Comment by ursAxZA

I might be missing something here as a non-expert, but isn’t chain-of-thought essentially asking the model to narrate what it’s “thinking,” and then monitoring that narration?

That feels closer to injecting a self-report step than observing internal reasoning.