Chain of Thought (CoT)

A prompting strategy that encourages an LLM to generate a step-by-step reasoning process before providing its final answer, significantly improving performance on complex tasks.

How it works

Chain of Thought prompting adds phrases like 'Let's think step by step' or provides worked examples showing intermediate reasoning before the final answer. This encourages the model to generate a sequence of reasoning steps before committing to an answer. The intermediate steps serve two functions: they guide the model's own computation, and they produce an output that can be checked for logical consistency. More advanced variants include tree-of-thought (branching exploration) and self-consistency (sampling multiple chains and majority-voting).

Why it matters

CoT prompting unlocked multi-step reasoning in LLMs that previously failed at arithmetic, logic puzzles, and complex question answering. A model that scores 20% on a hard maths benchmark with direct prompting might score 60%+ with CoT. It is a foundational technique in the 'reasoning models' movement: DeepSeek-R1 and OpenAI's o1 essentially automate CoT by training models to generate reasoning traces before answering, making CoT an internalized behaviour rather than a prompting trick.