两篇新的研究论文探讨了改进用于训练推理模型的RLVR（带可验证奖励的强化学习）的方法。第一篇论文介绍了REFT（首个Token多样化的Rollout探索），一种通过关注推理标记后的首个Token来使Rollout多样化的技术，从而在各种模型规模和难度级别上提高了性能。第二篇论文提出了HAPO（事后追溯感知策略优化），该方法通过基于奖励极性和Token熵分解Token更新来分析Token更新，表明持续的推理收益集中在高熵象限，并在数学推理基准测试中取得了有竞争力的结果。 AI
影响 这些论文引入了新颖的技术，通过改进的训练方法来增强LLM的推理能力，有望带来更强大、更有能力的AI系统。
排序理由 该集群包含两篇学术论文，详细介绍了改进LLM训练的新研究方法。
AI 生成摘要 · Google Gemini · 来自 3 个来源。 我们如何撰写摘要 →
arXiv:2605.28295v1 Announce Type: new Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) trains reasoning models without labeled trajectories, relying on grouped rollouts to expose the policy to alternative reasoning paths and a verifier to score them. Rollout divers…
Reinforcement Learning with Verifiable Rewards (RLVR) trains reasoning models without labeled trajectories, relying on grouped rollouts to expose the policy to alternative reasoning paths and a verifier to score them. Rollout diversity has accordingly emerged as a central bottlen…
arXiv:2604.11056v2 Announce Type: replace-cross Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) improves the reasoning ability of Large Language Models (LLMs), but sparse outcome rewards make token-level credit assignment difficult. We study token-level credit as …