Review Chain of Thought

Chain of thought is a term for intermediate reasoning steps used or elicited while solving a multi-step problem, whether those steps are internal to a model or presented as an explanation.

Breaking a problem into smaller steps can improve some results. A system may reason internally, provide a concise rationale, or generate a visible step-by-step explanation depending on its design and settings.

Visible reasoning can help a person inspect an answer, but it should not be treated as a guaranteed transcript of the model's internal process or as proof that the conclusion is correct.

A system solves a planning problem internally and returns the recommendation with a short list of assumptions, checks, and decisive factors for human review.

A fluent step-by-step explanation is not automatically faithful, complete, or correct. Verification still requires evidence or an independent method.

Reasoning text can aid review without proving the reasoning is faithful.

* [[reasoning-model|Reasoning Model]] * [[verification|Verification]] * [[confidence-vs-correctness|Confidence vs. Correctness]] * [[prompt-engineering|Prompt Engineering]]

* [Wei et al., *Chain-of-Thought Prompting Elicits Reasoning in Large Language Models* (NeurIPS 2022)](https://papers.neurips.cc/paper_files/paper/2022/file/9d5609613524ecf4f15af0f7b31abca4-Paper-Conference.pdf) — primary paper introducing chain-of-thought prompting and reporting its measured effects. * [Google Research, *Towards Understanding Chain-of-Thought Prompting: An Empirical Study of What Matters*](https://research.google/pubs/towards-understanding-chain-of-thought-prompting-an-empirical-study-of-what-matters/) — empirical analysis of which properties of generated reasoning steps affect performance.

* Pair any rationale with [[verification|Verification]] rather than treating explanation as evidence.