The paper introduces COPC, a Coupled Off-Policy Correction method for asynchronous reinforcement learning of large language models. COPC coordinates policy-side and advantage-side corrections by combining token-level ratio masking with two-sided clipped-ratio weighting of TD residuals, addressing both policy mismatch and advantage staleness. Experiments show COPC outperforms existing asynchronous baselines on tool-integrated mathematical reasoning and search tasks, while maintaining training stability and minimal overhead.
By Zicheng Hu, Zhijian Zhou, Xuan Zhang, Yuchen Liu, Cheng Chen, Yuan Li, Qi Gu, Yan Feng, Hongyan Hao, Chao Qu
The paper introduces SORL, a framework that stabilizes off‑policy reinforcement learning for long‑horizon large language model agents. It identifies two key instability sources—token‑level policy granularity mismatches and high‑variance off‑policy updates—and proposes turn‑level importance sampling and clipping‑triggered normalization to align optimization with multi‑turn interactions. Two instantiations, SO‑PPO and SO‑GRPO, are evaluated on open‑domain, multi‑hop, and medical QA benchmarks, as well as on asynchronous RL for mathematical reasoning, showing improved robustness without the need for early stopping or heuristic tuning.
By Chenliang Li, Adel Elmahdy, Alex Boyd, Zhongruo Wang, Siliang Zeng, Alfredo Garcia, Parminder Bhatia, Taha Kass-Hout, Cao Xiao, Mingyi Hong
arXiv:2606. 30789v1 Announce Type: new Abstract: Group Relative Policy Optimization (GRPO) has become a standard tool for improving the reasoning ability of large language models, yet its training dynamics are still described empirically: reward trajectories are fit with low-parameter functional forms whose constants carry no mechanistic meaning, and hyperparameter choices remain a matter of trial and error.
By Rajat Ghosh, Datta Nimmaturi, Aryan Singhal, Vaishnavi Bhargava, Henry Wong, Johnu George, Debojyoti Dutta
arXiv:2602. 04879v3 Announce Type: replace-cross Abstract: Reinforcement learning (RL) has become a cornerstone for fine-tuning Large Language Models (LLMs), with Proximal Policy Optimization (PPO) serving as the de facto standard algorithm.
By Penghui Qi, Xiangxin Zhou, Zichen Liu, Tianyu Pang, Chao Du, Min Lin, Wee Sun Lee
The paper investigates how enrichment—adding intermediate guidance to sparse rewards in reinforcement learning—affects policy-gradient updates. It shows that omitting or truncating importance-weight corrections implicitly reweights rewards, and decomposes the resulting gradient error into scale, rotation, and variance components. The authors propose a sequential Monte Carlo weight correction that mitigates variance growth, and apply their analysis to a fine-tuned Qwen3-1.7B model on OpenMathReasoning, illustrating how enrichment helps prevent collapse in sparse domains.
By Jinwoo Kim, Shraddha Barke
The paper investigates why Group Relative Policy Optimization (GRPO) benefits from per‑prompt normalization by examining the local curvature of the sequence‑level policy gradient. It shows that standard deviation normalization acts as an adaptive gradient, yielding a provably faster convergence rate than unnormalized REINFORCE under mild conditions, with the improvement tied to the average within‑prompt reward standard deviation. The authors also propose IS‑GRPO, an importance‑sampling variant that maintains alignment with the full gradient and offers a tighter convergence guarantee, and empirically validate these theoretical insights on GSM8K and MATH datasets at 1.5B and 7B model scales.
By Cheng Ge, Caitlyn Heqi Yin, Hao Liang, Jiawei Zhang
The paper introduces Anchored Neighborhood Optimization (ANO), a new policy‑optimization method that directly designs a smooth, bounded gain field for the probability‑ratio surrogate objective. ANO anchors the identity map at a ratio of one, peaks at a specified trust‑region boundary, and limits the influence of extreme off‑policy samples while providing a bounded, redescending pull on outliers. Empirical results show ANO consistently outperforms existing methods on Atari and MuJoCo benchmarks, and it remains robust under aggressive learning‑rate settings.
By Yiheng Zhang, Yiming Wang, Kaiyan Zhao, Zhenglin Wan, Jiayu Chen, Leong Hou U
arXiv:2609.36816v1 Announce Type: new
Abstract: Reinforcement learning (RL) has become a cornerstone for improving the reasoning capabilities of large language models (LLMs), but the need for on-poli...
By Ruichuan Huang, Jinghan Liu, Congliang Chen
arXiv:2609.36945v1 Announce Type: new
Abstract: We study the learning dynamics of fine-tuning a policy model on self-generated and reward-weighted data, with particular focus on a generalized version...
By Zhiwei Wang, Yanxi Chen, Yaliang Li, Bolin Ding
Gibbs‑Gap Policy Optimization (GGPO) is a new off‑policy reinforcement learning method that avoids high‑variance learner‑to‑behavior importance weights by minimizing a log‑mean‑exp contrast of KL‑adjusted rewards across responses to the same prompt. Derived from KL‑regularized policy optimization, GGPO guarantees that a learner’s single‑response expected reward is at least the behavior policy’s best‑of‑K reward minus bounded terms, and it includes a reference‑refresh mechanism that empirically improves performance. Experiments on benchmark problems and mathematical reasoning tasks show GGPO outperforming GRPO and AGRO, especially under training‑inference mismatch and with fixed offline data.
whyItMatters":"The paper demonstrates a practical improvement in off‑policy RL for language models by reducing variance and ensuring better reward guarantees, which is directly supported by the reported empirical gains and theoretical bounds."
By Xuanchang Zhang, Tong Zhang
arXiv:2607. 28582v1 Announce Type: new Abstract: On-policy self-distillation (OPSD) is a promising approach to improve reasoning language models, but it remains brittle in practice: making it work reliably often requires substantial engineering effort.
By Jiawei Xu, Minghui Liu, Juzheng Zhang, Tom Goldstein, Furong Huang
arXiv:2607. 04470v1 Announce Type: cross Abstract: 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.
By Faid Keddouri, Sohaib Houhou, Aissa Boulmerka, Nadir Farhi