Meta Chain-of-Thought is a class of techniques that operates at a meta-level above standard Chain-of-Thought prompting: instead of directly supplying example stepwise solutions, it constructs, learns, or selects abstractions, heuristics, or controllers that shape how an LLM generates intermediate reasoning traces. Practically, Meta-CoT can take the form of learned prompt templates that summarize reasoning strategies, retrieval-and-selection systems that choose diverse exemplar chains optimized for a given instance, or small meta-models that propose which decomposition or inference pattern to apply. The approach formalizes reasoning generation as a two-stage process — a meta-policy over reasoning trajectories followed by the generation of a concrete chain — linking ideas from meta-learning, prompt engineering, and latent-variable modeling of reasoning paths. In application, Meta-CoT improves generalization across problem types, reduces the need for hand-crafted exemplar design, mitigates brittle failures of single-chain outputs (when combined with ensembling or self-consistency), and scales to complex multi-step tasks such as mathematical problem solving, program synthesis, planning, and compositional question answering.
First documented uses of meta-level prompting and controller-style selection for reasoning began appearing in the research literature and community experiments around 2022–2023, with the "Meta-CoT" framing and more systematic empirical studies gaining traction through 2023–2024 as large LLMs and Chain-of-Thought methods demonstrated their value for downstream reasoning tasks.
Key contributors include the researchers who established Chain-of-Thought techniques (e.g., Wei et al., 2022) and subsequent zero-shot and prompting advances (e.g., Kojima et al., 2022), along with a broad set of teams advancing meta-learning, prompt selection, and reasoning robustness — notably groups at Google Research, OpenAI, and academic labs exploring meta-prompting, exemplar selection, and controller/meta-policy designs that together shaped the practical and theoretical development of Meta-CoT.