LLM Assertiveness can be Mechanistically Decomposed into Emotional and Logical Components

Hikaru Tsujimura, Arush Tagade

Large Language Models (LLMs) often display overconfidence, presenting information with unwarranted certainty in high-stakes contexts. We investigate the internal basis of this beha…