RL for Language Model Reasoning

Evaluation of multimodal reasoning models is typically reduced to a single accuracy score, implicitly treating reasoning as a unitary capability. We introduce MathLens, a benchmark of textbook-style geometry problems that exposes this assumption by operationally decomposing performance into perception, reasoning, and multimodal-specific components. Each problem is derived from a symbolic specification and accompanied by visual diagrams, text-only variants, multimodal questions, and targeted perceptual probes, enabling controlled measurement of each component. Using this decomposition, we show that common training strategies are associated with systematically different capability profiles that are invisible under aggregate accuracy. Across the models we study, reinforcement learning primarily improves perceptual grounding and robustness to diagram variation, while our analysis suggests that textual SFT yields gains through reflective reasoning. In contrast, as perception and reasoning improve, a growing fraction of remaining errors fall outside these components and are categorized as multimodal-specific. These results suggest that apparent progress in multimodal reasoning reflects shifting balances among subskills rather than uniform advancement, motivating evaluation beyond scalar accuracy.

Recent thinking models are capable of solving complex reasoning tasks by scaling test-time compute, but this scaling should be allocated in line with task difficulty. On one hand, short reasoning (underthinking) leads to errors on harder problems that require extended reasoning steps; but, excessively long reasoning (overthinking) can be token-inefficient by generating unnecessary steps even after reaching a correct intermediate solution. We refer to this as under-adaptivity, where the model fails to modulate its response length appropriately given problems of varying difficulty. To address under-adaptivity and strike a balance between under- and overthinking, we propose TRAAC (Think Right with Adaptive, Attentive Compression), an online post-training RL method that leverages the model's self-attention to identify key steps and prune redundant ones. TRAAC also estimates difficulty and incorporates it into training rewards, thereby learning to allocate a reasoning budget commensurate with example difficulty. Across a variety of tasks (AIME, AMC, GPQA-D, BBEH), TRAAC (Qwen3-4B) achieves an average absolute accuracy gain of 8.4% with a relative reduction in reasoning length of 36.8% compared to the base model, and a 7.9% accuracy gain paired with a 29.4% length drop compared to the best RL baseline. TRAAC generalizes well, with accuracy and efficiency gains on out-of-distribution non-math datasets like GPQA-D, BBEH, and OptimalThinkingBench. Our analysis shows that TRAAC learns to adjust its thinking budget based on difficulty and that a combination of task-difficulty calibration and attention-based compression yields gains across diverse tasks.

Reinforcement Learning with Verifiable Rewards (RLVR) replaces costly human labeling with automated verifiers. To reduce verifier hacking, many RLVR systems binarize rewards to $\{0,1\}$, but imperfect verifiers inevitably introduce false negatives (rejecting correct answers) and false positives (accepting incorrect ones). We formalize verifier unreliability as a stochastic reward channel with asymmetric noise rates $ρ_0$ and $ρ_1$ -- the FP rate and the FN rate, respectively. From this abstraction we derive two lightweight corrections: (i) a backward correction that yields an unbiased surrogate reward and thus an unbiased policy-gradient estimator in expectation, and (ii) a forward correction that reweights score-function terms so the expected update aligns with the clean gradient direction and requires only the FN rate. We implement both as lightweight hooks in a group relative policy optimization pipeline, both corrections improve RLVR for math reasoning under synthetic and real verifier noise, with the forward variant being more stable under heavier noise. Finally, an appeals mechanism with a lightweight LLM verifier estimates the FN rate online and further improves performance.

Existing multi-agent learning approaches have developed interactive training environments to explicitly promote collaboration among multiple Large Language Models (LLMs), thereby constructing stronger multi-agent systems (MAS). However, during inference, they require re-executing the MAS to obtain final solutions, which diverges from human cognition that individuals can enhance their reasoning capabilities through interactions with others and resolve questions independently in the future. To investigate whether multi-agent interaction can enhance LLMs' independent problem-solving ability, we introduce ILR, a novel co-learning framework for MAS that integrates two key components: Dynamic Interaction and Perception Calibration. Specifically, Dynamic Interaction first adaptively selects either cooperative or competitive strategies depending on question difficulty and model ability. LLMs then exchange information through Idea3, an innovative interaction paradigm designed to mimic human discussion, before deriving their respective final answers. In Perception Calibration, ILR employs Group Relative Policy Optimization (GRPO) to train LLMs while integrating one LLM's reward distribution characteristics into another's reward function, thereby enhancing the cohesion of multi-agent interactions. We evaluate the effectiveness of ILR across three LLMs from two model families of varying scales on five mathematical, one coding, one general question answering, and one scientific reasoning benchmarks. Experimental results show that ILR consistently outperforms single-agent learning, yielding an improvement of up to 5% over the strongest baseline. We further discover that Idea3 can enhance the robustness of stronger LLMs during multi-agent inference, and dynamic interaction types can boost multi-agent learning compared to pure cooperative or competitive strategies.

