Скачать или смотреть Learning How to Think: Meta Chain-of-Thought (Meta-CoT)

The primary source proposes Meta Chain-of-Thought (Meta-CoT), an extension of standard Chain-of-Thought designed to improve large language model (LLM) reasoning by explicitly modeling the underlying thought process, effectively working toward System 2 reasoning. The text explores empirical evidence suggesting state-of-the-art models exhibit behaviors consistent with in-context search and backtracking, especially in complex mathematical reasoning problems where simple step-by-step methods fail. Key methodologies discussed for achieving Meta-CoT capabilities include using Process Reward Models (PRMs) for intermediate step supervision, synthetic data generation, and Reinforcement Learning with Execution Feedback (RLEF) to enhance search efficiency and capability.