chain of thought (CoT)

Chain-of-Thought (CoT) prompting is a technique for large language models (LLMs) of sufficient scale in which the prompt is structured to encourage the model to generate intermediate reasoning steps (a chain of thought) before producing a final answer to a multi-step or complex problem.

The prompt typically includes either an instruction like Let’s think step by step or few-shot demonstration examples that explicitly show reasoning steps leading to the answer.

A notable variation, self-consistency, enhances CoT by sampling multiple reasoning paths and then selecting the answer that’s most consistent across those paths.

A chain of thought isn’t a guaranteed window into how a model reached its answer. Faithfulness research from Anthropic has found that models often rely on cues they never mention in their visible reasoning, so the printed steps are a useful intermediate output rather than a verified explanation.

Reasoning models internalize this behavior. They’re trained with reinforcement learning to produce a chain of thought before answering, so vendors now advise against explicit CoT prompting with them. OpenAI’s guidance recommends avoiding think step by step instructions outright, and Anthropic treats manual CoT as a fallback for when extended thinking is off.