Reinforcement Learning with Verifiable Reward (RLVR) effectively solves complex tasks but demands extremely long context lengths during training, leading to substantial computational costs. While multi-stage training can partially mitigate this, starting with overly short contexts often causes irreversible performance degradation, ultimately failing to reduce overall training compute significantly. In this paper, we introduce Thinking-Free Policy Initialization (TFPI), a simple yet effective adaptation to RLVR that bridges long Chain-of-Thought (CoT) distillation and standard RLVR. TFPI employs a simple ThinkingFree operation, explicitly discarding the thinking content via a direct </think> append, to reduce token usage during inference. Training with ThinkingFree-adapted inputs improves performance and lowers token consumption, even in the original slow-thinking mode. Extensive experiments across various benchmarks have shown that TFPI accelerates RL convergence, achieves a higher performance ceiling, and yields more token-efficient reasoning models without specialized rewards or complex training designs. With TFPI only, we train a 4B model to reach 89.0% accuracy on AIME24 and 65.5% on LiveCodeBench using less than 4K H20 hours.

Recent advancements have endowed Large Language Models with impressive general reasoning capabilities. However, these reasoning models often perform worse than non-reasoning models on personalization tasks. While some methods use outcome-based RL to improve personalization reasoning, they fail to supervise the reasoning process. As a result, models may reach correct answers through flawed reasoning chains, limiting further improvement. To address this, we propose TagPR, a novel framework that adds semantic tags to the reasoning process for step-by-step guidance. TagPR first automatically generates a structured, tagged dataset for Supervised Fine-Tuning. It then employs a multi-stage RL process guided by a composite reward signal, which integrates tag-based process supervision with a novel Personalization Reward Model with User Embeddings to achieve fine-grained alignment with user-specific logic. Extensive experiments on public LaMP, LongLaMP, PGraphRAG, and a self-constructed dataset demonstrate that our approach achieves state-of-the-art results, delivering an average improvement of 32.65% over the base model across all LaMP benchmark tasks. Our work demonstrates that tag-guided process supervision is an effective approach for personalization reasoning.

Large language models (LLMs) can improve through reinforcement learning with verifiable rewards (RLVR), but many settings lack reference answers or external verifiers. Label-free self-improvement instead constructs training signals from the model's own outputs. Because the same policy generates both the candidate responses and the signals derived from them, each update also changes the candidates available for future training, creating a self-reinforcing feedback loop. We identify a failure mode of this loop, which we call selection-induced diversity collapse: repeated self-selection can progressively narrow the candidate pool and make alternative correct solutions harder to sample. To address this problem, we propose EVOL-RL, which balances selection and variation. Majority voting provides a stable selection signal, while embedding-based novelty refines credit among responses with the same selection status; entropy regularization and asymmetric clipping further support this process by maintaining diverse rollouts and preserving rare, high-advantage updates. Across multiple training distributions, model scales, backbones, and matched label-free baselines, EVOL-RL consistently improves pass@n while generally maintaining or improving pass@1. For example, when trained without labels on AIME24, Qwen3-4B-Base improves over TTRL on held-out AIME25 from 4.6% to 17.1% in pass@1 and from 18.5% to 42.0% in pass@16. Training analyses further show that when the current majority is wrong, EVOL-RL better preserves alternative correct solutions. The code is available at: https://github.com/YujunZhou/EVOL-RL.

Medical question answering requires advanced reasoning that integrates domain knowledge with logical inference. However, existing large language models (LLMs) often generate reasoning chains that lack factual accuracy and clinical reliability. We propose Ranked Preference Reinforcement Optimization (RPRO), a novel framework that combines reinforcement learning with preference-driven reasoning refinement to enhance clinical chain-of-thought (CoT) performance. RPRO distinguishes itself from prior approaches by employing task-adaptive reasoning templates and a probabilistic evaluation mechanism that aligns model outputs with established clinical workflows, while automatically identifying and correcting low-quality reasoning chains. Unlike traditional pairwise preference methods, RPRO introduces a groupwise ranking optimization based on the Bradley--Terry model and incorporates KL-divergence regularization for stable training. Experiments on PubMedQA, MedQA-USMLE, and a real-world clinical dataset from Far Eastern Memorial Hospital (FEMH) demonstrate consistent improvements over strong baselines. Remarkably, our 2B-parameter model outperforms much larger 7B--20B models, including medical-specialized variants. These findings demonstrate that combining preference optimization with quality-driven refinement provides a scalable and clinically grounded approach to building more reliable medical LLMs.

