Large Language Model Reinforcement Learning

Follow this topic Ask about this topic

Momentum

17 papers in the last four weeks, up 42% on the four weeks before. 0.2% of all new papers.

Jul 6Week of Sep 21

Latest papers 194

Jul 5, 2026cs.LG

Regime-Conditional Stabilisation of LLM-Augmented Cooperative Multi-Agent Reinforcement Learning

Large Language Models (LLMs) offer a natural interface for translating human objectives into reward signals for cooperative multi-agent reinforcement learning (MARL), yet the training-time dynamics of this integration remain poorly understood. We show that dynamically updating LLM-generated reward weights during off-policy MARL violates the stationarity assumption of Potential-Based Reward Shaping (PBRS) and contaminates the experience replay buffer, whose stored transitions carry reward labels computed under stale shaping weights. We characterise the result as a regime-dependent failure whose severity depends on how competent the unshaped baseline already is. To control it we propose two stabilisation strategies: a Phase-Based Freeze Schedule that enforces strict stationarity within training phases, and Exponential Moving Average (EMA) smoothing that bounds per-episode weight drift. We evaluate across three cooperative environments and five random seeds with QMIX, complemented by an exploratory VDN extension, yielding a three-regime taxonomy. In the augmentative regime (Simple Spread), where the baseline is functional (74.4 %), EMA significantly improves success to 86.7 % (

+12.3

pp,

p<0.01

) while naive dynamic updates collapse it to 15.2 %. In the essential regime (Level-Based Foraging), where the baseline is broken (0.1 %), any shaping unlocks the task (95.9 % under EMA). In the supplementary regime (SMAC 3m), where the baseline is near-saturated (98.8 %), stabilised shaping preserves performance (99.9 %) while unstabilised shaping adds variance without gain. These findings establish reward-signal stationarity as a necessary design constraint and indicate that regime placement is a practical predictor of whether dynamic LLM shaping helps or harms.

Faid Keddouri, Sohaib Houhou, Aissa Boulmerka +1

National School of Artificial Intelligence (ENSIA), Sidi Abdellah Campus, Algiers, Algeria · Cosys-Grettia, Univ Gustave Eiffel, F-77454 Marne-la-Vallee, France.

Potential-Based Reward Shaping Multi-Agent Reinforcement Learning Large Language Model Reinforcement Learning Instability Stationarity Training Time

Jul 5, 2026cs.LG

On the effectiveness of reward functions in reinforcement learning for confidence calibration of large language models

In this paper, we consider the setting where large language models (LLMs) are trained using reinforcement learning (RL) to simultaneously improve reasoning accuracy and verbalize its confidence. Our reward scheme uses two functions for rewarding confidence verbalized by the LLM: one when the LLM is correct and a different one when the LLM is incorrect. With a poorly designed reward scheme, the LLM may be incentivized to answer incorrectly so that it can be confident that its answer is indeed incorrect, a phenomenon that we call confidence reward hacking. We propose the concept of non-hackable confidence reward schemes and define a spectrum of such reward schemes for RL confidence calibration training in LLMs. We demonstrate that selective confidence reward hacking can occur in practical datasets with reward schemes that are not designed to be non-hackable. We also demonstrate that the reward scheme with the best calibration to accuracy tradeoff depends on the dataset and the application, and propose using the reward scheme as a hyperparameter to optimize the tradeoffs in accordance to what is important for the application. The code of our experiments is available in https://anonymous.4open.science/r/rl-confidence-calibration-9ED4/README.md.

Chee Heng Tan, Zhuoyi Lin, Mehul Motani +1

School of Computing National University of Singapore · Institute of Advanced Intelligence and Computing Agency for Science, Technology and Research · Department of Electrical & Computer Engineering National University of Singapore

Large Language Model Reinforcement Learning Large Language Model Reliability Confidence Estimation Confidence Calibration Reward Functions

Jul 3, 2026cs.LG

Aligning Language Models with Selective Prediction

Large language models (LLMs) are increasingly deployed as critical decision-making components in high-stakes real-world AI systems, rendering LLM reliability a foremost practical concern. In this paper, we focus on enhancing LLM reliability through selective prediction (SP), a strategy that allows an LLM to only predict for inputs where it is likely to be correct (i.e., coverage) and hence reduce the error rate (i.e., risk) on that portion of inputs -- flagging the remaining inputs for future human discretion. In other words, SP improves LLM reliability by balancing the risk-coverage trade-off and enabling seamless human-AI collaboration. To integrate SP into LLMs, we focus on the LLM post-training alignment stage and propose to align LLMs with SP performance metrics, in contrast with existing LLM alignment methods that focus primarily on correctness or calibration metrics. Specifically, we propose a novel alignment framework, Reinforcement Learning for Selection Reward (RLSR), which targets the area under the risk-coverage curve (AURC) -- a popular SP performance metric -- as its alignment objective. RLSR achieves substantially better risk-coverage trade-off compared to multiple alignment baselines on both in-domain and out-of-domain tasks.

Gaoxiang Luo, Yifan Wu, Sinian Zhang +2

Department of Computer Science and Engineering · 2Bioinformatics and Computational Biology Program · 3Division of Biostatistics and Health Data Science

Large Language Model Alignment Large Language Model Reliability Large Language Model Reinforcement Learning Post-Training

Jul 3, 2026cs.AI

Oyster-II: Reinforcement Learning for Constructive Safety Alignment in Large Language Models

Large language models (LLMs) have demonstrated remarkable capabilities across diverse applications, yet ensuring their simultaneous safety, helpfulness, and trustworthiness remains a persistent challenge. Conventional refusal-oriented alignment strategies mitigate harmful content generation but systematically fail to serve legitimate user needs, often withholding information that could safely and constructively address the underlying intent of sensitive queries. Building upon the constructive safety paradigm pioneered by Oyster-I, which moves beyond blanket refusal toward thoughtful, response-oriented safety alignment, we identify two critical limitations of its Supervised Fine-Tuning (SFT)-based scheme: insufficient safety generalization to out-of-distribution scenarios and a phenomenon we term safety chain-of-thought (CoT) over-generalization, wherein safety-oriented reasoning patterns are excessively applied to benign queries, degrading helpfulness and user experience. To address these limitations, we propose Oyster-II, a reinforcement learning (RL)-based constructive safety alignment framework that adopts a Zero-RL paradigm combined with a multi-stage reinforcement learning strategy.Evaluated across extensive benchmarks, Oyster-II comprehensively surpasses both Qwen3-14B and its predecessor Oyster-I on safety dimensions, achieving cross-scale performance comparable to Qwen3-Max and Qwen3.5-397B.

Jiyang Guan, Yong Xie, Jun Chen +6

Alibaba AAIG

Safety Alignment Large Language Model Alignment Large Language Model Reinforcement Learning Offline Reinforcement Learning Fish

Jul 2, 2026cs.AI

Scaling with Confidence: Calibrating Confidence of LLMs for Adaptive Test Time Scaling

Training large language models (LLMs) with reinforcement learning (RL) has significantly advanced their performance on reasoning and question-answering tasks. However, prevailing RL reward designs typically prioritize response correctness, neglecting to incentivize models to express their confidence accurately. This leads to a critical problem: performance gains are often accompanied by poor calibration between confidence and accuracy, misleading models to overconfidently hallucinate when uncertain. To address this limitation, we propose

\textbf{C}

orrectness and

\textbf{C}

onfidence

\textbf{C}

alibration

\textbf{R}

einforcement

\textbf{L}

earning (

\textbf{C3RL}

), a novel RL algorithm integrating correctness, calibration and dataset-informed reference accuracy rewards together. Comprehensive evaluation across 8 text and multimodal datasets demonstrates that C3RL enhances calibration without sacrificing accuracy, outperforming the current state-of-the-art method in both performance and calibration metrics. Utilizing the well-calibrated verbalized confidence from C3RL, we further introduce

\textbf{C}

onfidence-based

\textbf{A}

daptive Test Time

\textbf{S}

caling (

\textbf{CAS}

), an adjustable inference-time strategy that allocates computational resources based on response confidence. Experiments show that CAS surpasses majority voting on both in-domain and out-of-domain datasets while reducing the inference budget by up to 12.33 times. We believe the synergy of C3RL and CAS paves the way for deploying more reliable and resource-efficient LLMs. The code, data and models will be released.

Xuqing Yang, Yi Yuan, Shanzhe Lei +1

Shanghai Jiao Tong University · Southeast University · Shanghai AI Laboratory

Confidence Estimation Large Language Model Reliability Large Language Model Reinforcement Learning Scaling Responses

Jun 30, 2026cs.CL

SLIM-RL: Risk-Budgeted Random-Masking RL for Diffusion LLMs Without Trajectory Slicing

Reinforcement learning for diffusion large language models (dLLMs) has largely moved to trajectory-aware methods. The current state of the art, TraceRL, holds that random masking is mismatched with the model's inference trajectory, and it reconstructs that trajectory during training by slicing each rollout into up to K/s trajectory-aligned training samples, a cost that grows with the block size K. We show that this mismatch can be mitigated without reconstructing the trajectory. Our method, SLIM-RL, bounds the commit risk of each rollout step with a tau-budget decoder, reducing aggregate commit risk in the training data. During optimization, SLIM-RL trains on these risk-controlled rollouts with a trace-free random-masking objective that adapts variance-reduction tools, combining sequence-level importance sampling, deterministic quadrature over masking levels under a mean-preserving, monotonically decreasing per-block mask schedule that we introduce. On SDAR-4B, SLIM-RL matches TraceRL's best MATH500 accuracy on only 0.46x its training samples at block size 16, improving over TraceRL by 6.32% on MATH500 and 11.05% on GSM8K under matched dynamic sampling. At block size 4, the 4B SLIM-RL surpasses the larger LLaDA-8B and Dream-7B dLLMs on math, exceeding LLaDA-8B by 10.76% on MATH500 while staying below the autoregressive Qwen2.5-7B. On code, it improves over TraceRL by 4.20% on MBPP and 3.65% on HumanEval. The tau-budget decoder transfers training-free across LLaDA, Dream, and SDAR. The source code is available at https://github.com/laolaorkkkkk/SLIM-RL .

Ruikang Zhao, Zhenting Wang, Han Gao +1

Technical University of Denmark · MBZUAI Institute of Foundation Models · Iowa State University +1

Diffusion Language Models Large Language Model Reinforcement Learning Diffusion-Based Reinforcement Learning Methods Token Budget Allocation Traces

Jun 30, 2026cs.LG

Addressing Over-Refusal in LLMs with Competing Rewards

Safety training on language models often induces over-refusal: improved safety on harmful prompts at the cost of increased refusal on harmless ones. Though this trade-off can be mitigated by training models with reinforcement learning (RL) to reason before answering, it does not remove the underlying problem that reasoning can often be a "rubber stamp" for a predetermined response. In this paper, we address the safety-refusal trade-off by rethinking how models are trained to reason about safety. Our key insight is that unsafe reasoning can itself serve as a useful exploratory signal. Rather than preemptively blocking harmful thoughts, we encourage the model to sufficiently explore unsafe reasoning but produce a safe response. The harmful exploration improves the model's ability to distinguish harmful from harmless prompts by resolving ambiguity, allowing it to remain safe while complying only when appropriate. We cast this as an adversarial optimization problem in which a reasoning player explores strategies for producing an unsafe response and an answer player ensures that the final output is safe. We train a single model with dense rewards to play both roles within one chain-of-thought, across different segments. To achieve this, we find that process rewards are crucial for stable optimization of competing objectives. Our resulting model SEAR deliberately engages in harmful reasoning as exploration while reliably flipping back to a safe answer. We demonstrate that this behavior helps mitigate over-refusal and defend against attacks that directly manipulate the reasoning to be harmful.

