The paper introduces ModeBench, a benchmark for multi-solution tasks that evaluates both correctness and the diversity of solutions found by a language model. It shows that reinforcement learning with verifiable rewards (RLVR) tends to collapse solution modes, concentrating probability on fewer correct answers even while accuracy remains high. To counter this, the authors propose Re:Max, a replay-based method that stores one verified example per discovered mode and trains uniformly on these stored modes, improving both success rates and solution diversity across various model scales and RL objectives.
By Liv G. d'Aliberti, Marwa Abdulhai, Sofiia Druchyna, Peter Henderson, Manoel Horta Ribeiro
arXiv:2510. 11686v2 Announce Type: replace-cross Abstract: Reinforcement learning (RL) promises to expand the capabilities of language models, but it is unclear if current RL techniques promote the discovery of novel behaviors, or simply sharpen those already present in the base model.
By Jens Tuyls, Dylan J. Foster, Akshay Krishnamurthy, Jordan T. Ash
arXiv:2510. 21978v2 Announce Type: replace-cross Abstract: Reinforcement learning with verifiable rewards (RLVR) has delivered impressive gains in mathematical and multimodal reasoning and has become a standard post-training paradigm for contemporary language and vision-language models.
By Hoang Phan, Xianjun Yang, Yuanshun Yao, Jingyu Zhang, Shengjie Bi, Xiaocheng Tang, Madian Khabsa, Lijuan Liu, Deren Lei
The paper introduces Stable-MM-R1, a framework that stabilizes reinforcement learning for multimodal reasoning by addressing training instability and entropy collapse. It proposes Potential‑Aware Query Mining (PAQM) to filter data toward high‑potential samples and Hybrid Stratified Replay (HSR) to restructure batches using path entropy and reward stratification, reusing stability anchors and hard negatives. The method demonstrates superior performance on complex reasoning tasks compared to strong baselines.
By Yimeng Ye, Shuang Chen, Wenxuan Huang, Manyuan Zhang, Kaituo Feng, Zhangquan Chen, Jiayu Chen, Yucheng Zhou, Yicheng Xiao, Zhiyuan Feng, Tianyu Shi
arXiv:2606. 24994v1 Announce Type: new Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) for language-model reasoning can fail at both extremes of task difficulty: easy prompts often produce all-correct, low-diversity rollout groups with little gradient signal, while hard prompts can produce all-incorrect groups with no positive reward.
By Wenyang Hu, Junxiang Jia, Zhen Shu, Daniel Dahlmeier, See-Kiong Ng, Bryan Kian Hsiang Low
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:2609.14896v1 Announce Type: cross
Abstract: A notable byproduct of LLM alignment training is mode collapse: the progressive loss of output diversity that narrows a model's expressivity at infer...
By Jiayi Yuan, Hangoo Kang, James Jihao Liu, Yejin Choi, Vikram Iyer, Liwei Jiang, Natasha Jaques
arXiv:2605. 01248v3 Announce Type: replace Abstract: Reinforcement learning (RL) post-training has enabled newer capabilities in models, such as agentic tool-use for search.
By Harsh Goel, Akhil Udathu, Susmija Jabbireddy, Pradnesh Kalkar, Atharva Parulekar
arXiv:2601. 09085v2 Announce Type: replace-cross Abstract: Group Relative Policy Optimization (GRPO) has become a standard approach for training mathematical reasoning models; however, its reliance on multiple completions per prompt makes training computationally expensive.
By Kangda Wei, Ruihong Huang
The paper introduces POISE, a reinforcement learning algorithm that uses a model’s internal states as a value estimator to reduce variance in reinforcement learning with verifiable rewards (RLVR). By employing a lightweight probe that reads internal signals during the forward pass, POISE predicts baselines online and uses a cross‑rollout construction to keep gradients unbiased. Experiments on Qwen3‑4B and OLMo3‑7B‑Instruct‑DPO across six domains show POISE outperforms existing RLVR baselines, offering more stable training and a value model that generalizes across tasks and scales with the policy.
By Yunho Choi, Jongwon Lim, Woojin Ahn, Minjae Oh, Jeonghoon Shim, Yohan Jo
arXiv:2603.16157v2 Announce Type: replace-cross
Abstract: While Reinforcement Learning (RL) enhances Large Language Model reasoning, on-policy algorithms like GRPO are sample-inefficient as they disc...
By Long Li, Zhijian Zhou, Tianyi Wang, Weidi Xu, Zuming Huang, Wei Chu, Zhe Wang, Shirui Pan, Chao Qu, Yuan Qi
arXiv:2505. 09655v5 Announce Type: replace-cross Abstract: Post-training LLMs with Reinforcement Learning, specifically Group Relative Policy Optimization (GRPO), has emerged as a paradigm for enhancing mathematical reasoning.
By Xiwen Chen, Wenhui Zhu, Peijie Qiu, Xuanzhao Dong, Hao Wang, Haiyu Wu, Huayu Li, Aristeidis Sotiras, Yalin Wang, Abolfazl Razi