How it works

The paper analyzes the gap between visible chain-of-thought text and underlying model mechanics, arguing surface outputs are an unreliable proxy.