作者位次 #3/5ACL 2026Poster· Natural Language Processing
Calibration-Aware Policy Optimization for Reasoning LLM s
Ziqi Wang · Xingzhou Lou · Meiqi Wu · Zhengqi Wen · Junge Zhang
摘要
Group Relative Policy Optimization (GRPO) enhances LLM reasoning but often induces overconfidence, where incorrect responses yield lower perplexity than correct ones, degrading relative calibration as described by the Area Under the Curve (AUC). Existing approaches either yield limited improvemen...