Taeyoun Kim, Aviral Kumar

Carnegie Mellon University · SynthLabs

Large Language Model Safety LLM Reasoning Strategies Refusals Large Language Model Reinforcement Learning Reasoning Skills Adversarial Training

Jun 29, 2026cs.LG

DRIFT: Difficulty Routing Self-DIstillation with Rhythm-Gated Exploration and Success BuFfer Training

Enabling large language models to achieve stable self-improvement without external expert supervision remains a central challenge in complex reasoning tasks. Existing self-distillation and reinforcement learning methods lack explicit mechanisms for tracking problem-level learning progress and adapting optimization strategies accordingly. Consequently, training may over-optimize easy problems, receive weak supervision from hard problems, and fail to sufficiently explore borderline cases. To resolve these issues, we propose DRIFT, an online self-evolution policy optimization framework for large language models. DRIFT regulates the model's self-improvement process through the joint use of Difficulty Routing and Rhythm Gating. The former identifies the model's learning state at the problem level and dynamically allocates self-distillation and reinforcement learning signals, while the latter refines policy updates at the token level, concentrating exploration on critical reasoning positions. By further incorporating a success buffer and a two-stage curriculum learning strategy, DRIFT preserves high-quality historical experience while progressively guiding the model from reliable behavior acquisition toward stable policy evolution. Evaluated across five benchmarks and three model scales, DRIFT surpasses the peak performance of both GRPO and SDPO across all evaluated metrics. On the average score over the five benchmarks, DRIFT achieves 79.5

\%

, outperforming GRPO by 9.5

\%

and SDPO by 7.5

\%

, establishing a new state-of-the-art result. Notably, on ToolUse, DRIFT reaches an accuracy of 79.2

\%

, improving over GRPO by 13.5

\%

and SDPO by 10.7

\%

, setting a new state-of-the-art and substantially outperforming all concurrent methods.

Haisen Luo, Yiwei Liu, Haoning Wang +13

3Beike Language and Intelligence · 2ENS Paris-Saclay · 1Tsinghua University

Unsupervised On-Policy Self-Distillation Large Language Model Reinforcement Learning On-Policy Self-Evolution Recursive Self-Improvement

Jun 29, 2026cs.CL

ARKD: Adaptive Reinforcement Learning-Guided Bidirectional KL Divergence Distillation for Text Generation

Knowledge distillation (KD) is a key technique for compressing Large Language Models (LLMs), yet methods relying on a single KL objective often fail to balance primary distribution fitting with long-tail probability modeling, limiting both generation quality and generalization. To address this, we analyze the complementary roles of forward and reverse KL divergence (FKL/RKL) in distribution alignment from theoretical and empirical perspectives. We then propose a reinforcement-learning-based adaptive KL-weighted distillation framework, in which a policy network dynamically assigns weights to FKL and RKL based on teacher-student distributional characteristics, guided by immediate reward signals to achieve dual alignment on principal and long-tail modes. Extensive experiments demonstrate consistent improvements across Rouge-L and BertScore metrics, surpassing greedy heuristics by 0.4-0.6 points and outperforming other baseline methods on diverse benchmarks.

Zilong Liu, Xuewen Zhang, Jinrui Xing +3

Li Auto, Beijing, China · School of Software and Microelectronics, Peking University, China

Probe-Logit Distillation Kullback-Leibler Divergence Large Language Model Reinforcement Learning Text Generation

Jun 28, 2026cs.LG

The Mirage of Optimizing Training Policies: Monotonic Inference Policies as the Real Objective for LLM Reinforcement Learning

Reinforcement learning (RL) has gained growing attention in large language model (LLM) post-training, yet RL training remains fragile and can suffer from instability or collapse. One vital cause is training-inference mismatch: LLM adopts separate inference and training engines for generation efficiency and training precision, which in practice exhibits inconsistent probabilities for the same trajectories on training and inference sides, even with synchronized model parameters. This naturally induces a special type of off-policyness ever existing and poisoning the training. Prior works have made various efforts in addressing the off-policyness to stabilize the training policies under the mismatch. In this paper, we point out the objective misalignment neglected by existing works that an effective update to the policy in the training engine not necessarily ensures the improvement of the inference policy, i.e., the one used in deployment. To this end, we propose a new policy optimization objective for LLM RL, named Monotonic Inference Policy Improvement (MIPI). Following this principle, we introduce Monotonic Inference Policy Update (MIPU), a two-step LLM RL framework that constructs sampler-referenced candidate updates and selectively accepts synchronized candidates using an inference-side gap proxy. Experiments conducted on two model scales under high mismatch show that MIPU improves average reasoning performance and training stability.

Jing Liang, Hongyao Tang, Yi Ma +9

1Tianjin · 2Alibaba

Large Language Model Reinforcement Learning Training-Inference Mismatch Boltzmann Policies

Jun 25, 2026cs.DC

RolloutPipe: Overlapping Pipelined Rollout and Training in Disaggregated On-Policy LLM Reinforcement Learning

Large language model (LLM) post-training for reasoning increasingly relies on reinforcement learning with verifiable rewards (RLVR), where models learn from ground-truth feedback on mathematical, logical, and scientific tasks. To enable flexible resource allocation and support heterogeneous training setups, modern RLVR systems adopt disaggregated architectures that decouple rollout generation and policy training across independent GPU pools. However, existing synchronous on-policy GRPO (Group Relative Policy Optimization) RLVR systems finish an entire rollout before starting training, leaving the trainer GPU pool idle while rollout is still ongoing. Asynchronous RL pipelines overlap the two stages, but at the cost of training on stale data. To address these challenges, we propose RolloutPipe, a post-training framework for disaggregated RLVR systems, which turns the fixed-weight rollout into a complete-group pipeline where trainable groups move to the trainer while later groups are still being generated. RolloutPipe achieves this through two techniques including complete-group pipelining (CGP) and frontier-group dispatch (FGD). CGP dispatches each trainable complete group to the trainer FIFO as soon as group materialization finishes, and FGD is an admission policy on the Rollout node that first admits requests for the frontier groups needed to form the next training batch, so that trainer-ready groups arrive earlier and more steadily. The design starts training before the rollout completes while maintaining on-policy correctness. Evaluated on Qwen3-1.7B across four reasoning and science benchmarks and twelve rollout settings, RolloutPipe shortens the rollout-to-train-end time by 30.7%-42.3%, and lowers the trainer waiting ratio by 37%-76% compared to Slime, a state-of-the-art rollout and training system.

Rongjian Chen, Jianmin Hu, Kejiang Ye +1

Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China · University of Chinese Academy of Sciences, Beijing, China · Southern University of Science and Technology, Shenzhen, China

Large Language Model Reinforcement Learning Rollout Reinforcement Learning With Verifiable Reward

Jun 25, 2026cs.LG

GEOALIGN: Geometric Rollout Curation for Robust LLM Reinforcement Learning

Online reinforcement learning is widely used to align large language models (LLMs) with reward signals, yet training can be unstable under noisy or misspecified rewards. We identify a failure mode we call directional inconsistency: within a batch, a small set of high-reward rollouts induces representation-space preference directions that sharply disagree with the batch majority, resulting in high-variance and destabilizing updates. We propose geoalign, a lightweight plug-in for rollout curation in iterative policy optimization. Geoalign (i) forms within-prompt preference pairs, (ii) learns an online projector on per-rollout hidden states to concentrate reward-ordered displacement directions, and (iii) detects directionally inconsistent rollouts via their angular deviation from a batch consensus prototype and rectifies them with within-prompt stable alternatives. Geoalign is forward-pass only and adds negligible overhead. Across dialogue alignment with a learned reward model and mathematical reasoning with binary verified rewards, Geoalign improves final performance and reduces training oscillation, outperforming PF-PPO, PAR, PODS, and Seed-GRPO. These results suggest latent directional consensus as an effective reliability signal for online LLM RL.

Ting Zhou, Zhenqing Ling, Yiyang Zhao +2

Sun Yat-Sen University, Guangzhou, China · Alibaba Group, Hangzhou, China · Guangdong Tobacco Guangzhou Co., Ltd., Guangzhou, China

Large Language Model Alignment Large Language Model Reinforcement Learning Offline Reinforcement Learning Latent Variable

Jun 24, 2026cs.CL

Why Multi-Step Tool-Use Reinforcement Learning Collapses and How Supervisory Signals Fix It

Tool use enables large language models (LLMs) to perform complex tasks, and recent agentic reinforcement learning (RL) methods show promise for enhancing model capabilities. However, RL alone often leads to instability or limited gains in tool-use tasks. In our experiments, some models exhibit catastrophic collapse, where performance abruptly drops and tool-invocation structures fail. The analysis reveals that these failures stem from unexpected probability spikes in specific control tokens, disrupting structured execution, yet the underlying tool-use capability remains intact, merely obscured by specific formats. To address this, we systematically investigate a diverse set of supervisory signals, including off-policy supervision, hint-based guidance, erroneous example supervision, and others, applied under both synchronous and interleaved training schemes. We find that interleaving supervised fine-tuning (SFT) with RL substantially improves stability, but exhibits degraded performance under format and content out-of-distribution (OOD) evaluation. We also analyze the impact of learning rates and generalization across settings. These results highlight the importance of understanding RL failures and demonstrate how diverse supervisory signals can guide exploratory learning, enabling robust training of LLMs for complex, multi-step tool-use tasks. Our Code is available at https://github.com/hypasd-art/Tool-RL-Box.

Yupu Hao, Zhuoran Jin, Huanxuan Liao +2

1The Key Laboratory of Cognition and Decision Intelligence for Complex Systems, Institute of Automation, Chinese Academy of Sciences, Beijing, China · School of Artiﬁcial Intelligence, University of Chinese Academy of Sciences, Beijing, China

Large Language Model Reinforcement Learning Agentic Reinforcement Learning Tool Invocation Collapse

Jun 24, 2026cs.CL

OPERA: Aligning Open-Ended Reasoning via Objective Perplexity-based Reinforcement Learning

Reinforcement Learning (RL) has enabled LLMs to excel in objective reasoning tasks such as mathematics and code generation. However, applying RL to open-ended tasks, such as creative writing, remains challenging because LLM-as-a-judge reward models often exhibit stylistic biases and positional inconsistencies, leading to unstable supervision. To address this, we propose OPERA (Objective Perplexity-based Reflective Alignment), which replaces unreliable external judges with intrinsic rewards derived from perplexity dynamics. Specifically, we derive an intrinsic reward signal from perplexity dynamics, quantifying uncertainty reduction at critical reflective states. During the cold-start phase, we introduce a data synthesis method that leverages carefully designed guiding words to generate diverse reasoning traces, along with perplexity-prioritized rollouts that utilize internal log-probabilities to identify logically consistent reasoning branches. This pipeline yields a large-scale dataset comprising 20,000 high-quality reasoning trajectories. Empirical evaluations consistently demonstrate the scalability and efficacy of our approach in alignment for open-ended tasks. Implementing OPERA on Qwen3-8B establishes a new state-of-the-art among open-source models, achieving parity with or surpassing proprietary models like Gemini2.5 and MiniMax-M2.5 in some open-ended tasks. The code is available at https://github.com/pangpang-xuan/OPERA.

