Chain-of-Thought
Getting a language model to write out intermediate reasoning steps before its final answer, which markedly improves results on math and logic problems.
Also CoT · step-by-step reasoning · let's think step by step · scratchpad
Wei et al., 2022
If you solve 17 × 24 in your head you may slip up. Write the steps on paper and you get it right. Letting the model 'write on paper' helps it the same way.
Where it fits
+All relationships as a list
How it works
An LLM produces each token with a fixed amount of computation. If a question needs several reasoning steps, jumping straight to the answer leaves too little room to compute it.
Chain-of-thought prompting asks the model to show its work:
Q: A cafe had 23 apples. It used 20 for lunch and bought 6 more. How many does it have? A: It started with 23 and used 20, leaving 3. Then it bought 6, so 3 + 6 = 9. The answer is 9.
Two ways to trigger it:
- Few-shot: include worked examples with reasoning in the prompt (in-context learning).
- Zero-shot: just add “Let’s think step by step.”
Each step the model writes becomes part of its context, so later steps can build on earlier ones, like using scratch paper.
A caution: a model’s written reasoning is not necessarily a faithful account of what drives its answer, which is an active topic in interpretability and safety. Building on this idea, reasoning models are trained to think at length before answering.
Everyday use cases
- Case 01
Math word problems
Asking for the steps before the answer greatly reduces arithmetic and logic slips.
- Case 02
Debugging
Having a model walk through code line by line before proposing a fix.
Strengths, weaknesses and when to use it
Strengths
- Large accuracy gains from a small prompt change; the steps make errors easier to spot.
Weaknesses
- More tokens means slower and costlier answers.
- The written reasoning is not guaranteed to reflect how the model actually reached its answer.
When to use itMulti-step problems (math, planning, logic). Less useful for simple lookups.
Compare with
Check yourself
- Q01
Why does writing out steps help an LLM?
Show answer
A: Each generated step becomes context the model can build on, giving it more computation per answer