SCARL (Sequence Confidence-Aligned Reinforcement Learning) addresses a misalignment between a reinforcement learning policy’s confidence and its reward signal in language-model reasoning. By reshaping each advantage with an alignment gap that reflects the difference between a response’s group-relative confidence and the outcome it certifies, SCARL drives responses with equal advantage toward equal confidence and mitigates entropy collapse. Experiments on Qwen3‑4B‑Base show a 2.4‑point lift in Pass@32 over DAPO and consistent improvements across multiple estimators and backbones while preserving policy entropy.
By Zhenghong Ke, Jingbo Zhou, Hua Wu
arXiv:2609.11149v3 Announce Type: replace-cross
Abstract: How fast does a language model degrade when trained on its own outputs? Theory traces it to gradually accumulating errors, while experiments...
By Yangze Liu, Zhongyi Han
arXiv:2606. 15455v1 Announce Type: cross Abstract: Reinforcement learning with verifiable rewards (RLVR) has become a key approach for enhancing the reasoning abilities of large language models.
By Suqin Yuan, Jinkun Chen, Jiyang Zheng, Muyang Li, Lei Feng, Dadong Wang, Tao Xiang, Tongliang Liu, Bo An
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
The paper critiques the use of pass@$k$ as a sole metric for evaluating reinforcement‑learning post‑training on language models, showing that it ignores diversity in model outputs. Experiments on Qwen2.5‑1.5B‑Instruct trained with Group Relative Policy Optimization (GRPO) and rejection‑sampling fine‑tuning (RFT) reveal that these methods shift token‑level, answer‑level, and unique‑answer diversity in opposite directions, yet pass@$k$ fails to capture these differences. Even when restricting to verifier‑correct completions, diversity gaps persist, and pass@$k$ does not consistently identify the better performing arm on hard problem subsets, indicating that the metric can mask true capability changes.
By Subham Rath, Raj Dandekar, Rajat Dandekar, Sreedath Panat
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
AEM: Adaptive Entropy Modulation for Multi‑Turn Agentic Reinforcement Learning proposes a supervision‑free credit assignment technique that adjusts entropy dynamics at the response level during RL training. By aligning uncertainty estimation with the granularity of LLM agent actions and rescaling advantages based on a response‑level uncertainty proxy, AEM improves the exploration‑exploitation balance without relying on dense intermediate supervision. Experiments on ALFWorld, WebShop, and SWE‑bench‑Verified show consistent performance gains across models from 1.5B to 32B, including a 1.4% improvement in a state‑of‑the‑art software‑engineering RL framework.
By Haotian Zhao, Songlin Zhou, Yuxin Zhang, Stephen S. -T. Yau, Wenyu Zhang, Lun Tian, Tianshu Zhu, Yifeng Huang, Yucheng Zeng, Jingnan Gu, Daxiang Dong, Jianmin Wu
EliteReplay is a post‑training method that combines GRPO with a quality‑diversity archive to preserve diverse solutions in reinforcement learning. By collecting distinct strategies from a frozen base model before training and replaying them during training, the approach improves accuracy while maintaining a broader set of correct solutions. On text‑to‑SQL tasks, EliteReplay outperforms baselines, achieving higher pass@128 scores and up to 77% more distinct correct strategies, and it better handles ambiguous questions by covering multiple interpretations.
By Hanane Nour Moussa, Doeun Lee, Sachin Kumar
arXiv:2601. 12401v2 Announce Type: replace-cross Abstract: Reinforcement learning (RL) has emerged as a powerful paradigm for fine-tuning large-scale generative models, such as diffusion and flow models, to align with complex human preferences and user-specified tasks.
By Jinmei Liu, Haoru Li, Zhenhong Sun, Chaofeng Chen, Yatao Bian, Bo Wang, Daoyi Dong, Chunlin Chen, Zhi Wang
The paper introduces new methods for deciding which self‑evolving language model updates to keep. HeirCommit freezes each updated model and evaluates it on fresh questions before deciding to inherit it, while EvoHeir separates parameter inheritance from experience inheritance to repair failures and preserve successes. These techniques reduce harmful updates and improve performance on mathematics and general‑reasoning tasks compared to existing loops such as R‑Zero, Agent0, and Skill Self‑Play.
By Haoyi Yang, Kshitij Mishra, Nils Lukas, Salem Lahlou
arXiv:2606. 01561v1 Announce Type: new Abstract: Aligning Large Language Models (LLMs) with human preferences is often formulated via Direct Preference Optimization (DPO).
By Xiwen Chen, Wenhui Zhu, Jingjing Wang, Peijie Qiu, Zhipeng Wang, Huayu Li, ZhengXiao He, Xuanzhao Dong, Prayag Tiwari, Mingkun Xu, Yujian Xiong, Feng Luo, Abolfazl Razi, Brendan Hogan Rappazzo, Anderson Schneider, Yuriy Nevmyvaka
arXiv:2509. 24372v3 Announce Type: replace-cross Abstract: Fine-tuning large language models (LLMs) for downstream tasks is an essential stage of modern AI deployment.
By Xin Qiu, Yulu Gan, Conor F. Hayes, Qiyao Liang, Yinggan Xu, Roberto Dailey, Elliot Meyerson, Babak Hodjat, Risto Miikkulainen