Wenxuan Jiang, Zining Fan, Zijian Zhang +6

The Hong Kong Polytechnic University · Peking University · Meituan Longcat Team +2

LLM Reasoning Strategies Large Language Model Reinforcement Learning Reasoning Trajectory Large Language Model Alignment Llm-As-A-Judge Value Alignment+1

Jun 24, 2026cs.LG

Learning with a Single Rollout via Monte Carlo Pass@k Critic

Estimating token-level advantages in reinforcement learning (RL) for language models remains challenging because scaling up episodic experience collection is expensive. The difficulty intensifies for baseline advantage estimation methods, where repeated sampling causes trajectories to diverge into substantially different reasoning prefixes. In this context, RL algorithms such as GRPO prove limited: an outcome reward is too sparse to be attributed to specific actions like intermediate steps, and comparisons across sampled traces are non-trivial because they are heterogeneous. To mitigate both the computational cost of repeated sampling and the difficulty of credit assignment, we study single-rollout proximal policy optimization (SR-PPO) featuring token-level credit assignment in RL for language models. Instead of estimating advantages by normalizing episodic returns within the candidate group, we train a calibrated token-level credit critic using Monte Carlo outcomes from one rollout per prompt. Specifically, we use the critic to predict the Pass@k success probability at the prompt prefix, which is derived from a Pass@1 attempt. This choice yields a more selective learning signal than Pass@1: it discounts easily solved prefixes while prioritizing hard ones whose success probability remains marginal. We show that as

k

increases, Pass@k converges to a reachability indicator, reflecting whether a prefix can lead to at least one successful continuation. In an explicit state graph, the limit (

k \rightarrow \infty

) can be computed in

O(|V|+|E|)

time, offering a promising surrogate for direct credit assignment without the need to sample contrastive traces. As an initial validation, SR-PPO exhibits stable learning dynamics, along with consistent gains in Pass@128 success rates on mathematical reasoning benchmarks such as HMMT26 and AIME24.

Fengdi Che, Yang Liu, Lei Yu +4

University of Alberta · BIGAI · University of Toronto +4

Large Language Model Reinforcement Learning Credit Assignment Step-Level Advantage Estimation Critic Monte Carlo

Jun 22, 2026cs.LG

EvoRubrics: Dynamic Rubrics as Rewards via Adversarial Co-Evolution for LLM Reinforcement Learning

Rubric-based rewards offer interpretable and fine-grained optimization signals for reinforcement learning in open-ended tasks where verifiable answers are unavailable. However, pre-constructed rubrics remain static throughout training, creating a fundamental mismatch with the evolving policy: fixed criteria gradually lose discriminative power as the model improves, leading to reward saturation and potential hacking. Recent dynamic rubric methods partially address this but rely on external frontier models or ground-truth answers, and update rubrics only at coarse granularity. We propose EvoRubrics, a co-evolutionary RL framework where a Policy LLM and a Rubric Generator jointly improve through adversarial interaction within each training step. As the policy improves under the rubric generator's guidance, the rubric generator adapts its criteria to remain discriminative and informative, enabling evaluation to track the policy in real time and naturally inducing an automatic curriculum. Experiments show that EvoRubrics consistently outperforms static and dynamic rubric baselines across benchmarks. The learned Rubric Generator further generalizes as a transferable reward model. Notably, even a fully self-supervised variant without any external supervision achieves meaningful gains, suggesting that co-evolution between generation and evaluation alone can provide sufficiently rich learning signals. Our code is publicly available at https://anonymous.4open.science/r/EvoRubrics-2155/.

Hongxin Ding, Baixiang Huang, Yue Fang +6

National Engineering Research Center of Software Engineering, Peking University, China · School of Computer Science, Peking University, Beijing, China · Key Laboratory of High Confidence Software Technologies, Ministry of Education +2

On-Policy Self-Evolution Large Language Model Reinforcement Learning Co-Evolution Adversarial Robustness

Jun 21, 2026cs.LG

Curriculum Reinforcement Learning Can Incentivize Reasoning Capacity in LLMs Beyond the Base Model

Reinforcement learning with verifiable rewards (RLVR) is widely viewed as a promising path toward continuously improving large language models. Recent works, however, suggest that mainstream RLVR often reallocates sampling probabilities among trajectories already present in the base model: it can improve sampling efficiency, reflected by higher pass@1 scores, but yields limited gains, and can even decrease pass@k scores when k is large, and therefore may fail to expand the base model's reasoning capacity boundary. In this paper, we present a boundary-aware Curriculum RL approach to move beyond the base model's reasoning capacity boundary. Our approach first uses pass@k sampling to locate the current reasoning capacity boundary, then applies targeted teacher guidance to examples near or beyond that boundary, and finally uses RL to consolidate the newly introduced reasoning patterns. Across Qwen, Llama, and DeepSeek base models, boundary-aware Curriculum RL improves both pass@1 scores and pass@256 scores, with pass@1 reflecting one-attempt performance and pass@256 serving as an empirical proxy for the reasoning capacity boundary. In our experiments, average pass@256 improves by 9.8 percentage points over the base models and by 10.3 percentage points over Vanilla RLVR. These results suggest that boundary-aware Curriculum RL can provide a scalable route for LLMs to continuously improve beyond the base model's empirical reasoning capacity boundary.

Pengxiang Cai, Tianchen Fang, Xiaohan Li +3

The Hong Kong University of Science and Technology (Guangzhou) · Nanyang Technological University · Zhejiang University

Curriculum Reinforcement Learning Large Language Model Reinforcement Learning LLM Reasoning Strategies Reinforcement Learning With Verifiable Reward Deepseek

Jun 21, 2026cs.CL

Learning at the Right Pace: Adaptive Data Scheduling Improves LLM Reinforcement Learning

Large Language Models (LLMs) achieve remarkable reasoning capabilities through reinforcement learning (RL) post-training. However, existing RL post-training commonly relies on uniform data sampling, which ignores the semantic structure of the training data and the changing capability of the training policy. To address these limitations, we propose Adaptive Data Scheduling (ADS), a dual-level data scheduling framework for pacing RL post-training that replaces uniform sampling with an adaptive distribution over semantic clusters and policy-boundary sample selection. At the cluster level, ADS organizes samples according to semantic patterns and maintains an adaptive inter-cluster distribution to solidify current training progress. At the sample level, ADS performs intra-cluster scheduling to continuously sample policy-boundary samples, which provides informative relative advantages. Experimental results across three LLMs and seven reasoning benchmarks demonstrate that ADS improves average accuracy by 5.2% over Group Relative Policy Optimization (GRPO). Notably, ADS consistently improves RL methods with different objective designs, highlighting its potential as a general data scheduling strategy for LLM RL post-training. The source code is available at: https://github.com/Richard-zrx/ADS.

Zicheng Xu, Ruixuan Zhang, Yu-Neng Chuang +7

Johns Hopkins University · Rice University · University of Haifa +2

Large Language Model Reinforcement Learning Large Language Model Training Reasoning Benchmark

Jun 20, 2026cs.LG

Modularized Reinforcement Learning on LLMs: From MDP Creation to Exploration and Learning

Reinforcement learning (RL) has become central to LLM post-training, yet the methods that dominate current pipelines, PPO and GRPO, represent only a narrow slice of what RL offers. Understanding why these methods prevail, and what alternatives exist, requires a principled examination of the design decisions that underlie any RL algorithm. This survey organizes that examination around three stages of algorithm construction. We begin with MDP creation: how the reward function, state space, action space, termination condition, and discount factor are, or could be, defined for LLM training. We then turn to exploration, covering temperature sampling, entropy regularization, intrinsic motivation, tree search, and curriculum learning. Finally, we address learning along four classical RL dimensions: model-free versus model-based, value-based versus policy-based versus actor-critic, on-policy versus off-policy, and credit assignment, including both Monte Carlo methods, which rely on full return estimates, and bootstrapping methods, which update estimates using other learned predictions. Mapping the LLM literature onto this taxonomy reveals a strikingly non-uniform distribution of research effort. Critic-free policy gradients and Monte Carlo credit assignment are densely populated, while value-based methods, off-policy actor-critic training, and bootstrapping-based credit assignment remain largely unexplored despite well-established counterparts in classical RL. These gaps represent concrete opportunities for transferring proven RL techniques to LLM training. By making these gaps explicit alongside the methods that have proven effective, this survey offers researchers in both RL and LLMs a shared framework for understanding current practice and identifying promising directions for future work.

Zhao Yang, Yuxuan Jiang, Ting-Chih Chen +18

VU Amsterdam, The Netherlands · University of Maryland, Baltimore County, USA · Leiden University, The Netherlands +4

Large Language Model Reinforcement Learning Offline Reinforcement Learning Large Language Model Training Policy Gradient Credit Assignment Markov Decision Processes+1

Jun 19, 2026cs.LG

Depth-Entropy Guided Sampling for Training-Free LLM Reasoning

Reinforcement learning (RL) has become the dominant paradigm for improving the reasoning capabilities of large language models, but it requires expensive training, curated data, and reward signals. Recent work shows that sampling from sharpened base-model distributions at test time recovers much of the RL gain, yet existing methods rely solely on output-layer likelihoods and ignore the transformer's internal forward-pass dynamics. We introduce Depth-Entropy Guided Sampling (DEGS), a training-free, test-time method that exploits layer-wise entropy collapse as an intrinsic quality signal. We observe that stronger reasoners -- including RL-posttrained variants -- exhibit a distinctive "late collapse": logit-lens decoded entropy stays elevated until deeper layers before converging. We define a per-sequence collapse depth

D(\mathbf{x})

and a joint objective

π(\mathbf{x}) \propto p(\mathbf{x})^α\exp(βD(\mathbf{x}))

that combines sequence likelihood with this depth-entropy structure, instantiated inside an MCMC power-sampling framework (DEGS-MCMC). Across three open-weight models and four reasoning benchmarks, this near-chance per-candidate signal compounds over the sampling trajectory into state-of-the-art training-free accuracy, with gains largest out of domain and on the harder splits -- exactly where likelihood alone falls short -- at single-digit-percent wall-clock overhead. DEGS narrowly trails an in-house GRPO reference on the math splits GRPO was trained for, yet surpasses it out of domain on GPQA for all three models, without any training, reward model, or labeled data.

Zibin Meng, Peng Xie, Kani Chen

1The Hong Kong University of Science and Technology

LLM Reasoning Strategies Recursive Reasoner Large Language Model Reinforcement Learning Diffusion Sampling

Jun 18, 2026cs.LG

Manifold Bandits: Bayesian Curriculum Learning over the Latent Geometry of Large Language Models

Reinforcement learning (RL) is a central approach for improving reasoning capabilities in large language models (LLMs), where training efficiency depends critically on how problems are sampled during optimization. Existing adaptive curriculum learning methods typically prioritize prompts of intermediate difficulty, treating problem selection as a standard bandit problem with independent arms and overlooking the structured, heterogeneous nature of the task space. In this work, we frame problem sampling as a manifold-structured bandit problem with endogenous non-stationarity: problems are related through the model's latent representation space, and sampling decisions can steer how learning signals evolve across that space. To operationalize this perspective, we introduce Bayesian Manifold Curriculum (BMC), a structure-aware framework that organizes problems into a hierarchical task tree and applies Bayesian learning to guide sampling. Empirically, we find that different sampling strategies induce non-trivial tradeoffs between productivity (learning signal), diversity (coverage of the task manifold), and utility (evaluation relevance). These results show that prioritizing difficulty alone is insufficient for strong downstream performance, highlighting the importance of incorporating structure and type-awareness into problem sampling.

