What is chain of thought? Getting an AI to write out its steps one by one before it gives the answer

Getting an AI to write out its steps one by one before it gives the answer.

Chain of thought (CoT) means having an AI write out its in-between steps before the final answer. Each step supports the next, so it makes fewer mistakes on problems that need several steps.

In 2022, a Google team showed that adding a few worked examples to a prompt sharply raised scores on math word problems. Another 2022 study found that simply adding "Let's think step by step" helped too.

Reasoning models are trained to do this on their own. But research suggests the written steps don't always show how the model really reached its answer.

Think of it like this

It's your teacher saying, "Show your work." Writing each step makes mistakes easier to catch, and you're less likely to skip something important.

For example

Ending a question with "Think it through step by step" is the simplest chain-of-thought prompt.

Common mix-ups

Mix-upTo get step-by-step working, you have to retrain the model.
ActuallyIn a 2022 Google study, just eight worked examples were enough for a 540-billion-parameter model to set a record on GSM8K, a test of primary-school maths word problems. The effect was clear in big models, though, and was reported to show up poorly in small ones.
Mix-upThe working an AI writes shows exactly how it reached the answer.
ActuallyWhen researchers secretly slipped in a hint pointing to one answer, models would follow the hint yet write working that never mentioned it, giving some other plausible reason instead.

Try it in class

For teachers at the start of a lesson — or for students working on their own.

Question: Why does writing out the working reduce mistakes on multi-step problems?
Activity: Pick one multi-step maths word problem and ask an AI twice — once plainly, once with "Solve it step by step" added. Put the two answers side by side, and in groups re-check every step yourselves, marking each one right or wrong.

Records about this

Source-checked moments from AI history.

May 24, 2022Takeshi Kojima et al. (University of Tokyo, Google Research)

Kojima et al. present Large Language Models are Zero-Shot Reasoners

CoT without examples. The original of every later zero-shot reasoning prompt

Prompt EngineeringPaper

Jan 28, 2022Jason Wei et al. (Google Brain)

Chain-of-Thought Prompting Elicits Reasoning in Large Language Models

The starting point of the entire lineage of changing reasoning performance through prompts alone. With 8 CoT exemplars, the 540B model reached SOTA on GSM8K

Prompt EngineeringPaper