Chain of Thought

Chain of thought (CoT) is a prompting technique that asks a language model to write out intermediate reasoning steps before giving a final answer. Instead of jumping straight to a result, the model produces a sequence of smaller statements, each conditioned on the ones before it. Because a transformer generates one token at a time, those extra tokens give the model more computation to spend on a hard problem, and they expose the path it took. Accuracy on multi-step arithmetic, logic puzzles, and word problems improves noticeably compared with answering directly. In practice a prompt may simply instruct the model to work through the problem step by step, or include worked examples that demonstrate the reasoning format. Recent reasoning models from OpenAI, Anthropic, and Google build this behavior in, generating an internal reasoning pass before the visible reply and trading latency and token cost for reliability. A common pitfall is treating the written reasoning as a faithful record of how the model actually reached its answer. The text can look sound while carrying an early error forward through every later step, so critical outputs still need verification against a tool or an external source. CoT pairs well with retrieval, which supplies the facts the reasoning operates on.