Chain of thought

Background

Chain of thought is a prompting technique that gets a language model to generate intermediate reasoning steps as text before it commits to a final answer, rather than jumping straight to the output. Because a transformer produces one token at a time and each token can attend to everything generated so far, the reasoning it writes out becomes part of its own working context. The model effectively uses the generated tokens as a scratchpad, so a multi-step problem gets decomposed into smaller sub-conclusions that condition the tokens that follow. For a task like a word problem, the model writes 'there are 3 boxes with 4 items each, so 12 items, minus 2 removed, leaves 10' before stating '10', and each stated quantity constrains the next step. You trigger it either with an instruction such as 'think step by step,' with few-shot examples that themselves show worked reasoning, or with models trained to reason by default. It reliably helps on arithmetic, logic, and multi-hop questions where a single forward pass has too little room to compute the answer implicitly. When building software with AI, it matters because it trades tokens and latency for accuracy on hard reasoning, and because the visible trace is inspectable — you can see where the model went wrong. A caveat: the written explanation is not guaranteed to be the true cause of the answer, so treat it as an aid, not a proof.