[论文] Reinforcement Learning without Ground-Truth Solutions can Improve LLMs

论文概要

研究领域: ML
作者: Yingyu Lin, Qiyue Gao, Nikki Lijing Kuang
发布时间: 2026-06-27
arXiv: 2606.27369

中文摘要

使用可验证奖励的强化学习（RLVR）训练LLM通常依赖真实答案来分配奖励。我们引入RiVER框架，使用确定性执行反馈作为连续值监督，在无需真实解的情况下训练LLM，扩展了RLVR的应用范围。

原文摘要

Reinforcement learning with verifiable rewards (RLVR) for training LLMs typically rely on ground-truth answers to assign rewards, limiting their applicability to tasks where the ground-truth solution is unknown. We introduce a \textbf{R}anking-\textbf{i}nduced \textbf{VER}ifiable framework (RiVER) that trains LLMs on score-based optimization tasks without ground-truth solutions, using deterministic execution feedback as continuous-valued supervision. When applying group-relative RL to such conti...

自动采集于 2026-06-27

#论文 #arXiv #ML #小凯