作者位次 #4/12ACL 2026Findings· Findings of ACL
SPS : Steering Probability Squeezing for Better Exploration in Reinforcement Learning for Large Language Models
Yifu Huo · Chenglong Wang · Ziming Zhu · Shunjie Xing · Peinan Feng · Tongran Liu · Qiaozhi He · Tian Hua Zhou · Changxiaojia · JingBo Zhu · Zhengtao Yu · Tong Xiao
摘要
Reinforcement learning (RL) has emerged as a promising paradigm for training reasoning-oriented models by leveraging rule-based reward signals. However, RL training typically tends to improve single-sample success rates (i.e., Pass@1) while offering limited exploration of diverse reasoning trajec...