With respect to improving the reasoning accuracy of LLMs, the representative reinforcement learning (RL) method - Group Relative Policy Optimization (GRPO) - has achieved critical success, yet it still suffers from the issue of insignificant reward signals. This paper introduces ReST-RL, a unified Reinforced Self-Training (ReST) policy-value framework that reconnects policy optimization and value-guided search to improve LLM reasoning ability. Firstly, ReST-GRPO adopts an optimized ReST-style algorithm to reshape the policy-induced trajectory distribution by increasing the reward variance of GRPO sampling and exposing the policy to more informative partial states, thereby improving training efficiency and effectiveness. Then, we further introduce a decoding optimization method, VM-MCTS, which trains a Value Model (VM) from self-collected Monte-Carlo Tree Search (MCTS) targets and deploys it through an adapted MCTS algorithm to provide precise process signals and verification scores, further enhancing reasoning accuracy. These two stages are internally dependent - ReST-GRPO yields higher-quality trajectories for value learning with VM-MCTS, which in turn enables more effective inference-time search. We validate our framework on multiple coding benchmarks (e.g., APPS, BigCodeBench, and HumanEval), where it significantly outperforms other reinforcement training baselines (naive GRPO, DAPO, and ReST-DPO), as well as decoding and verification baselines (e.g., PRM-BoN and ORM-MCTS), indicating its power to strengthen LLM reasoning capability. Moreover, we further evaluate ReST-RL on out-of-domain math and science reasoning tasks, where it achieves improved performance without target-domain tuning and favorable end-to-end efficiency trade-offs, providing preliminary transfer evidence beyond our primary coding domain.

Advanced reasoning in LLMs on challenging domains like mathematical reasoning can be tackled using verifiable rewards based reinforced fine-tuning (ReFT). In standard ReFT frameworks, a behavior model generates multiple completions with answers per problem, for the answer to be then scored by a reward function. While such RL post-training methods demonstrate significant performance improvements across challenging reasoning domains, the computational cost of generating completions during training with multiple inference steps makes the training cost non-trivial. To address this, we draw inspiration from off-policy RL, and speculative decoding to introduce a novel ReFT framework, dubbed Nested-ReFT, where a subset of layers of the target model acts as the behavior model to generate off-policy completions during training. The behavior model configured with dynamic layer skipping per batch during training decreases the inference cost compared to the standard ReFT frameworks. Our theoretical analysis shows that Nested-ReFT yields unbiased gradient estimates with controlled variance. Our empirical analysis demonstrates improved computational efficiency measured as tokens/sec across multiple math reasoning benchmarks and model sizes. Additionally, we explore three variants of bias mitigation to minimize the off-policyness in the gradient updates that allows for maintaining performance that matches the baseline ReFT performance.

Large Reasoning Models (LRMs) can exhibit step-by-step reasoning, reflection, and backtracking, but these behaviors are often unregulated, leading to overthinking. As a result, LRMs continue generating redundant reasoning even after reaching high-confidence conclusions. This increases inference cost and latency, limiting practical deployment. The root cause is the absence of an intrinsic mechanism to monitor the reasoning state and decide when to continue, backtrack, or stop. We propose MERA, a meta-cognitive reasoning framework that decouples reasoning from control to enable independent optimization of control strategies. MERA constructs high-quality reasoning-control supervision data via a takeover-based pipeline, and transforms long-horizon traces into structured reasoning-control alternating sequences for training. The model is trained with supervised fine-tuning to internalize the structured separation, and further optimized with Control-Segment Policy Optimization (CSPO), which combines segment-wise GRPO with control masking to focus learning on control segments. Experiments across reasoning benchmarks show that MERA improves both efficiency and accuracy.

