Training
Chain-of-Thought
A technique where a language model writes out a series of intermediate reasoning steps before its final answer, sharply improving accuracy on complex, multi-step tasks.
Chain-of-Thought (CoT) is a technique in which a large language model generates a series of intermediate reasoning steps before its final answer — effectively “thinking out loud.” Instead of emitting a solution immediately, the model breaks a problem into smaller steps, which sharply improves accuracy on arithmetic, logical, and multi-step tasks.
The technique was described in 2022 by Google Brain researchers (Wei et al.). In its original form it is a kind of in-context learning: the prompt includes worked examples that show the reasoning steps. Researchers later found that simply appending “let’s think step by step” elicits zero-shot CoT without any examples at all.
Chain-of-thought is the foundation of modern reasoning models such as OpenAI’s o1/o3 series, DeepSeek R1, and Claude with extended thinking, which during 2024-2026 turned CoT from a prompting trick into a trained property of the model. This opened a new scaling axis — more compute at inference in exchange for accuracy.