principles.fyi · the brain · concept

chain of thought

Having the model write out its reasoning steps before giving a final answer.

prompt -> reasoning steps -> final answer

Chain-of-thought is prompting (or training) a model to produce intermediate reasoning — working through the problem step by step — before committing to its final answer. It helps on hard reasoning tasks because each generated step becomes part of the context the next step conditions on, so the model effectively gives itself room to compute rather than guessing in one leap. It is the canonical form of test-time compute: the extra reasoning tokens are exactly the extra computation spent at answer-time.