Prompting a model to write out intermediate reasoning steps before its final answer, either by showing worked examples or by adding an instruction such as "Let's think step by step". It improves accuracy on multi-step problems such as arithmetic and logic, at the cost of more output tokens.
Card 1 of 292. Answer: Prompting a model to write out intermediate reasoning steps before its final answer, either by showing worked examples or by adding an instruction such as "Let's think step by step". It improves accuracy on multi-step problems such as arithmetic and logic, at the cost of more output tokens.
About · Quizzes · History of AI · To suggest new questions or report an error send me a dm.
Wei and colleagues at Google introduced it in 2022 with few-shot prompts whose examples included the reasoning, not just the answers. The gains appeared mainly in large models; in small ones the written reasoning was often wrong. Kojima and colleagues then showed the same year that a single sentence, "Let's think step by step", gets much of the effect with no examples, which became known as zero-shot chain of thought.
The numbers that made it famous: on the GSM8K grade-school math benchmark, PaLM 540B went from 17.9% with standard prompting to 56.9% with chain of thought. Self-consistency (Wang and colleagues, 2022), which samples several reasoning chains and takes a majority vote on the final answer, raised it to about 74%. Zero-shot chain of thought took InstructGPT from 17.7% to 78.7% on the MultiArith benchmark.
Follow-ups: reasoning models such as OpenAI's o1 and DeepSeek-R1 are trained with reinforcement learning to produce long chains on their own, so explicit prompts help them less; the written chain is not guaranteed to reflect how the model actually reached its answer (Turpin and colleagues, 2023); and the extra tokens add latency and cost.
