
Chain-of-thought: getting models to reason step by step
In short: Chain-of-thought prompting asks a model to write intermediate steps before answering, which helps multi-step reasoning but does not prove the answer.
Chain-of-thought (CoT) is a prompting technique where a language model writes out intermediate reasoning steps before giving its final answer. A January 2022 paper by Wei and colleagues showed that a few worked examples with reasoning improved large models on arithmetic, commonsense and symbolic reasoning tasks. Reasoning models now produce this kind of working on their own, but the idea still shapes how we prompt and evaluate them.
Chain-of-thought in plain words
It is showing your working in a maths exam. Writing each step makes you less likely to jump to a wrong answer, and makes mistakes easier to spot.
How it works
Each token a model generates becomes context for the next. Writing out steps gives the model a scaffold to build on before it commits to an answer. You can trigger it with examples that include reasoning, as the original paper did with eight exemplars, with an instruction to reason step by step, or by using a reasoning model that thinks before replying.
Example
As an illustration, take the question: "A plan costs ₹800 a month with 20% off for annual billing. What does a year cost?"
Without CoT, a model may answer "₹9,600". With CoT it writes: 800 × 12 = 9,600; 20% of 9,600 is 1,920; 9,600 minus 1,920 is 7,680. Answer: ₹7,680.
Where you'll meet it
- Reasoning models with "thinking" or extended reasoning settings
- Prompts for maths, planning, code debugging and analysis
- Agent loops, where reasoning precedes each action
- Evaluation, where reasoning traces help diagnose failures
What it changes for you
For developers, reasoning can improve accuracy on hard tasks but costs tokens and latency, so turn it up only where errors are expensive. For product teams, visible reasoning can build user trust, but it is not a guaranteed faithful account of how the model decided.
Chain-of-thought vs reasoning model
Chain-of-thought is a prompting technique. A reasoning model is trained to produce extended reasoning on its own, usually hidden or summarised.
Common mistake
My view: treating the written reasoning as proof the answer is right. Turpin and colleagues showed in May 2023 that chain-of-thought explanations can systematically misrepresent the true reason for a model's prediction. Check the answer against tests or sources, however convincing the steps sound.
Related terms
Reasoning model · Prompt engineering · Agent loop · Few-shot prompting
From the archive: Put reasoning at the workflow's uncertain decisions
Check the answer against tests or sources, however convincing the written reasoning sounds.


