Chain-of-Thought
Chain-of-thought is the technique of having a model work through intermediate steps before answering. On counting and logic tasks it measurably improves accuracy compared with answering straight away.
Why it works
A model spends roughly the same computation on every token. Answering immediately forces the whole problem into a single step. Writing out the reasoning gives it more steps for the same problem, each building on the previous one now written down.
The intermediate steps act as external memory. The model does not hold the calculation "in its head"; it sees it in the text in front of it, which cuts arithmetic and logic slips noticeably.
A side benefit is checkability. Seeing the reasoning, you can find where it went wrong instead of guessing from a single final number.
How to use it
- Ask for steps before the conclusion"Work through it step by step, then give the answer" works even on simple models.
- Not for every taskOn translation, summarising and copy-editing, reasoning only lengthens the answer and raises the bill.
- The reasoning can be hiddenAsk it to reason and then output only the conclusion: the quality stays and the answer is short.
- Plausible reasoning can still be wrongWritten-out steps do not guarantee correctness. They make checking possible; they do not replace it.