How does chain-of-thought prompting work?
Two variants, same mechanism. Few-shot: the prompt includes a handful of worked examples whose solutions show their steps, and the model imitates the pattern. Zero-shot: the prompt just appends an instruction like “let’s think step by step.” Either way, the model’s answer is conditioned on reasoning it has already generated, which keeps multi-step problems from being solved in one unsupported leap. It is one of the foundational moves of prompt engineering.
What did the research actually show?
The 2022 paper by Wei et al. at Google, Chain-of-Thought Prompting Elicits Reasoning in Large Language Models, showed that eight worked examples pushed a 540-billion-parameter model to then state-of-the-art accuracy on the GSM8K math benchmark (Wei et al., 2022, read 12 October 2026). The effect scaled with model size: small models gained little, large models gained a lot, an early clue that spending more tokens on thinking buys accuracy, the idea that became test-time compute.
Is chain-of-thought prompting obsolete now?
As a trick, mostly; as an idea, it won. Reasoning models internalized the technique: they generate extended internal reasoning without being asked, and the dial moved from prompt wording to an API parameter, reasoning effort. Chain-of-thought stopped being a prompt hack the day models started doing it unprompted. Explicit CoT still earns its keep when you need the steps visible and in your format: a calculation a reviewer must check, or a decision trail kept for audit.
What are the limits of chain-of-thought?
The written reasoning is an output, not a window into the model’s actual computation: a model can produce a plausible chain and a wrong answer, or a correct answer with a confabulated chain. Treat the steps as checkable work, not as proof, and verify the ones that matter, because a fluent chain of thought can dress up a hallucination very convincingly.