作者位次 #5/7ACL 2026Findings· Findings of ACL
Unlocking Exploration in RLVR : Uncertainty-aware Advantage Shaping for Deeper Reasoning
Can Xie · Ruotong Pan · Xiangyu Wu · Zhang yunfei · Jiayi Fu · Tingting Gao · Guorui Zhou
摘要
Reinforcement Learning with Verifiable Rewards (RLVR) has shown significant promise for enhancing the reasoning capabilities of large language models (LLMs). However, prevailing algorithms like GRPO broadcast a uniform advantage signal across all tokens in a sequence. This coarse-grained approach...