Darrien McKenzie, Nicklas Hansen, Xiaolong Wang

University of California, San Diego

Curriculum Reinforcement Learning Large Language Model Reinforcement Learning Adaptive Sampling Large Language Model Training Curriculum Stochastic Multi-Armed Bandits

Jun 17, 2026cs.CL

Beyond Reward Engineering: A Data Recipe for Long-Context Reinforcement Learning

Long-context reasoning is an essential capability for large language models, particularly when they are deployed as autonomous agents that must reason over lengthy trajectories. Reinforcement learning (RL) has recently emerged as a dominant paradigm for improving this ability, yet existing work largely focuses on reward engineering while diverse training data remains scarce. We revisit this problem from a data-centric perspective and show that a simple yet effective data recipe alone, paired with a minimal outcome-based GRPO setup, suffices to substantially improve long-context reasoning. Our recipe targets three complementary task families -- retrieval, multi-evidence synthesis, and reasoning -- for which we construct and curate eight datasets totaling ~14K examples. Experiments on three models (Qwen3-4B/8B/30B-A3B) yield average gains of +7.2/+3.2/+6.4 points across seven long-context benchmarks, surpassing prior RL training sets. We further demonstrate that these gains transfer to agentic tasks, where continuing RL training on an agent-tuned model with our data recipe improves GAIA by +4.8 and BrowseComp by +7.0 points. We will release our datasets to facilitate future research.

Xiaoyue Xu, Sikui Zhang, Xiaorong Wang +2

OpenBMB · Tsinghua University

Large Language Model Reinforcement Learning Long-Context Reasoning LLM Reasoning Strategies Better Call Group Relative Policy Optimization Recipe

Jun 16, 2026cs.LG

From Reasoning Traces to Reusable Modules: Understanding Compositional Generalization in Language Model Reasoning

Post-training pipelines that combine supervised fine-tuning (SFT) with reinforcement learning (RL) have emerged as the key recipe for transforming large language models (LLMs) into robust reasoners. We argue that this combined success is driven by compositional generalization, which we formalize through a hierarchical latent selection model. In this framework, reasoning traces are generated by a cascade of discrete latent selection variables corresponding to reusable atomic modules, including both skills (local operations) and routing mechanisms (how intermediate information is selected, reused, and composed). Within this model, we theoretically show that SFT and RL play asymmetric, complementary roles: SFT supplies the raw module materials in compositional traces, and RL decomposes those traces to identify the latent atomic modules and enable compositional generalization. We design controlled experiments to validate this theory. Our results demonstrate that RL can extract atomic modules from compound traces supplied by SFT and recombine them to solve new configurations. Moreover, we find that training on compound traces yields stronger generalization than training on isolated atomic modules. Finally, we investigate the relationship between SFT and RL data and identify an effective protocol in which SFT ensures coverage of all atomic modules through compositional traces, while RL focuses on novel compositions outside the SFT support to drive exploration.

Lingjing Kong, Xin Liu, Guangyi Chen +9

Carnegie Mellon University · Mohamed bin Zayed University of Artificial Intelligence · Institute of Foundation Models +1

LLM Reasoning Strategies Compositional Generalization Large Language Model Reinforcement Learning Reasoning Traces Recursive Reasoner Reuse

Jun 16, 2026cs.AI

Shattering the Autoregressive Curse: Dynamic Epistemic Entropy Orchestrated Erasable Reinforcement Learning for LLMs

Although reinforcement learning (RL) has expanded the cognitive boundaries of large language models (LLMs), it often remains vulnerable to the autoregressive curse in long-horizon logical reasoning: small epistemic perturbations introduced early in generation can propagate irreversibly along the Markov decision process flow, triggering cascading failures that drive the reasoning trajectory toward collapse. To overcome this autoregressive cascade, in which a single early mistake can compromise all subsequent reasoning steps, we propose dynamic epistemic entropy orchestrated erasable reinforcement learning (

\text{E}^3\text{RL}

).

\text{E}^3\text{RL}

eliminates reliance on external signals by grounding the model's endogenous local autoregressive cross-entropy as an intrinsic coordinate of epistemic uncertainty. By introducing segment-level adaptive dynamic thresholds and advantage allocation,

\text{E}^3\text{RL}

enables the model to precisely excise localized logical defects while reusing historical key-value (KV) cache streams, thereby endowing the reasoning process with a self-healing capability. We train

\text{E}^3\text{RL}

on the DeepMath-103k dataset. Experimental results show that

\text{E}^3\text{RL}

reshapes the exploration efficiency of long-sequence reasoning and improves sample efficiency while maintaining linear memory overhead. On mathematical reasoning benchmarks such as AIME,

\text{E}^3\text{RL}

achieves substantial performance gains, with the 4B and 8B parameter models surpassing previous state-of-the-art (SOTA) results by 5.349% and 6.514%, respectively. These findings suggest that

\text{E}^3\text{RL}

shatters the autoregressive curse in long-sequence reasoning and establishes a theoretical and systems-level foundation for the next generation of self-healing artificial general intelligence (AGI).

Ziliang Wang, Kang An, Faqiang Qian +5

1SenseTime · 2Shanghai Jiao Tong University

Large Language Model Reinforcement Learning LLM Reasoning Strategies Offline Reinforcement Learning Autoregressive Model Sequence Modeling

Jun 16, 2026cs.CL

From Trainee to Trainer: LLM-Designed Training Environment for RL with Multi-Agent Reasoning

Reinforcement learning pipelines for Large Language Model (LLM) training often rely on manually redesigned environments between stages, requiring practitioners to heuristically infer which configuration will best improve the current policy. To automate this process, we propose the LLM-as-Environment-Engineer framework in which the current policy model analyzes failure trajectories together with contextual information and proposes modifications to the next-stage training environment configuration. We also introduce MAPF-FrozenLake, a controllable testbed whose generator exposes multi-dimensional environment configurations, making it suitable for studying and benchmarking environment redesign. On this testbed, we condition the environment engineer on structured summaries of policy behavior, failure cases, and environment statistics, from which it produces the configuration for the next training stage. With Qwen3-4B as the backbone, our framework achieves the strongest aggregate performance on our benchmarks, outperforming larger proprietary LLMs (e.g., GPT, Gemini) and fixed-environment training baselines. We further analyze which forms of context are most effective, finding that successful environment updates rely on failure evidence and preserve configurations that already work. Interestingly, the current RL checkpoint serves as a better environment engineer than the original base model, suggesting that policy learning improves the model's ability to diagnose its remaining weaknesses.

Chao Chen, Chengzu Li, Zhiwei Li +2

1LARK, HKUST (GZ) · University of Cambridge · 3HKUST

Large Language Model Reinforcement Learning

Jun 15, 2026cs.CL

The Value Axis: Language Models Encode Whether They're on the Right Track

We investigate whether language models internally track the value of their current trajectory, defined as the likelihood that their ongoing strategy will achieve their goals. Using synthetic, in-context reinforcement learning data, we construct a "value" axis for Qwen3-8B. We find that activations along this axis distinguish between high vs. low verbalized confidence, rollouts without and with backtracking, and correct vs. corrupted code. Steering towards high value causally suppresses self-correction and reduces explanatory verbosity, while steering towards low value induces backtracking and exploration. We demonstrate that direct preference optimization (DPO) can increase the internal value of rewarded behaviors (e.g. use a certain word), causing the model to act more confidently after exhibiting them. Finally, we apply the value axis to study in-the-wild settings. For example, we find that Qwen assigns low value to politically sensitive chat queries after post-training and that supervised fine-tuning increases internal confidence within the training domain. Our results suggest that language models linearly encode an estimate of expected goal success that modulates their confidence in pursuing a direction.

Nick Jiang, Isaac Kauvar, Jack Lindsey

1Stanford University · 2Anthropic

Large Language Model Reinforcement Learning Value

Jun 15, 2026cs.LG

ExpRL: Exploratory RL for LLM Mid-Training

Sparse reward reinforcement learning (RL) has become a standard tool for improving LLM reasoning, but its success depends critically on the coverage present in the base model. In practice, models are often primed for RL through \emph{mid-training} on curated reasoning traces that teach useful primitive skills such as decomposition, verification, or self-correction. Although effective, this strategy requires manually specifying what the model should learn, and it remains unclear whether such primitive coverage is enough for much harder problems, which require combining these skills into broader solution strategies. We study a more automated approach: \emph{RL-based mid-training} using large corpora of human-written question-answer data. Rather than treating reference solutions as targets to imitate, our method, ExpRL, uses them as \emph{reward scaffolds}: references are hidden from the policy and used only to construct problem-specific grading rubrics for judging on-policy reasoning traces. The policy samples from the original problem prompt, while an LLM judge compares the sampled reasoning trace against the reference solution and assigns outcome-level or process-level dense rewards. This lets ExpRL reinforce partial progress, useful intermediate reductions, and productive reasoning behaviors that sparse final-answer rewards often fail to upweight. On challenging math reasoning tasks, ExpRL yields stronger RL priming than SFT, sparse-reward GRPO, and self-distillation, and provides a better initialization for subsequent sparse-reward RL. Additional mixed-domain experiments further suggest that ExpRL can extend beyond the original math-only setting.

Violet Xiang, Amrith Setlur, Chase Blagden +2

1Stanford University · 2Carnegie Mellon University · 3OpenAI +1

Large Language Model Reinforcement Learning LLM Reasoning Strategies Expression

Jun 14, 2026cs.CV

OmniOPSD: Rationale-Privileged On-Policy Self-Distillation for Affective Computing

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

84.19

, and ablations further support the value of rationale-privileged teacher guidance.

Zebang Cheng, Shuimu Chen, Boxue Yang +7

Shenzhen University · Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ) · Tsinghua University +6

Unsupervised On-Policy Self-Distillation Large Language Model Reinforcement Learning Omni

Jun 14, 2026cs.LG

Reinforcement Learning for LLM-based Event Forecasting

We use Group Relative Policy Optimization (GRPO), a recently devised sample and memory efficient reinforcement learning method, to finetune pretrained LLMs in the range of 1.5B to 14B parameters equipped with the ability to get current information through the use of a Wikipedia revisions tool, or news summaries, to forecast real events beyond the knowledge cutoff of the LLM, as well as problems made to simulate different aspects of the dynamics of that training. We use the results of these experiments to comment on the scaling capability of LLMs for forecasting, as well as classify how judgmental forecasting fits into the verifiable/unverifiable domain taxonomy, considering the impact of the inherent aleatoric uncertainty when forecasting future events (e.g. the roll of a die). As a result of the GRPO training, we manage to bring a 1.5B parameter transformer (Qwen 2.5 1.5B) to forecasting performance superior to Claude Sonnet 3.5 over the same dataset as measured by cross entropy from the market agreed probabilities. We also discuss various dead ends on the path to this result.

