作者位次 #9/11ACL 2026Poster· Natural Language Processing
Escaping the Echo Trap: On Credit Assignment Failure in Multi-turn LLM Self-Reflection
Linxuan Du · Guangquan Xue · Xiaobo Liang · Qipeng Huang · Yuyang Ding · Xinyu Shi · Zhang Yijun · Ji Qi · Wenpeng Zhu · Juntao Li · Min Zhang
摘要
Despite the potential of multi-turn self-reflection to improve LLM reasoning, its effectiveness in practice is severely constrained by a failure mode we term the Echo Trap.Specifically, this phenomenon gives rise to two coupled problems: (1) the model becomes limited by its inherent capabilities...