作者位次 #4/8ACL 2026Findings· Findings of ACL
Revisiting Entropy Regularization: Adaptive Coefficient Unlocks Its Potential for LLM Reinforcement Learning
Xiaoyun Zhang · Xiaojian Yuan · Di Huang · Wang You · Chen Hu · Jingqing Ruan · Kejiang Chen · Xing Hu
摘要
Reasoning ability has become a defining capability of Large Language Models (LLMs), with Reinforcement Learning with Verifiable Rewards (RLVR) emerging as a key paradigm to enhance it. However, RLVR training often suffers from policy entropy collapse, where the policy becomes overly deterministic...