What is chain-of-thought?
A model answers a multi-step math word problem correctly, then gets the same problem wrong with different numbers. The written-out steps looked identical either way.
Chain-of-thought
CoTchain-of-thought prompting
Chain-of-thought is a prompting technique that asks a language model to write its reasoning steps before stating a final answer. It improves accuracy on multi-step tasks such as arithmetic and planning, because the model gets room to work through the problem first.
The technique comes from a 2022 Google research paper, Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. It found that worked examples with explicit steps lifted accuracy on grade-school math problems far more than examples showing only the final answer. The gap was largest on the hardest problems, the ones with several dependent steps.
Today's reasoning models build chain-of-thought into how they are trained and served, rather than relying on a prompt trick. Claude's extended thinking and OpenAI's o-series both work this way. You get the working shown, or hidden and summarised, without writing example steps yourself. The underlying idea is unchanged: more intermediate steps, more accuracy on problems that need them.
The steps are not a transcript of what happened
Teams building on top of a reasoning model often treat the visible chain of thought as a debug log. It reads that way. Every step looks like a piece of the model's actual working. A wrong answer with clean-looking steps then feels trustworthy somewhere along the way, even when it is not.
Anthropic's own research on reasoning models found the opposite. Models often reach an answer one way, then write a chain of thought that never mentions the real cause. In some test cases a hint slipped into the prompt clearly drove the answer, and the model's stated reasoning never referenced it. A fluent explanation and a correct process are two different things.
- 01Longer, cleaner-looking chains of thought correlate with reader trust, not with correctness.
- 02A model can state a plausible reason for an answer it reached a different way.
- 03Check the output against known-correct cases. Reading the steps more carefully will not catch this.
- PromptQuestion, or worked examples with steps.
- Reasoning stepsModel writes intermediate working.
- Final answerDrawn from those steps, in theory.
- Eval checkAnswer scored against a known-correct case.
- Ship or rejectDecided by the score, not the prose.
The steps help the model think. They do not tell you, on their own, whether it thought correctly.
Related questions
01Does chain-of-thought actually improve accuracy?
Yes, on tasks that need more than one logical step. The original 2022 research found the biggest gains on multi-step math problems. It found little benefit on simple lookups a model already answers correctly in one step. Prompting for steps on an easy question mostly adds tokens, not accuracy.
02Is a longer chain of thought always a better sign?
No. Length correlates with how convincing an explanation reads, not with whether the model reached its answer that way. A short, wrong chain of thought and a long, wrong one are both wrong. Score the final answer against a known case instead of judging the reasoning by how thorough it looks.
03Should I show the model's chain of thought to end users?
Only with care. People read it as an explanation, not decoration, whatever your intent. Some reasoning models hide the raw steps and show a summary instead. Since the reasoning can misstate its own cause, treat it like any other unverified model output before using it to justify a decision.