Amit Arnold Levy

University of Oxford

Large Language Model Reinforcement Learning Data-Driven Forecasting

Jun 13, 2026cs.CL

Replay What Matters: Off-Policy Replay for Efficient LLM Reinforcement Unlearning

LLM unlearning has emerged as a cost-effective alternative to full retraining for removing hazardous knowledge from pretrained models while preserving general utility. Recent RL-based methods such as RULE reformulate unlearning as learning a refusal behavior, but their on-policy optimization repeatedly samples from the same forget and retain/boundary prompts throughout training. We identify a critical inefficiency in this process: easy cases quickly converge and provide little useful gradient signal, while hard cases near the forget/retain boundary continue to produce low-reward rollouts that are discarded after a single use. To address this issue, we propose ReRULE, an off-policy replay enhancement for reinforcement unlearning. ReRULE stores low-reward hard-case rollout groups in a replay buffer during early GRPO training and reuses them in later stages through importance-sampled off-policy updates, redirecting computation toward boundary cases that still require learning. Theoretically, we show that ReRULE yields a tighter hard-case convergence bound than pure on-policy RULE. Empirically, ReRULE improves MUSE-Books Retain Quality from 46.3 to 56.2 while adding only 5--11% training time across benchmarks. Its limited improvement on the simpler TOFU setting further supports the intended conditional behavior: replay is most beneficial when the hard/easy disparity is pronounced.

Zirui Pang, Chenlong Zhang, Haosheng Tan +3

The Hong Kong University of Science and Technology (Guangzhou) · Institute of Automation, Chinese Academy of Sciences · University of Glasgow

Large Language Model Unlearning Large Language Model Reinforcement Learning Replay Convergence

Jun 11, 2026cs.AI

ReSum: Synergizing LLM Reasoning and Summarization with Reinforcement Learning

Reinforcement Learning with Verifiable Rewards (RLVR) is a central technique for improving long-horizon reasoning in Large Language Models (LLMs). However, existing RLVR methods often encourage unnecessarily long reasoning rollouts, which can degrade reasoning coherence and exhaust the available context budget. Existing approaches to long-context organization often depend on external mechanisms to organize rollouts, rather than enabling the model to manage its own reasoning trajectory. To address this limitation, we propose ReSum, a novel RLVR framework that enables LLMs to compress and organize their reasoning trajectories through self-summarization. Our pilot studies show that self-summarization stabilizes generation by lowering token-level entropy, and that introducing a ``summarization'' phrase can substantially mitigate errors propagated from an incorrect rollout prefix. Motivated by these findings, ReSum adopts a summarization-aware adaptive rollout mechanism that contrastively evaluates whether self-summarization benefits the ongoing reasoning process. Specifically, when the model spontaneously triggers self-summarization, ReSum masks the summarization phrase to create a contrastive branch; for non-summarization positions, it instead randomly injects the phrase to create a matched branch. We further design a summarization-aware advantage to enable finer-grained comparison between contrastive rollout trajectories. Extensive experiments show that ReSum improves performance at an average of 4% while reducing rollout length by 18.6%.

Xucong Wang, Ziyu Ma, Yong Wang +5

University of Science and Technology of China · 2AMAP, Alibaba Group

Large Language Model Reinforcement Learning LLM Reasoning Strategies Reinforcement Learning With Verifiable Reward Summarization

Jun 10, 2026cs.LG

Breaking Entropy Bounds: Accelerating RL Training via MTP with Rejection Sampling

Reinforcement learning (RL) has become a key component in modern large language models, yet the rollout stage remains the key bottleneck in RL training pipelines. Although Multi-Token Prediction (MTP) offers a natural solution to accelerate rollouts through speculative decoding, many studies have observed that MTP acceptance rates degrade significantly during RL training, leading to limited speedup performance. To address this bottleneck, we present Bebop, a systematic study of MTP in LLM post-training, and offer practical recipes to integrate MTP into large-scale RL pipelines. First, we reveal that the MTP acceptance rate is fundamentally bounded by the fluctuation of model entropy, which demonstrates a clear negative linear relationship with the rise of entropy in the RL stage. Second, we show that probabilistic rejection sampling largely alleviates the disturbance introduced by entropy in RL compared to greedy draft sampling. We further identify that the conventional MTP training objectives (cross-entropy or KL) are suboptimal in such settings, and therefore we propose a novel end-to-end TV loss that directly optimizes multi-step rejection sampling acceptance rate, yielding ~10% acceptance rate improvements, achieving up to 95% acceptance rates and up to 25% extra inference throughput gains across mathematical reasoning, code generation, and agentic tasks. Third, we test various online MTP training strategies during RL and show that pre-RL MTP training with e2e TV loss and rejection sampling achieves a consistent acceptance rate and speedup throughout the entire RL, eliminating the need for costly online MTP updating. We provide extensive experiments and analysis that validate our findings. Experimental results show our method achieves up to 1.8x end-to-end acceleration in async RL training of Qwen3.5, Qwen3.6, and Qwen3.7 models.

Yucheng Li, Huiqiang Jiang, Yang Xu +14

Qwen Team, Alibaba Inc.

Large Language Model Reinforcement Learning Offline Reinforcement Learning Rejection Sampling Multi-Token Prediction Rejection

Jun 10, 2026cs.LG

ReCal: Reward Calibration for RL-based LLM Routing

Large language model (LLM) routing has emerged as an effective paradigm for leveraging the complementary strengths of multiple LLMs through dynamic model and reasoning-strategy selection. Recent reinforcement learning (RL)-based routing methods further improve routing quality by optimizing routing policies from interaction feedback. However, they still struggle to provide informative and comparable learning signals under heterogeneous tasks with varying difficulty. In practice, multiple objectives (e.g., correctness, format behavior) are aggregated into a single scalar reward, leading to ambiguous credit assignment and conflicting optimization signals. Moreover, reward signals exhibit significant variability across instances, where some instances produce higher or more variable rewards, introducing optimization bias that favors trivial samples over informative ones. To address these issues, we propose \textbf{ReCal}, a \textbf{\underline{Re}}ward \textbf{\underline{Cal}}ibration framework for RL-based LLM routing. We first introduce a hierarchical reward decomposition mechanism with component-wise advantage estimation. We further propose a distribution-aware optimization strategy that calibrates optimization variability through variance-aware reweighting and per-dataset normalization. Experiments on seven datasets demonstrate that ReCal consistently improves routing performance, and training stability over baselines. Code is available at https://anonymous.4open.science/r/ReCal.

Qihang Yu, Hanwen Tong, Zhengqi Zhang +5

Zhejiang University · Ant Group · Shanghai AI Laboratory

Large Language Model Routing Large Language Model Reinforcement Learning Recast

Jun 10, 2026cs.AI

Architecture-Aware Reinforcement Learning Makes Sliding-Window Attention Competitive in Math Reasoning

The rapid progress of reasoning and agentic large language models (LLMs) has increased the demand for long-context inference, but self-attention (SA) scales quadratically with context length. To address this, we study SWARR (Sliding-Window Attention with Reinforced Adaptation for Math Reasoning), a practical recipe for adapting SWA models to mathematical reasoning. SWARR has two stages: (1) efficient conversion from a pretrained SA model to SWA with supervised fine-tuning (SFT), which avoids pretraining a new base model, and (2) policy adaptation with reinforcement learning (RL). We find that SWA still underperforms SA after SFT, and we hypothesize that this gap is caused in part by a data-architecture mismatch: most SFT data are prepared for SA models and may contain long-range dependencies that are difficult for SWA to model. Because on-policy RL optimizes self-generated trajectories under the SWA constraint, it can adapt trajectories to better match SWA. Experiments on mathematical reasoning benchmarks show that this recipe substantially narrows the gap between SWA and SA, recovering much of the accuracy lost during SWA conversion while preserving the efficiency benefits of linear-complexity attention. Our central contribution is the empirical finding that RL changes the conclusion one would draw from conversion and SFT alone about SWA's viability for math reasoning.

Kai Liu, Peijie Dong, Xinchen Xie +5

Shanghai Research Institute for Intelligent Autonomous Systems, Tongji University · Hong Kong University of Science and Technology (Guangzhou) · Shanghai AI Laboratory

Mathematical Reasoning Benchmarks Large Language Model Reinforcement Learning Kimi Delta Attention Rapid Adaptation

Jun 9, 2026cs.LG

How Does Reasoning Flow? Tracing Attention-Induced Information Flow for Targeted RL in LLMs

Token-level credit assignment remains a key obstacle for reinforcement learning (RL) in large language models (LLMs), where RL recipes typically treat all tokens equally, failing to distinguish decisive reasoning steps from routine formatting or fluent filler. Recent attempts leverage model-internal signals to assign finer-grained credit, but these are often point-wise heuristics that ignore the global structure of information propagation. We propose FlowTracer, an RL framework that traces answer-targeted reasoning flow on an attention-induced directed acyclic graph in which nodes correspond to tokens and edge capacities come from aggregated attention weights and derives token credit from this global structure. The edge capacities are reweighted to retain only the influence that can reach the answer region, while enforcing local flow conservation so intermediate tokens neither lose nor gain effective mass due to path length or irrelevant branches. On this graph, FlowTracer extracts an information-flow backbone connecting the question to the answer and scores tokens by flow throughput, revealing high-impact hubs and aggregation checkpoints that mediate long-range dependencies. These derived importances are used to shape token-level rewards, enabling learning signals to focus precisely on the tokens that route information toward (or away from) correct answers and delivering consistent performance gains across a range of reasoning tasks.

Zhichen Dong, Yang Li, Yuhan Sun +9

1Shanghai Jiao Tong University · 2Alibaba Group · 3Shanghai Artificial Intelligence Laboratory

Large Language Model Reinforcement Learning LLM Reasoning Strategies Credit Assignment Information-Flow Control Offline Reinforcement Learning

Jun 9, 2026cs.AI

Reasoning or Memorization? Direction-Aware Diversity Exploration in LLM Reinforcement Learning

Reinforcement learning has become a key paradigm for eliciting reasoning abilities in large language models, where exploration is crucial for discovering effective solution trajectories. Existing exploration methods typically encourage diversity in semantic or gradient spaces, without distinguishing what drives this diversity. A trajectory may appear novel because it follows a new reasoning process, or because it varies memorized patterns and shortcuts. Rewarding both cases equally may steer exploration toward memorization rather than genuine reasoning improvement. In this paper, we propose DiRL, a Direction-Aware Reinforcement Learning framework that anchors exploration to an internal reasoning-memorization direction of the policy. Specifically, DiRL extracts this direction from model representations, constructs direction-weighted gradient features to characterize rollout updates, and shapes rewards to amplify reasoning-aligned exploration while suppressing memorization-aligned variations. DiRL integrates seamlessly into standard Group Relative Policy Optimization (GRPO). Extensive experiments on mathematical and general reasoning benchmarks demonstrate the effectiveness of DiRL, showing significant improvements over various existing exploration methods.

Jiangnan Xia, Yucheng Shi, Yu Yang +3

University of Georgia · Tencent AI Lab · The Education University of Hong Kong +1

Large Language Model Reinforcement Learning Memorization

Jun 8, 2026cs.LG

Rethinking the Divergence Regularization in LLM RL

Reinforcement learning (RL) has become a key component of post-training large language models (LLMs). In practice, LLM RL is often off-policy because of training-inference mismatch and policy staleness, making trust-region control essential for stable optimization. Mainstream methods such as PPO and GRPO approximate this control with a ratio-clipping mechanism, but the importance ratio can be a poor proxy for distributional shift in long-tailed vocabularies. Recent work such as DPPO addresses this mismatch by replacing ratio-based clipping with a divergence-based mask, yielding a trust region defined by the sampled token's absolute probability shift. However, DPPO still relies on a hard mask: once a token crosses the trust-region boundary in a harmful direction, its gradient is discarded rather than corrected. To address this, we propose Divergence Regularized Policy Optimization (DRPO), which replaces the hard mask with a smooth advantage-weighted quadratic regularizer on policy shift. DRPO preserves the same trust-region geometry as DPPO while inducing bounded, continuous gradient weights that attenuate diverging updates and provide corrective signals beyond the boundary. Experiments across model scales, architectures, and precision settings show that DRPO improves the stability and efficiency of LLM RL training.

