作者位次 #5/7ACL 2026Poster· Natural Language Processing
S tep H int: Multi-level Stepwise Hints Enhance Reinforcement Learning to Reason
Kaiyi Zhang · Ang Lv · Jinpeng Li · Yongbo Wang · Feng Wang · Haoyuan hu · Rui Yan
摘要
Reinforcement learning with verifiable rewards (RLVR) is a promising approach for improving the complex reasoning abilities of large language models (LLMs). However, current RLVR methods face two significant challenges: the near-miss reward problem, where a small mistake can invalidate an otherwi...