Tag
This paper investigates why chain-of-thought prompting improves language model accuracy at probe time, finding that gains arise primarily from local token co-occurrence and lexical activation rather than global logical derivation.