一作ACL 2026Findings· Findings of ACL
Where C o T Reasoning Commits: Entropy Traces Identify Interpretable Attention Heads
Tianhe Zhang · Yonghong Deng · Ping Jian · Zhen Yang · Boyang Wang · Xinyue Zhang
摘要
While LLMs demonstrate impressive reasoning capabilities, their internal decision dynamics remain opaque. To render these process interpretable and intervenable, we propose Dynamic Entropy Tracing, a mechanism-aware framework that interprets the evolving “choice state” of attention heads during C...