Large reasoning models (LRMs) often exhibit overthinking, producing verbose Chain-of-Thought (CoT) traces that increase inference cost and obscure the underlying reasoning process. Existing CoT compression methods mainly rely on global length rewards, which conflate necessary intermediate reasoning with redundant text and may therefore compromise reasoning fidelity. This paper revisits overthinking from a semantic-efficiency perspective and decomposes CoT redundancy into two distinct forms: internal redundancy, defined as informational stagnation before the first correct answer, and external redundancy, defined as superfluous continuation after the first correct answer. Based on this decomposition, we propose a dual-penalty reinforcement learning framework that separately optimizes reasoning progress and termination behavior. Specifically, a sliding-window semantic similarity metric penalizes low-progress reasoning segments, while a normalized external-redundancy metric discourages post-answer continuation. Experiments on GSM8K, MATH500, and AIME24 across different model scales show that our method reduces average reasoning length by 41.3% on the 1.5B model and 40.1% on the 7B model, while preserving competitive accuracy and achieving the best overall accuracy-efficiency score among evaluated baselines. The learned compression behavior further transfers to out-of-domain reasoning tasks, including GPQA and LiveCodeBench. More importantly, our analysis reveals a clear asymmetry between the two redundancy types: external redundancy can be largely removed with little performance loss, whereas internal redundancy compression follows a sensitive accuracy-efficiency trade-off. These results suggest that effective CoT compression should optimize semantic efficiency rather than sequence length alone, offering a principled route toward more concise, efficient, and interpretable LRMs.

The proliferation of Large Language Models (LLMs) in medicine has enabled impressive capabilities, yet a critical gap remains in their ability to perform systematic, transparent, and verifiable reasoning, a cornerstone of clinical practice. This has catalyzed a shift from single-step answer generation to the development of LLMs explicitly designed for medical reasoning. This paper provides the first systematic review of this emerging field. We propose a taxonomy of reasoning enhancement techniques, categorized into training-time strategies (e.g., supervised fine-tuning, reinforcement learning) and test-time mechanisms (e.g., prompt engineering, multi-agent systems). We analyze how these techniques are applied across different data modalities (text, image, code) and in key clinical applications such as diagnosis, education, and treatment planning. Furthermore, we survey the evolution of evaluation benchmarks from simple accuracy metrics to sophisticated assessments of reasoning quality and visual interpretability. Based on an analysis of 60 seminal studies from 2022-2025, we conclude by identifying critical challenges, including the faithfulness-plausibility gap and the need for native multimodal reasoning, and outlining future directions toward building efficient, robust, and sociotechnically responsible medical AI.

Large Language Models (LLMs) often produce plausible but poorly-calibrated answers, limiting their reliability on reasoning-intensive tasks. Recent research suggests that Chain-of-Thought (CoT) reasoning paths are inherent in pre-trained LLMs and can be elicited by simply altering the decoding process, where the presence of a CoT path correlates with higher answer confidence. Building on these insights, we present Reinforcement Learning from Self-Feedback (RLSF), a post-training stage that utilises the model's intrinsic confidence as a self-generated reward. By generating multiple CoT decoding beams from a frozen LLM, we compute the confidence of each final answer span and rank the resulting traces accordingly to create synthetic preferences. These preferences are subsequently utilised to fine-tune the policy through standard preference optimisation, requiring no human labels, gold answers, or externally curated rewards. RLSF simultaneously (i) refines the model's probability estimates--restoring well-behaved calibration--and (ii) strengthens step-by-step reasoning, yielding improved performance on arithmetic reasoning and multiple-choice question answering. By converting a model's own uncertainty into structured self-feedback, RLSF affirms reinforcement learning on intrinsic model behaviour as a principled and data-efficient component of the LLM post-training pipeline. Our results demonstrate that leveraging these inherent reasoning capabilities provides a robust path for enhancing model reliability without manual prompt engineering or external supervision.

Existing informal language-based (e.g., human language) Large Language Models (LLMs) trained with Reinforcement Learning (RL) face a significant challenge: their verification processes, which provide crucial training signals, are neither reliable nor scalable. In fact, the prevalent large proprietary models could hardly generate verifiable programs. A promising yet largely uncharted alternative is formal language-based reasoning. Grounding LLMs in rigorous formal systems where generative models operate in formal language spaces (e.g., Dafny) enables the automatic and mathematically provable verification of their reasoning processes and outcomes. This capability is pivotal for achieving large-scale, reliable formal software verification. It is a common practice to employ human-annotated chain-of-thought and answers to induce the reasoning and coding capabilities of LLMs. Unfortunately, it becomes unacceptably all-consuming to provide such priors for supervising complex programming tasks. In this work, we systematically explore ways to reduce human annotations with the formal language, Dafny, as the main environment for our pilot study. Our pipeline mainly relies on introducing an automatic and scalable data curation pipeline, and careful RL designs integrated with feedback from the formal language verifier. We introduce DafnyComp, a benchmark of compositional formal programs with auto-formalized specifications for specification reasoning. Our supervised fine-tuning (SFT) stage enables even small models (e.g., 0.5B) to generate syntactically valid and verifiable Dafny code, surpassing proprietary models. RL with regularization further improves performance, achieving stronger generalization to out-of-domain tasks and outperforming all strong baselines on the challenging DafnyComp benchmark.

