Getting reasoning out of the model and keeping it out of the answer
Chain-of-thought trades tokens for accuracy on multi-step problems. It helps on arithmetic, logic and planning — and costs you on lookups.
| Technique | How | Best for |
|---|---|---|
| Zero-shot CoT | Append "Think step by step." | A cheap first try on any reasoning task |
| Few-shot CoT | Show worked examples including the reasoning | Domain problems with a house method |
| Structured CoT | Ask for reasoning inside tags, answer outside | Anything whose output is parsed |
| Self-consistency | Sample n times, take the majority answer | High-stakes arithmetic; n× the cost |
| Least-to-most | Ask it to decompose, then solve each part | Problems that nest |
| Task | Why CoT hurts |
|---|---|
| Factual lookup | Reasoning about a fact you either know or do not adds a chance to talk yourself out of it |
| Classification into few labels | The extra tokens mostly rationalise the first instinct |
| Creative writing | Planning aloud flattens voice |
| Latency-critical paths | You are paying real milliseconds for the tokens |
On multi-step reasoning — arithmetic, logic, planning — usually yes. On factual recall and simple classification it often makes them worse, because it gives the model room to talk itself out of a correct first instinct.
Ask for it inside tags and parse them off, or use a model with a native reasoning channel. Never ask it to "think but do not show your thinking" — that just moves the reasoning into the answer.