvement on underrepresented domains like rare diseases while incurring substantial costs from generating complex reasoning chains. To efficiently enhance medical reasoning, we propose MedSSR, a Medical Knowledge-enhanced data Synthesis and Semi-supervised Reinforcement learning framework. Our framework first employs rare disease knowledge to synthesize distribution-controllable reasoning questions. We then utilize the policy model itself to generate high-quality pseudo-labels. This enables a two-stage, intrinsic-to-extrinsic training paradigm: self-supervised RL on the pseudo-labeled synthetic data, followed by supervised RL on the human-annotated real data. MedSSR scales model training efficiently without relying on costly trace distillation. Extensive experiments on Qwen and Llama demonstrate that our method outperforms existing methods across ten medical benchmarks, achieving up to +5.93% gain on rare-disease tasks. Our code is available at https://github.com/tdlhl/MedSSR.

Reinforcement learning for open-ended text generation is constrained by the lack of verifiable rewards, necessitating reliance on judge models that require either annotated data or powerful closed-source models. Inspired by recent work on unsupervised reinforcement learning for mathematical reasoning using confidence-based endogenous rewards, we investigate whether this principle can be adapted to open-ended writing tasks. We find that directly applying confidence rewards leads to Triviality Bias: the policy collapses toward high-probability outputs, reducing diversity and meaningful content. We propose TCER (Triviality Corrected Endogenous Reward), which addresses this bias by rewarding the relative information gain between a specialist policy and a generalist reference policy, modulated by a probability-dependent correction mechanism. Across multiple writing benchmarks and model architectures, TCER achieves consistent improvements without external supervision. Furthermore, TCER also transfers effectively to mathematical reasoning, validating the generality of our approach across different generation tasks.

Reinforcement Learning with Verifiable Rewards (RLVR) improves the reasoning ability of Large Language Models (LLMs), but sparse outcome rewards make token-level credit assignment difficult. We study token-level credit as a reward-conditioned shift from the behavior policy to a hindsight posterior. In autoregressive RLVR, this shift can be expressed through Conditional Mutual Information (CMI), which shows that token entropy upper-bounds possible hindsight credit. Entropy, however, indicates capacity rather than update direction, so we introduce the Four Quadrant Decomposition to separate updates by reward polarity and token entropy. Controlled interventions show that these two factors jointly shape token updates. Sustained reasoning gains concentrate in signed high-entropy quadrants, whereas low-entropy updates saturate quickly. Based on this analysis, we propose Hindsight-Aware Policy Optimization (HAPO), a sign-preserving modification to GRPO that performs capacity-guided advantage reallocation. Experiments on mathematical reasoning benchmarks in two model settings show that HAPO achieves competitive performance among entropy-aware baselines.

On-policy reinforcement learning has become the dominant paradigm for reasoning alignment in large language models, yet its sparse, outcome-level rewards make token-level credit assignment notoriously difficult. On-Policy Distillation (OPD) alleviates this by introducing dense, token-level KL supervision from a teacher model, but typically applies this supervision uniformly across all rollouts, ignoring fundamental differences in signal quality. We propose Signal-Calibrated On-Policy Distillation Enhancement (SCOPE), a dual-path adaptive training framework that routes on-policy rollouts by correctness into two complementary supervision paths. For incorrect trajectories, SCOPE performs teacher-perplexity-weighted KL distillation to prioritize instances where the teacher demonstrates genuine corrective capability, while down-weighting unreliable guidance. For correct trajectories, it applies student-perplexity-weighted MLE to concentrate reinforcement on low-confidence samples at the capability boundary rather than over-reinforcing already mastered ones. Both paths employ a group-level normalization to adaptively calibrate weight distributions, accounting for the intrinsic difficulty variance across prompts. Extensive experiments on six reasoning benchmarks show that SCOPE achieves an average relative improvement of 11.42% in Avg@32 and 7.30% in Pass@32 over competitive baselines, demonstrating its consistent effectiveness.

Reinforcement Learning with Verifiable Rewards (RLVR) is an essential paradigm that enhances the reasoning capabilities of Large Language Models (LLMs). However, existing methods typically rely on static policy optimization schemes that misalign with the model's evolving reasoning capabilities. To address this issue, we propose Adaptive Power-Mean Policy Optimization (APMPO), which comprises two main innovations: Power-Mean Policy Optimization (PMPO) and Feedback-Adaptive Clipping (FAC). Specifically, PMPO introduces a generalized power-mean objective. This enables the model to adaptively transition from the signal-amplifying behavior of the arithmetic mean to the consistency-enforcing behavior of the geometric mean. FAC adaptively adjusts clipping bounds based on real-time reward statistics to overcome the limitations of static mechanisms. Capitalizing on these innovations, APMPO improves learning dynamics and reasoning performance. Extensive experiments on nine datasets across three reasoning tasks showcase the superiority of APMPO over state-of-the-art RLVR-based baselines. For instance, APMPO boosts the average Pass@1 score on mathematical reasoning benchmarks by 3.0 points compared to GRPO when using Qwen2.5-3B-Instruct.