Chain-of-Meta-Thought Separates Abstract Strategy from Problem Execution in LLM Training

CoMT splits post-training into two stages mirroring human cognition: first learning abstract meta-strategies via supervised fine-tuning, then applying them to specific problems via RL. The approach targets the entanglement of generalizable reasoning patterns with problem-specific execution that plagues standard SFT+RL pipelines.