Summary
Chain-of-thought (CoT): ask the model to reason step by step before answering. On math, logic, and multi-step tasks it improves accuracy by 10-40 percentage points. On simple tasks it wastes tokens. Learning when to use it is the discipline.
Objectives
- 01State the chain-of-thought result.
- 02Recognize tasks where CoT helps versus tasks where it does not.
- 03Distinguish zero-shot CoT from few-shot CoT.
The Lesson
The result
Wei et al. 2022: adding 'Let's think step by step' or few-shot CoT examples improved GSM8K math accuracy from ~18% to ~57% with the same underlying model. The result generalizes to logic, planning, and multi-step reasoning.
When not to use it
Simple classification, lookup, or single-fact extraction — CoT wastes tokens and can degrade accuracy by overthinking. Modern reasoning models internalize CoT during RL; explicit CoT prompts are less useful with them.
Key Ideas
- CoT trades tokens for accuracy on multi-step tasks.
- Not a universal win; unnecessary on simple tasks.
References
- Wei et al., 'Chain-of-Thought Prompting Elicits Reasoning in Large Language Models' (2022)