Jiarui Yao, Xiangxin Zhou, Penghui Qi +3

1Tencent Hunyuan · 2UIUC · 3NUS

Large Language Model Reinforcement Learning Trust Region Gradient Clipping Distribution Shifts

Jun 8, 2026cs.CL

Gradient-Guided Reward Optimization for Inference-time Alignment

Ensuring the reliability of Large Language Models (LLMs) under distribution drift requires inference-time adaptation. While inference-time alignment methods such as Best-of-

N

and rejection sampling are widely used, they frame the task as a sampling-intensive, reward-guided search, leading to two key limitations: their performance is bounded by the base model's generation quality, and their reliance on imperfect reward models makes them vulnerable to reward hacking. To address these challenges, we introduce Gradient-Guided Reward Optimization (GGRO), a lightweight inference-time method that performs targeted, minimal intervention during decoding via gradient guidance. Specifically, GGRO monitors token-level entropy to identify high-uncertainty regions indicative of drift or misalignment. Upon detection, it responds by injecting nudging tokens, generated using gradient signals from an off-the-shelf reward model, to steer the generation trajectory rather than merely re-ranking samples. Experiments show that GGRO consistently improves inference-time alignment across safety, helpfulness, and reasoning benchmarks. It also increases coverage of high-quality responses and robustness to reward hacking, with minimal computational overhead. Code is available at https://github.com/lhk2004/GGRO.

Hankun Lin, Ruqi Zhang

Department of Computer Science, Purdue University

Large Language Model Alignment Reward Gradients Large Language Model Reinforcement Learning LLM Inference Optimization

Jun 8, 2026cs.LG

3SPO: State-Score-Supervised Policy Optimization for LLM Agents

Training large language models (LLMs) as autonomous agents via reinforcement learning (RL) has enabled frontier models to achieve superhuman performance in long-horizon tasks. However, existing RL algorithms operate at the trajectory level, performing policy optimization only after collecting complete episode rollouts. This coarse-grained approach faces fundamental challenges in multi-turn agent settings where rewards are sparse, delayed, and credit assignment across individual steps is critical. In this work, we propose \textbf{State-Score-Supervised Policy Optimization (3SPO)}, a novel RL algorithm that performs post-step policy optimization with dynamic state score supervision. At each step, 3SPO computes the state score based on historical success rates, supervising step-wise credit assignment, adaptive rollout and post-step policy optimization without requiring value function estimation or additional auxiliary models. Theoretically, under a per-state bandit abstraction, we show that the proposed score-supervised allocation mechanism achieves logarithmic allocation regret and provide sample-complexity guarantees for action identification, score distinguishability, and filtering stability. Experiments on ALFWorld and WebShop with Qwen2.5-1.5B/7B-Instruct show that 3SPO consistently outperforms GRPO by

+22.6\%

on ALFWorld and

+15.6

points on WebShop, while using comparable resources to achieve

2.4\times

more state exploration and

1.8\times

faster convergence. Code is available at https://github.com/genalyu/3SPO.

Yu Han, Kailing Li, Yang Jiao +4

School of Computer Science and Technology, East China Normal University · Fudan University · KAUST +1

Large Language Model Reinforcement Learning Offline Reinforcement Learning Large Language Model Agents

Jun 7, 2026cs.LG

Reformulate LLM Reinforcement Learning for Efficient Training under Black-box Discrepancy

Reinforcement Learning (RL) has emerged as a pivotal post-training paradigm, yet it frequently suffers from unpredictable sub-optimum performance or even training collapses. Recent findings attribute these failures to a hidden train-inference discrepancy (or mismatch), stemming from the disparate underlying engines and architecture. We find that the training policy can actively self-correct such a discrepancy when provided with an appropriate learning signal. Then, we further empirically identify a discrepancy tolerance region: within this region, aggressively narrowing the discrepancy can suppress policy exploration and reduce learning efficiency, whereas outside this region, reducing excessive discrepancy improves optimization consistency and raises the achievable local performance ceiling. According to such findings, we formulate this problem as a Discrepancy-Constrained Markov Decision Process (DCMDP), where reward maximization is coupled with a constraint that aligns training-Inference behavior, achieving stable dual-objective optimization. To adaptively balance performance improvement and discrepancy control, we introduce a Lagrangian relaxation mechanism that dynamically adjusts the relative weight of the two objectives according to the current degree of discrepancy violation. This enables stable dual-objective optimization: the policy is allowed to explore freely within the tolerance region, while being guided back when the discrepancy exceeds the safe boundary. Empirically, DCMDP significantly improves the performance of 8B dense model (Qwen-3-8b) and 30B Mixture-of-Expert model (Qwen-3-30bA3b), and enables a heterogeneous training paradigm, where LLMs can be optimized in high-fidelity training setup while being explicitly aligned for low-cost, resource-constrained inference deployment.

Jiashun Liu, Runze Liu, Xu Wan +3

1Hong Kong University of Science and Technology · 2Zhejiang University · 3Tianjin University

Large Language Model Reinforcement Learning Training-Inference Mismatch Discrepancy Post-Training Markov Decision Processes Convex Relaxation

Jun 7, 2026cs.CL

Back on Track: Aligning Rewards and States for Reasoning in Diffusion Large Language Models

Reinforcement learning (RL) holds immense promise for enhancing the reasoning capabilities of diffusion large language models (dLLMs). However, progress is fundamentally constrained by a dual misalignment between authentic generation trajectory and the gradient update process: (i) Process-reward misalignment. Sparse, terminal rewards are indiscriminately assigned to all intermediate steps of the generation process, failing to provide discriminative credit assignment. (ii) State-trajectory misalignment. Policy updates are often diverted toward artificial, out-of-trajectory states, squandering gradients on less informative samples. To address these limitations, we introduce Process Aligned Policy Optimization (PAPO), a novel framework that holistically aligns the RL update with the dLLM's generative trajectory via Step-Aware Process Rewards (SPR) that transform sparse terminal rewards into dense, step-wise credit, and Entropy-Guided Historical Re-enactment (EHR) that replays authentic trajectories at high-uncertainty steps. Extensive experiments on four benchmarks demonstrate that PAPO significantly outperforms baselines, achieving gains of up to 4.5% on GSM8K, 4.8% on MATH500, 42.2% on Countdown and 16.1% on Sudoku.

Yawen Shao, Jie Xiao, Kai Zhu +6

University of Science and Technology of China · Tongyi Lab · Northeastern University

Large Language Model Reinforcement Learning Diffusion Language Models Diffusion-Based Reinforcement Learning Methods Reward Gradients Large Language Model Alignment

Jun 7, 2026cs.AI

GIFT: LLM-Guided State-Reward Interface for Financial Reinforcement Learning

Financial portfolio trading is naturally formulated as a reinforcement learning problem, where an agent sequentially rebalances assets under changing market conditions to balance return, risk, and transaction costs. Yet in non-stationary markets, raw OHLCV states and short-horizon return rewards often provide an under-specified learning interface, motivating large language models as a way to inject financial knowledge into state and reward design while constraining open-ended generation. To this end, we propose GIFT, an LLM-guided framework for state-reward interface design in PPO-based financial reinforcement learning. Rather than using the LLM to make trading decisions, GIFT uses Factor-guided State Enhancement to generate state features from financial-factor primitives, Risk-rule-guided Reward Shaping to generate auxiliary rewards from portfolio-risk rules, and Diagnostic-guided Refinement to revise candidate interfaces using PPO rollout diagnostics. After refinement, GIFT fixes the selected state-reward interface before evaluation, with no further LLM queries or interface updates at test time. Comprehensive rolling-window experiments across diverse market regimes and portfolio scenarios demonstrate that GIFT improves learning-signal quality and out-of-sample risk-adjusted portfolio performance over baselines. Code and data are available at: https://github.com/KAG778/GIFT .

Yanyan Wu, Boyi Zhang, Yanlin Liu +10

East China University of Science and Technology · University of Science and Technology of China · Southwestern University of Finance and Economics +5

Large Language Model Reinforcement Learning Offline Reinforcement Learning Response-Level Rewards Transcriptomics

Jun 7, 2026cs.LG

Sparrow: Sparse Rollout for Stable and Efficient Long-context RL of Large Language Models

Despite being powerful, reinforcement learning with verifiable rewards (RLVR) induces extremely long COT, making it computationally expensive. Since RLVR per-step cost is dominated by long-context rollout generation, sparse attention offers a promising way to accelerate dense rollout. However, sparse rollouts require a delicate stability-efficiency tradeoff: overly aggressive sparsity causes collapse, while overly lenient sparsity gives insufficient speedup. In this work, we study this tradeoff through sparse-to-dense actor-policy mismatch. We first observe that sparse rollout collapse is not driven by uniform degradation across tokens: most sparse tokens align perfectly with dense even under aggressive sparsity. Motivated by this, we hypothesize that sparse rollout training remains stable if the lower tail of per-token actor-policy mismatch stays above a critical threshold throughout the trajectory. We introduce a dynamic sparsity schedule that keeps this tail statistic constant during generation and validate our hypothesis. Across Qwen3 thinking-family models, keeping the tail mismatch statistic near a consistent threshold generally enables stable training. We then use a cost model to find the sparsity schedule for maximum speedup under this mismatch threshold, achieving 2.2x, 2.4x, and 2.0x rollout speedups when training Qwen3-1.7B, Qwen3-4B, and Qwen3-8B. Empirically, we show the thresholds generalize to a larger model (Qwen3-14B) and another RL domain (coding). Finally, our analysis naturally motivates DistillSparse: lightweight LoRA-based distillation on sparse rollout lets more aggressive sparsity reach the same sparse-to-dense mismatch threshold, yielding higher speedup.

Yang Zhou, Ranajoy Sadhukhan, Zhaofeng Sun +7

Carnegie Mellon University · Cornell University · Intel +1

Autoregressive Rollout Large Language Model Reinforcement Learning Reinforcement Learning With Verifiable Reward Sparsity Bird

Jun 6, 2026cs.CL

Rewrite to Translate, Translate to Reward: Reinforcement Learning for Source Rewriting in Machine Translation

Rewriting source text with large language models (LLMs) before translation has been shown to improve machine translation (MT) quality. However, we find that prompt-based rewriting can degrade translation quality rather than improve it, particularly when smaller LLMs, such as 4B-parameter models, are used. We argue that this limitation stems from the difficulty of controlling rewriting behavior through natural-language prompts alone: a rewrite is useful only if it improves downstream translation, yet existing prompt-based methods do not explicitly optimize for this signal. To address this issue, we propose RLSR (Reinforcement Learning for Source Rewriting), a reinforcement learning framework that trains the rewriting model with a reward based on the downstream translation-quality improvement produced by each rewrite. Experiments across six MT systems and 16 language pairs show that our 4B RLSR-trained rewriting models significantly outperform both the no-rewriting baseline and prompt-based rewriting baselines at the same model scale, while remaining competitive with baselines that use a 235B LLM.

