Reasoning that Travels: Dissecting How Chain-of-Thought Transfers Across Models

In three linesStudy of chain-of-thought (CoT) transfer across models using a provider-receiver framework. Full traces often transfer successfully, but mechanisms vary: answer extraction (AIME), receiver competence (MMLU-Pro), or partial structured information (ZebraLogic). In free-generation mode, partial CoTs improve performance, suggesting guidance for continued reasoning.

Summary generated by Claude — human-verified