Recent advances highlight Reinforcement Learning with Verifiable Rewards (RLVR) as a promising approach to enhancing LLM capabilities. However, it remains unclear whether the current practice of RLVR truly expands models' reasoning boundary or mainly amplifies high-reward outputs that base models already know. We introduce empirical support dynamics, a framework that quantifies how RLVR preserves, expands, and shrinks access to verified solutions. Applied across models, reasoning tasks, and training checkpoints, it matches verified successes question by question to reveal changes that aggregate accuracy can conceal. Across the evaluated model-level benchmark aggregates, the shrinkage of observed support outweighs its expansion: more questions lose a verified success than gain one at the stated sampling budgets. These losses occur alongside gains in pass@1. Some RLVR models increase average token-level entropy while reducing final-answer entropy: their token choices become more uncertain on average even as their final answers become more concentrated. Together, these findings reveal an ``invisible leash'' in current RLVR practice: it can make existing successes more reliable while losing access to other verified solutions. Empirical support dynamics make this imbalance measurable and motivate exploration strategies that reach underrepresented solutions while retaining prior successes.

Next-word prediction (NWP) trains language models against a single observed continuation, even though many contexts admit multiple plausible next words. Recent RL-based next-word reasoning methods make this tension explicit: they reward a model for producing a rationale that supports one context-conditioned continuation, which can turn a pre-existing preference into a confident, self-justifying trajectory. We introduce BOW, an RL framework that instead trains models to produce self-contained, neutral, and comprehensive descriptions of the plausible next-word space. The policy generates a next-word reasoning trajectory from the full context, but a frozen scorer computes the core reward from that trajectory alone, without separately receiving the context. BOW-Reg adds a lightweight breadth regularizer to this core reward to discourage premature collapse. On two model backbones, BOW remains competitive with the original models and often outperforms trained baselines across ten general reasoning benchmarks. On benchmarks testing ambiguous references and word meanings, BOW-Reg achieves the highest SharedRef correctness and the lowest HoWN-Simple single-sense collapse on both backbones. Human evaluation further shows that BOW-Reg produces broader next-word reasoning trajectories, while direct next-word-prediction evaluation shows that these trajectories remain predictive.

Natural Language Inference (NLI) is a central task in natural language understanding with applications in fact-checking, question answering, and information retrieval. Despite its importance, current NLI systems heavily rely on supervised learning with datasets that often contain annotation artifacts and biases, limiting generalization and real-world applicability. In this work, we apply a reinforcement learning-based approach using Group Relative Policy Optimization (GRPO) for Chain-of-Thought (CoT) learning in NLI, eliminating the need for human-labeled rationales and enabling this type of training on challenging datasets such as ANLI. We fine-tune 7B, 14B, and 32B language models using parameter-efficient techniques (LoRA and QLoRA), demonstrating strong performance across standard and adversarial NLI benchmarks. At the 32B scale, GRPO-trained models generalize better than other supervised baselines in adversarial sets. With AWQ quantization, the 32B model fits within 22GB of CUDA memory. This work provides a scalable and practical framework for building robust NLI systems without sacrificing inference quality.

Large Language Models (LLMs) can exhibit factual inaccuracies during long-chain reasoning or under misleading prompts, limiting their reliability in historical analysis. We propose Kongzi, a language model for historical understanding and reasoning that combines continued pre-training on historical corpora, supervised fine-tuning, and fact-enhanced reinforcement learning. Its factuality reward compares entities and relations extracted from generated responses with historical references through knowledge-graph similarity. We evaluate 0.5B and 7B models on historical question answering and counterfactual reasoning, using reported ratings from three LLM judges, and on CMMLU-Chinese-History. Kongzi-7B receives the highest historical-accuracy ratings among the compared models under all three judges. Reinforcement learning improves the mean answer scores of both model sizes and their CMMLU-Chinese-History accuracy, although Kongzi-7B remains below the strongest reported baselines on that benchmark. These descriptive results support further study of factuality rewards for historical language models.

