#receiver head analysis

Thought Anchors is a new framework that improves understanding of reasoning processes in large language models by analyzing sentence-level contributions and causal impacts.