Boxuan Lyu, Haiyue Song, Zhi Qu +3

Institute of Science Tokyo · Preferred Networks, Inc. · Nara Institute of Science and Technology

Machine Translation Quality Machine Translation Rewrite Large Language Model Reinforcement Learning Language Modeling

Jun 5, 2026cs.LG

SAW: Stage-Aware Dynamic Weighting for Multi-Objective Reinforcement Learning in Large Language Models

Although multi-objective reinforcement learning (MORL) is central to aligning large language models with complex human preferences, the prevailing practice of static weighted summation overlooks a more fundamental phenomenon: reward learning is markedly asynchronous across objectives. Well-learned dimensions quickly produce homogeneous, low-variance signals whose residual noise contaminates the aggregated reward (in GRPO) or occupies a fixed share of the advantage budget (in GDPO), interfering with the scarce yet high-value signals carried by under-learned dimensions. To address this asynchrony, we propose Stage-Aware Dynamic Weighting (SAW), a lightweight, algorithm-agnostic dynamic weighting mechanism. SAW utilizes the coefficient of variation (CV) as a scale-invariant proxy for real-time informativeness, reweighting each dimension's reward or advantage contribution by its relative informativeness within the batch. Unlike gradient-based methods that require multiple forward and backward passes, SAW relies solely on batch-level statistics, introducing nearly negligible computational overhead. Experiments on tool-calling and text summarization tasks demonstrate that SAW consistently improves both training efficiency and final performance under both GRPO and GDPO frameworks, confirming it as a general-purpose plug-in for multi-reward LLM alignment. Our code is available at https://github.com/Zhaolutuan/SAW

Yuchen He, Baolong Bi, Shenghua Liu +7

Institute of Computing Technology, Chinese Academy of Sciences · University of Electronic Science and Technology of China

Multi-Objective Reinforcement Learning Large Language Model Reinforcement Learning Response-Level Rewards Large Language Model Alignment Summarization

Jun 3, 2026cs.CL

Read the Trace, Steer the Path: Trajectory-Aware Reinforcement Learning for Diffusion Language Models

Diffusion large language models (dLLMs) generate responses by iteratively unmasking and revising many positions in parallel. This process leaves a rich denoising trace depicting which tokens become confident, which remain unstable, and when commitments form. Existing dLLM reinforcement learning methods use this signal only weakly. Flat rollouts are cheap, but assign a single outcome reward to the whole trajectory. Tree rollouts provide finer, verifiable training signals by branching partial trajectories and propagating leaf rewards upward, but are compute intensive. We ask whether the denoising trace itself can provide tree-like supervision without tree-level compute. We introduce CAPR (Cached-Amortized Path Refinement), a dLLM-RL algorithm that summarizes the denoising trace into a compact path state, uses cached trajectory states to generate cheap sibling continuations, and trains a block-level value head for local block-wise supervision. Under a block-wise unmasking schedule, CAPR records path-state and block-progress features, then redistributes the final outcome reward across blocks according to the tokens revealed in each block. This trains the value head to convert one sparse reward into block-level PPO weights. CAPR therefore recovers much of the granularity of tree search while avoiding full tree expansion, reducing rollout-generation cost to roughly 0.75x that of flat rollouts and 0.6x that of tree rollouts (under standard settings). Across 4x4 Sudoku, Countdown, GSM8K, and Math500, on dense and mixture-of-experts LLaDA backbones, CAPR sets a new state of the art for RL-tuned dLLMs at 256- and 512-token budgets. On Sudoku, it matches the strongest tree-structured baseline at less than one third of the per-step compute.

Anant Khandelwal, Manish Gupta

Microsoft AI, India

Large Language Model Reinforcement Learning Diffusion Language Models Diffusion-Based Reinforcement Learning Methods Denoising Trajectory Trajectory-Aware Hidden-State Analyses Traces

Jun 2, 2026cs.LG

Efficient Hyperparameter Optimization for LLM Reinforcement Learning

Reinforcement learning (RL) for large language models (LLMs) is highly sensitive to hyperparameter configurations, making hyperparameter optimization (HPO) essential yet computationally expensive. Existing multi-fidelity HPO methods remain inefficient for LLM RL due to the massive model scale and resource-intensive training cycles. In this paper, we propose Joint Fidelity Hyperparameter Optimization (JF-HPO), which simultaneously adapts both model size and training budget as fidelity. JF-HPO is empowered by: (i) it leverages a small proxy model of the target LLM for efficient training and evaluation in each HPO trial; (ii) it integrates carefully designed early-stopping strategies based on training dynamics; (iii) it introduces an efficient checkpointing mechanism to eliminate redundant computations. Compared with existing HPO methods, JF-HPO significantly improves the computational efficiency of each trial (up to 14.9 times), while achieving better or competitive predictive accuracy under the same time budget. Notably, compared with utilizing hyperparameter configurations from the VeRL Recipe, JF-HPO delivers performance improvements ranging from 5.8% to 111.6%.

Minping Chen, Bowen Xiao, Du Liang +2

1The Hong Kong University of Science and Technology (Guangzhou) · 3China United Network Communications Group · 2The Hong Kong University of Science and Technology

Large Language Model Reinforcement Learning General Grid Search Framework

May 30, 2026cs.CL

Internalize the Temperature: On-Policy Self-Distillation as Policy Reheater for Reinforcement Learning

Reinforcement learning from verifiable rewards improves the reasoning ability of large language models, but often suffers from entropy collapse, in which increasingly concentrated policies reduce rollout diversity and useful learning signals. Existing remedies either constrain the RL objective (e.g., entropy regularization) or adjust sampling temperature during rollout collection, but these interventions remain external to the model parameters. We propose Temperature-Scaled On-Policy Self-Distillation (TS-OPSD), a lightweight policy reheating method that internalizes the exploratory effect of temperature into model parameters. Starting from an entropy-collapsed RL checkpoint, TS-OPSD constructs a self-teacher by applying high-temperature scaling to the model's own logits, then distills the resulting smoother distribution back into the student. This policy reheating requires no external teacher, privileged data, or additional inference cost. Experiments on Qwen3-4B-Base and Qwen3-8B-Base show that policy reheating yields a stronger initialization for continued RL than both standard continued RL and rollout-level temperature reheating. Further analyses show that TS-OPSD mainly reduces output sharpness while preserving intermediate representations, top candidate sets, and reasoning capability. These results suggest that entropy restoration can serve as a simple post-collapse intervention for extending reasoning-oriented RL.

Xuewei Yang, Jiachen Yu, Jie Wu +3

Tsinghua University

Unsupervised On-Policy Self-Distillation Large Language Model Reinforcement Learning Internalization Layer-Selective Rehearsal Strategy

May 29, 2026cs.LG

PR2: Predictive Routing Replay for MoE-Based LLM Reinforcement Learning

Mixture of Experts (MoE) Large Language Models (LLMs) achieve strong performance at scale. However, reinforcement learning (RL) on MoE-based LLMs often suffers from training instability. A root cause is router drift, i.e., expert activations can change drastically across model updates and differ between disaggregated rollout and training phases, causing large rollout--training mismatch and unstable importance sampling weights in PPO-style RL algorithms. Routing replay mitigates this issue by freezing the replay route within each reasoning trajectory, but it ignores how the router evolves under off-policy updates and thus causes router staleness. To address this limitation, we propose Predictive Routing Replay (PR2), which augments each router with a lightweight evolution predictor that learns to anticipate short-horizon router evolution. During the rollout phase, we use the predictive routing distribution to apply top-

k

routing, enabling gradients to reach experts that are likely to become active after updates. During the training phase, we replay the resulting predicted route to retain consistency for stable importance estimation. Theoretical analysis and experiments support that PR2 reduces routing-induced mismatch, improves RL stability, and yields stronger performance across various reasoning benchmarks.

Daize Dong, Junlin Chen, Haolong Jia +9

♣Rutgers University · ♦AMD · ⋆MBZUAI

Large Language Model Reinforcement Learning Replay

May 29, 2026cs.LG

ARCA: Adapter-Residual Credit Assignment When Token Signals Degenerate

Token-level credit assignment for language-model reinforcement learning is usually formulated as if the policy were fully trainable, while practical LLM-RL pipelines often rely on parameter-efficient fine-tuning, especially LoRA. We argue that this separation hides a structural failure mode. Under LoRA, the policy is restricted to a low-rank neighborhood of the reference model, so the per-token output-distribution differences used by common intrinsic credit signals, surprisal, entropy reduction, and policy divergence, can become degenerate after within-trajectory normalization, either approaching uniform weights or concentrating on a small set of task-agnostic positions. We formalize this behavior and propose measuring it directly with concentration diagnostics such as weight Gini and effective-token ratio. We then introduce \emph{Adapter-Residual Credit Assignment} (ARCA), a lightweight alternative that derives token salience from the adapter's own hidden-state residual,

\|h^{\text{adapted}}_t - h^{\text{base}}_t\|_2

. ARCA asks where the adapter actually changes the model, rather than where the output distribution appears uncertain or shifted, and requires no learned reward model, value head, or tree construction. In a compact MATH/Qwen3-1.7B GRPO sweep, ARCA exhibits the predicted non-degenerate middle-regime credit distribution under matched rollout budgets and remains competitive with rank-matched baselines.

Rodney Lafuente-Mercado

Scale AI

Credit Assignment Large Language Model Reinforcement Learning Single-Token Output Distributions Autoregressive Language Models Divergence

May 29, 2026cs.CL

Consolidating Rewarded Perturbations for LLM Post-Training

Post-training of language models is commonly framed as a sample-score-update loop implemented by gradient descent. A recent line of work, exemplified by RandOpt, relocates this loop to weight space, sampling Gaussian perturbations around a pretrained model and ensembling the top-K rewarded specialists at inference. While competitive with PPO and GRPO under matched training compute, this prediction-level ensemble incurs K forward passes per test example and does not extend cleanly to free-form generation. We ask whether the rewarded population can instead be folded into a single deployable model, replacing the inference-time ensemble with one consolidated update. A split-half analysis over 25 model-task pairs reveals reproducible low-rank structure in every case. We turn this geometry into CoRP (Consolidating Rewarded Perturbations), a gradient-free operator that combines reward-weighted aggregation, compatibility-aware reweighting, and a held-out validation gate, with no gradient flowing through the language model. Across five language models from 0.5B to 8B and five tasks covering math, code, and creative writing, CoRP improves the base model by 8.1 points on average. Using one tenth of RandOpt's perturbation budget, CoRP exceeds single-inference RandOpt by 6.5 points and recovers more than half of the gain of the 50-pass majority-vote ensemble, at one forward pass per test example.

Zheyu Zhang, Shuo Yang, Gjergji Kasneci

Technical University of Munich

Large Language Model Reinforcement Learning Post-Training Majority Voting

May 29, 2026cs.LG

DARTS: Distribution-Aware Active Rollout Trajectory Shaping for Accelerating LLM Reinforcement Learning

Reinforcement Learning (RL) has become pivotal for improving model capabilities yet suffers from rollout efficiency bottlenecks due to the long-tail response length distribution. While existing works mitigate the impact of long tails via prompt-level tail scheduling, we focus on the root source of inefficiency: the distribution itself. Specifically, we characterize the long-tail distribution at a finer granularity, identifying intra-prompt long tails, and revealing that they frequently consist of ineffective verbosity. To address this, we propose a novel paradigm of active distribution shaping to shape the rollout distribution towards conciseness and certainty, thereby fundamentally resolving tail-induced overheads. We achieve this through a distribution-aware trajectory sampling mechanism, which selects trajectories from a redundant exploration space for each prompt, and an adaptive redundancy allocation scheme to maximize both shaping effectiveness and system efficiency. Experiments demonstrate significant acceleration over state-of-the-art systems by up to 1.77x without compromising model performance.