Reinforcement finetuning (RFT) has shown great potential for enhancing the mathematical reasoning capabilities of large language models (LLMs), but it is often sample- and compute-inefficient, requiring extensive training. In this work, we introduce AdaRFT (Adaptive Curriculum Reinforcement Finetuning), a method that significantly improves the efficiency of RFT through adaptive curriculum learning. AdaRFT dynamically adjusts the difficulty of training problems based on the model's recent reward signals, ensuring that the model consistently trains on tasks that are challenging but solvable. This adaptive sampling strategy accelerates learning by maintaining an optimal difficulty range, avoiding wasted computation on problems that are too easy or too hard. AdaRFT requires only a lightweight extension to standard RFT algorithms like Proximal Policy Optimization (PPO), without modifying the reward function or model architecture. Experiments on competition-level math datasets demonstrate that AdaRFT improves convergence efficiency and reasoning performance. Given problem-level difficulty annotations, AdaRFT reduces RFT training time by up to 2 times across data distributions and model scales, offering a more scalable and effective RFT framework.

Natural Language to SQL (NL2SQL) aims to translate natural language queries into executable SQL statements, offering non-expert users intuitive access to databases. While recent approaches leveraging large-scale private LLMs such as GPT-4 have achieved state-of-the-art results, they face two critical challenges: the lack of openness and reproducibility, and the prohibitive computational cost of test-time scaling. To address these issues, we explore improving the model-level performance of small-scale public LLMs in NL2SQL under resource-constrained settings. Our exploratory experiments reveal the potential of task decomposition for enhancing NL2SQL performance, but also highlight the difficulty of enabling LLMs to decompose queries effectively. Motivated by these findings, we propose LearNAT, a novel framework designed to enhance decomposition capabilities of LLM. LearNAT introduces (1) a Decomposition Synthesis Procedure, which leverages AST-guided search with pruning strategies to generate verifiable and efficient decompositions, and (2) Margin-Aware Reinforcement Learning, which provides fine-grained preference optimization for multi-step reasoning beyond standard DPO. Extensive experiments on benchmark datasets demonstrate that LearNAT significantly improves the performance of small-scale LLMs, achieving results comparable to GPT-4 with only a 7B parameter model. These results validate the effectiveness of verifiable decomposition and fine-grained preference learning in advancing NL2SQL towards openness, transparency, and efficiency. Our code is publicly available at https://github.com/MrBlankness/LearNAT.

The reasoning abilities of large language models (LLMs) have improved with chain-of-thought (CoT) prompting, allowing models to solve complex tasks stepwise. However, training CoT capabilities requires detailed reasoning data, which is often scarce. The self-taught reasoner (STaR) framework addresses this by using reinforcement learning to automatically generate reasoning steps, reducing reliance on human-labeled data. Although STaR and its variants have demonstrated empirical success, a theoretical foundation explaining these improvements is lacking. This work provides a theoretical framework for understanding the effectiveness of reinforcement learning on CoT reasoning and STaR. Our contributions are: (1) criteria for the quality of pre-trained models necessary to initiate effective reasoning improvement; (2) an analysis of policy improvement, showing why LLM reasoning improves iteratively with STaR; (3) conditions for convergence to an optimal reasoning policy; and (4) an examination of STaR's robustness, explaining how it can improve reasoning even when incorporating occasional incorrect steps. We also run RL-STaR on GPT-2, Qwen2.5-0.5B and Phi-3-mini, and the measured return curves follow the ones the analysis predicts. This framework bridges empirical findings with theoretical insights, advancing reinforcement learning approaches for reasoning in LLMs.

Autoregressive large language models (LLMs) compress knowledge from their training data through next-token conditional distributions. This limits tractable querying of this knowledge to start-to-end autoregressive sampling. However, many tasks of interest -- including sequence continuation, infilling, and other forms of constrained generation -- involve sampling from intractable posterior distributions. We address this limitation by using amortized Bayesian inference to sample from these intractable posteriors. Such amortization is algorithmically achieved by fine-tuning LLMs via diversity-seeking reinforcement learning algorithms: generative flow networks (GFlowNets). We empirically demonstrate that this distribution-matching paradigm of LLM fine-tuning can serve as an effective alternative to maximum-likelihood training and reward-maximizing policy optimization. As an important application, we interpret chain-of-thought reasoning as a latent variable modeling problem and demonstrate that our approach enables data-efficient adaptation of LLMs to tasks that require multi-step rationalization and tool use.