Chain-of-thought prompting is described

Jason Wei, Denny Zhou and colleagues at Google described chain-of-thought prompting, a technique in which a handful of worked examples in a prompt — showing the intermediate reasoning steps that lead to an answer, not just the answer itself — substantially improved a large language model’s performance on tasks requiring multi-step reasoning. No retraining or fine-tuning was involved: the same frozen model, given a different style of example, reasoned better.

The paper’s central result used PaLM, Google’s 540-billion-parameter model: prompting it with just eight chain-of-thought exemplars reached state-of-the-art accuracy on GSM8K, a benchmark of grade-school math word problems, surpassing even a fine-tuned GPT-3 model augmented with an external verifier. The authors reported similar gains across arithmetic, commonsense and symbolic reasoning benchmarks, and argued that the ability to benefit from chain-of-thought reasoning emerged only in sufficiently large models — smaller models saw little or no improvement, and sometimes did worse.

The finding mattered because it was a capability elicited by prompting rather than by scale or training data alone, and because it suggested reasoning ability was latent in large models rather than absent from them. It reframed the question from “can the model solve this” to “can the model be asked in a way that lets it solve this,” and it set a template — externalising intermediate steps before producing a final answer — that later labs built directly into training. The reasoning-focused models that followed, trained explicitly to produce long chains of reasoning before answering rather than relying on prompting tricks to elicit them, trace their lineage to this observation.