Chain-of-Thought (CoT) is both a prompt-engineering technique and a mode of generation in which the model writes out the intermediate steps of its procedure before delivering the final answer. The reason is architectural: a language model spends a constant amount of computation on every token it generates, so a complex task cannot be solved all at once in a single step. By spelling the procedure out, the model creates external working memory inside its context window and spreads the computation over several steps. A distinction is drawn between few-shot CoT, where the prompt shows a handful of worked examples together with the reasoning, and zero-shot CoT, where an instruction such as “think step by step” is enough. A layer on top of this is self-consistency: several independent chains are generated and the most frequent conclusion is taken as the result. Newer reasoning models are trained on long chains of deliberation directly, and some of them keep those chains hidden from the user. The fundamental caveat is faithfulness – the chain that gets written out is a plausible verbalization of the procedure, not a documented record of the actual internal computation, so it can lead to a correct conclusion for the wrong reasons, and the other way round.
Try working out 17 × 24 in your head. Most people hesitate and guess. Write it out on paper, though – 17 × 20 is 340, 17 × 4 is 68, together 408 – and you will get it right almost every time. The paper did not make you any smarter; it merely let you break the task into steps instead of holding all of it in your head at once. A model works the same way: give it room to think out loud and it handles harder tasks more reliably. And just as with people, the explanation it finally writes down need not match what was really going through its head.