What Makes Chain-of-Thought Work at Probe Time? Local Co-occurrence Rather Than Global Derivation

In three linesStudy on chain-of-thought (CoT) mechanisms at probe time. Authors show performance gains arise primarily from lexical activation and short-range token co-occurrence (2-3 tokens), not global logical derivation. Even word-shuffled rationales substantially outperform no-rationale baselines.

Summary generated by Claude — human-verified