New Paper Finds LLM Self-Explanations Predict Behavior

A summary of a new paper argues existing faithfulness metrics are unsuitable for evaluating frontier LLMs and introduces a new metric; the authors report that LLM self-explanations help predict model behavior.