Moyan AI Training Institution LogoMoyan AI

Prompting · Intermediate

Chain-of-Thought Prompting

Also known as: CoT, Chain-of-Thought (CoT) Prompting

Chain-of-thought prompting supplies examples of intermediate reasoning steps alongside answers. The original paper studies how this changes performance on selected reasoning tasks.

What Chain-of-Thought Prompting is

Wei et al. study prompting with reasoning demonstrations on arithmetic, commonsense and symbolic tasks. Their reported gains depend on the models and evaluation setup; they are not a universal accuracy promise.

How it works

A demonstration contains a problem, intermediate steps and an answer. A new problem follows those demonstrations. This prompting method does not itself train new model weights or verify that the generated steps are true.

Why it matters

Intermediate steps can give a reader something to check. A fluent explanation can still contain a mistaken assumption or calculation, so compare the answer with independent evidence.

Common uses

  • →Study how worked demonstrations affect a narrowly defined task
  • →Check intermediate calculations against an answer key

Strengths

  • ✓Reported benefits in the original paper's evaluated settings

Watch for

  • ✓Not a guarantee of correct reasoning
  • ✓Current model behavior and cost have not been measured here

Proposed arithmetic check

Fictional exercise: three boxes each contain four pencils; two pencils are removed. Keep the reference answer, 10, separate from the prompt.

  1. Compare a direct-answer prompt with a prompt containing a separate worked demonstration.
  2. Check 3 × 4 − 2 = 10 independently. Record both wrong and right answers.
  3. Do not infer broad reasoning ability from this one simple question.

The arithmetic reference is 10 pencils. No model answers, gains or timings are reported: this comparison has not been run.

Continue exploring

More in this collection

Browse all AI Concepts

Sources & References

Wei et al. — Chain-of-Thought Prompting Elicits Reasoning in Large Language Models (2023 revision)