Yujie Wang, Siwei Chen, Longzan Luo +4

School of Computer Science & Beijing Key Laboratory of Software and Hardware Cooperative Artificial Intelligence Systems, Peking University, Beijing, China · School of Artificial Intelligence, Shanghai Jiao Tong University, Shanghai, China · Institute of Computational Social Science, Peking University (Qingdao), Qingdao, China

Large Language Model Reinforcement Learning Autoregressive Rollout Long-Tailed Distribution Robust Trajectory

May 28, 2026cs.CL

Configurable Reward Model for Balanced Safety Alignment

Aligning large language models (LLMs) to heterogeneous and rapidly evolving safety requirements remains a critical challenge. Existing instruction-tuned LLMs and standalone safety classifiers often fail to generalize to new safety configurations, motivating the need for Reward Models (RMs) that are explicitly configurable to changing specifications. We introduce the Configurable Safety Reward Model (CSRM), which is jointly optimized for calibrated safety compliance and reward modeling. Our approach is supported by configuration-targeted data augmentation that enforces instruction adherence while preserving relative severity structure. The resulting RM is sensitive to fine-grained safety configurations and conversational nuances, substantially improving generalization to previously unseen safety configurations. CSRM achieves state-of-the-art performance on recent configurable safety benchmarks, including CoSApien (94.6% F1) and DynaBench (75.8% F1), without requiring additional human annotation. When used for downstream safety alignment, CSRM yields LLMs with a significantly improved helpfulness-safety tradeoff compared to existing baselines.

Zhengping Jiang, Mehran Khodabandeh, Akash Bharadwaj +5

Johns Hopkins University · Meta · Meta Superintelligence Labs

Safety Alignment Large Language Model Alignment Safety Benchmarks Large Language Model Reinforcement Learning Progress Reward Modeling

May 28, 2026cs.LG

ESPO: Early-Stopping Proximal Policy Optimization

When a large language model under reinforcement learning commits a wrong reasoning step early in a trajectory, standard algorithms force it to keep generating until the maximum horizon, spending compute on tokens that never receive positive reward and polluting advantage estimates with post-failure noise. We propose ESPO (Early-Stopping Proximal Policy Optimization), which detects trajectory failure on-the-fly and terminates rollouts early. At each generation step, ESPO computes a surrogate regret using only the logits already computed during sampling, and terminates when the smoothed cumulative regret significantly exceeds its estimated values. Truncated trajectories are treated as absorbing failure states with a terminal reward, concentrating negative temporal-difference (TD) errors near the detected failure step without any additional reward model or human annotation. On DeepSeek-R1-Distill-Qwen-7B trained for mathematical reasoning, ESPO surpasses PPO on AIME~~2024 (46.28% vs. 45.25%), AMC~~2023 (85.83% vs. 82.94%), and MATH-500 (87.42% vs. 85.43%), while saving more than 20% rollout tokens cumulatively.

Zihang Li, Rui Zhou, Yingcheng Shi +8

1Tongyi Lab , Alibaba Group · 2Peking University

Large Language Model Reinforcement Learning Proximal Policy Optimization

May 28, 2026cs.LG

Hista and Numca: Estimate State Value Effectively for LLM Reinforcement Learning

Reinforcement learning (RL) refines large language models (LLMs) by directly optimizing model behavior through reward signals. While accurate state value estimation is critical for stable training in classical RL, it remains an underexplored challenge in LLM post-training. In this work, we introduce the State Value Estimation Benchmark (SVEB) to assess state estimation within existing RL frameworks and show that critics in standard approaches like PPO collapse to a coarse group-average baseline. To address this, we propose two techniques: Numca, which leverages numerical spans as gradable milestones for state value estimation, and Hista, a framework that uses LLM's hidden states as representation to weighted average disjoint rollouts and their return. Extensive experiments demonstrate that both methods yield more accurate state value estimates and enhance training performance across different RL algorithms and model sizes without incurring significant computational overhead.

Zizhe Chen, Jiqian Dong, Yizhou Tian +4

Department of Computer Science and Engineering, The Chinese University of Hong Kong · Huawei Technologies Ltd.

Large Language Model Reinforcement Learning State Estimation

May 28, 2026cs.AI

The Curse of Helpfulness: Inverse Scaling Law in Robustness to Distractor Instructions via DistractionIF

Large Language Models (LLMs) are increasingly deployed in agentic and retrieval-augmented generation (RAG) systems, where they must execute user-specified tasks over externally provided reference text. In practice, such context is often unstructured and contaminated with benign but instruction-like semantic noise, such as editorial comments and system traces, which should be treated strictly as data. We introduce DistractionIF, a benchmark designed to evaluate robustness against such distractor instructions in reference text. Across a broad range of models, we observe a consistent inverse scaling phenomenon: larger models are often less robust, with performance dropping by up to 30 points as scale increases. Mechanistically, our perplexity analysis reveals that scaling erodes the probabilistic boundary between robust and distracted behaviors, making models increasingly prone to over-interpreting noise as instructions. To address this, we demonstrate that reinforcement learning, specifically Group Relative Policy Optimization (GRPO), can restore this boundary, improving robustness by up to 15.5% without compromising general instruction-following capability. Our findings highlight a critical instruction-following robustness gap in reference-grounded tasks and establish reinforcement learning as a promising path for enforcing strict data-instruction separation at scale.

Zeli Su, Zhankai Xu, Tianlei Chen +4

Minzu University of China, Beijing, China · Ant Group, Hangzhou, China · Renmin University of China, Beijing, China +1

Large Language Model Reinforcement Learning Instruction-Tuned Models Distractor Scaling Assistant Scaling Laws

May 28, 2026cs.LG

GDSD: Reinforcement Learning as Guided Denoiser Self-Distillation for Diffusion Language Models

Reinforcement learning (RL) can be used to improve the policy (denoiser) of diffusion large language models (dLLMs), while being hindered by the intractability of the policy likelihood. A dominant and efficient family of methods replaces the likelihood in standard RL with its evidence lower bound (ELBO), estimated from randomly masked sequences. Despite being well aligned with pre-training, these approaches introduce bias through training--inference mismatch by using the ELBO as a likelihood surrogate, which can degrade performance. In this work, we propose Guided Denoiser Self-Distillation (GDSD) to directly distill the denoiser of dLLMs from an advantage-guided self-teacher, derived from the closed-form optimum of reverse-KL regularized RL. GDSD matches the dLLM's denoiser logits to the teacher's via a normalization-free objective, which reduces RL to likelihood-free self-distillation and thus bypasses the TIM biases. Recent ELBO-based methods emerge as instances of applying different distillation divergences, but with diagnosable pathologies that GDSD avoids. On planning, math, and coding benchmarks with LLaDA-8B and Dream-7B, GDSD consistently outperforms prior state-of-the-art ELBO-based methods with a more stable training reward dynamics, achieving test-accuracy improvements of up to

+19.6\%

. These results suggest that direct denoiser self-distillation, without relying on an ELBO likelihood surrogate, can provide a more stable and effective RL procedure for dLLMs. Code is available at https://github.com/GaryBall/GDSD.

Xiaohang Tang, Keyue Jiang, Che Liu +4

UNIST · University of Basel

Diffusion Language Models Large Language Model Reinforcement Learning Unsupervised On-Policy Self-Distillation Denoising Diffusion Probabilistic Model Lower Bounds

May 27, 2026cs.LG

When LLM Reward Design Fails: Diagnostic-Driven Refinement for Sparse Structured RL

For sparse, structured reinforcement-learning tasks with semantic reward-function interfaces, LLM-generated reward shaping is better framed as debugging than one-shot generation. We study PPO-trained agents using MiniGrid as core evaluation and MuJoCo as boundary stress test. Our audit finds two dominant one-shot failure modes -- reward flooding and semantic/API misunderstanding -- plus a rarer weak-shaping case. We propose diagnostic-driven iterative refinement, where training diagnostics and a failure-mode taxonomy guide targeted reward-function revision. Refinement improves DoorKey-8x8 from 2.3% to 97.6% and KeyCorridor from 31.2% to 86.7% with high seed-to-seed variance. Controls show these gains are not from retrying or extra training: metrics-only re-prompting yields large drops, while a static-vocabulary control recovers much of the gap (87.6%; 70.7%), showing the taxonomy prompt is a major mechanism and dynamic labels provide only partially isolated incremental evidence. Budget-matched and Best-of-3 comparisons separate refinement from selection and training-time effects. Component-removal tests, sensitivity analyses, and an audit against author labels provide converging evidence for the debugging interpretation while revealing calibration limits. Continuous-control results show the boundary: success-based diagnostics can misfire in dense-reward locomotion, and return-trend feedback removes one false-positive mechanism without robust gains. The low-call protocol is a cost contrast with population-based reward search, not a benchmark comparison. In four crossed-variance-design environments, point estimates suggest larger gains when LLM reward-function variance dominates but bootstrap intervals are wide. The method is bounded to sparse structured tasks with reliable interfaces under PPO; fields like event_text may help, hurt, or be neutral.

Youting Wang, Yuan Tang, Bowen Liu +2

Large Language Model Reinforcement Learning Refinement Latent Failure Patterns Iterative Refinement

May 27, 2026cs.AI

DenoiseRL: Bootstrapping Reasoning Models to Recover from Noisy Prefixes

Reinforcement learning has become a central paradigm for advancing reasoning in large language models, yet most existing methods still depend on stronger teacher models or heavily curated difficult datasets, limiting scalable capability improvement. In this paper, we introduce DenoiseRL, a reinforcement learning framework that substitutes external supervision with recovery-oriented optimization over failures from weak models. Instead of relying on stronger supervision or carefully engineered data, DenoiseRL learns directly from incorrect reasoning traces by converting them into opportunities for improvement, making training more scalable and less dependent on external resources. This yields a richer and more diverse learning signal, improving exploration efficiency from imperfect model behavior. As a result, DenoiseRL improves reasoning performance and overall training efficiency while reducing the need for expensive data curation or stronger teacher models. Empirically, DenoiseRL consistently outperforms strong on-policy RL baselines across competitive mathematical and general reasoning benchmarks and promotes stronger self-corrective behavior as training difficulty increases, highlighting an effective and scalable alternative pathway for improving reasoning in large language models.

Caijun Xu, Changyi Xiao, Zhongyuan Peng +1

1Fudan University · 2Shanghai Innovation Institute

Large Language Model Reinforcement Learning Large Reasoning Models Reinforcement Learning Baseline

May 26, 2026cs.LG

Guiding LLM Post-training Data Engineering with Model Internals from Sparse Autoencoders

Model internals encode rich information about how a large language model (LLM) processes its training data; however, post-training data engineering largely relies on external signals and ignores rich intrinsic signals lying in model internals. We propose SAERL, a data engineering framework for LLM reinforcement learning (RL). It models three intrinsic data properties: diversity, difficulty, and quality, using model internals extracted with Sparse Autoencoder (SAE), an advanced mechanistic interpretability tool. Each property grounds a concrete data engineering operation: SAE-space clustering with moderate batch mixing for batch diversity control, a difficulty proxy for easy-to-hard curriculum ordering, and a quality probe for data filtering. SAERL improves average accuracy by 3.00% over vanilla GRPO and reaches target accuracy with 20% fewer training steps on Qwen2.5-Math-1.5B, with consistent gains across model scales and RL algorithms. Experiments show that SAE transfers effectively across model families and scales, serving as a lightweight and reusable data engineering tool. These results demonstrate that model internals are a powerful and practical source of signals for post-training data engineering.

Yi Jing, Zao Dai, Jinwu Hu +4

Tsinghua University

Large Language Model Training Large Language Model Reinforcement Learning Mechanistic Interpretability