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
arXiv:2610.14440v1 Announce Type: cross
Abstract: Asynchronous reinforcement-learning post-training of large language models is computationally efficient but inherently off-policy: the learner update...
By Shayan Mohajer Hamidi, Kayhan Behdin, Wenhui Zhu, Aida Rahmattalabi, Jincheng Cao, Jelena Markovic-Voronov, Alborz Geramifard
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 identifies a failure mode in Proximal Policy Optimization (PPO) critics called Value Flattening, where state values vary sharply across intermediate states while critic predictions remain flat. The authors analyze this phenomenon theoretically and empirically, linking it to an implicit variance penalty and redundant updates from temporally correlated states. They propose Sparse Proximal Policy Optimization (SP³O), which applies the value loss to only a few well‑separated states, and demonstrate that this sparse supervision mitigates Value Flattening and improves policy performance on Qwen3‑Base across model sizes and evaluation suites.
By Yizhuo Li, Jianhao Yan, Yun Luo, Zhi Wang, Futing Wang, Rong-Xi Tan, Kanghui Tian, Ganqu Cui, Ning Ding, Peilin Zhao, Yafu Li, Yu Cheng
arXiv:2606. 08779v1 Announce Type: new Abstract: Reinforcement Learning (RL) has emerged as a pivotal post-training paradigm, yet it frequently suffers from unpredictable sub-optimum performance or even training collapses.
By Jiashun Liu, Runze Liu, Xu Wan, Jing Liang, Hongyao Tang, Ling Pan
arXiv:2608. 09805v1 Announce Type: cross Abstract: Exploration has been a focus of reinforcement learning research for a long time.
By Vatsal Venkatkrishna, Nico Daheim, Iryna Gurevych
arXiv:2610.15840v1 Announce Type: cross
Abstract: Reinforcement learning (RL) for large language model (LLM) agents remains challenging in long-horizon tasks with sparse outcome rewards, where assign...
By Kaijie Zhu, Weixiang Yan, Omar Attia, Abbas Kazerouni, Alexander Toshev, Andrew Szot
arXiv:2605. 17333v2 Announce Type: replace Abstract: Reinforcement Learning from Verifiable Rewards (RLVR) typically samples multiple responses per prompt and assigns binary rewards based on individual correctness, yet the collective structure of the group output, specifically the distribution of errors, is largely discarded.
By Wenpu Liu, Yuqi Xu, Weichu Xie, Yongfu Zhu, Shuai Dong, Ziyue Wang, Wenqi Shao, Xiaoying Zhang, Tong Yang, Nan Duan, Jiaqi Wang
arXiv:2606. 15333v1 Announce Type: cross Abstract: LLM unlearning has emerged as a cost-effective alternative to full retraining for removing hazardous knowledge from pretrained models while preserving general utility.
By Zirui Pang, Chenlong Zhang, Haosheng Tan, Zhuoran Jin, Jiaheng Wei, Zixin Zhong
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
The paper introduces Self‑Sparse Speculative Rollouts (S$^3$R), a method that uses sparse attention to propose tokens and then verifies them with full attention, thereby reducing KV‑cache traffic while preserving the dense‑policy distribution for reinforcement learning rollouts. S$^3$R eliminates the need for a separate drafter by sharing policy parameters between drafting and verification, and a KV‑aware scheduler limits cache thrashing to maintain parallelism. Experiments on long‑context mathematical reasoning with the Qwen3 model show that S$^3$R achieves 1.35–1.39× higher end‑to‑end training throughput with comparable task performance.
By Yuhang He, Hongyu Ge, Zihan Lin, Siyi Liu, Yongqi Zhang
arXiv:2605. 20256v2 Announce Type: replace Abstract: Reinforcement learning has become a cornerstone for aligning and unlocking the reasoning capabilities of large-scale models.
By Xikai Zhang, Yongzhi Li, Likang Xiao, Yingze Zhang, Yanhua Cheng, Quan Chen, Peng Jiang, Wenjun Wu, Liu Liu