Chain-of-Thought Prompting Elicits Reasoning in Large Language Models
2022 — linear step-by-step LLM reasoning (NeurIPS)
Chain-of-thought prompting elicits multi-step reasoning in large language models — the linear (path-graph) topology.
Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E., Le, Q., & Zhou, D. (2022). Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. Advances in Neural Information Processing Systems 35 (NeurIPS 2022); arXiv:2201.11903.
Why we cite it
Names the linear chain of thought (CoT) topology used in LLM prompting — one legible step after another. Our corpus uses CoT for the same shape on the event log (L-β assert), with batching and envelope fields defined in Thinking by Writing — not as endorsement of any particular model or benchmark result.
Claims we use
- Topology — CoT is a path graph of intermediate “thoughts” before the final answer.
- Legibility — surfacing steps aids audit and error localization (we extend this to tier-one capture and harness feedback).
Author (primary)
Jason Wei — full author list in the original paper.
Corpus stance
Context — external prompting method; compare Reasoning Topology for tree/graph variants and event-sourced broadcast.