Chain of thought: what it is, explained simply

In plain words

Try this in your head: 3 boxes of 12 mangoes, 20 are sold, how many are left? Most of us quietly do it in two steps: 36 first, then 16.

A language model writes one word at a time, and each word gets the same small amount of computing. If it has to blurt out the answer as its very first word, it must do all the working in that single step, and it often slips.

Chain of thought gives it room. When the model writes "3 × 12 = 36" first, that line is now part of the text it reads, so the next step is easy. Each step becomes a link in a chain that leads to the answer.

In 2022, researchers found that simply adding "Let's think step by step" to a question sharply improved results. Today's reasoning models are trained to do this on their own.

Everyday example

Working out a bill at the kirana store

The shopkeeper does not stare at a pile of items and announce a total. He writes each price down the side of a paper bag, then adds the column. If someone interrupts, he can look at what he has written and carry on. The model's written steps are that column on the paper bag.

Where the example breaks: The shopkeeper's column is exactly how he got the total. A model's written steps can look neat while the answer came from somewhere else, so the chain is helpful but not a guaranteed explanation.

How it actually works

Chain of thought began as a prompting technique and became a core part of how models are trained.

Few-shot chain of thought

In January 2022, Google researchers put a few worked examples, with their steps written out, into the prompt. The model copied the style and wrote steps for new problems too. A 540 billion parameter model prompted this way set a new best score on a test of school maths word problems.

Zero-shot: one magic line

Months later, a University of Tokyo and Google team showed that no examples were needed. Adding "Let's think step by step" lifted one model from 17.7% to 78.7% on an arithmetic test, and from 10.4% to 40.7% on harder maths word problems.

Why it works

Each token the model writes is fed back in, so written steps act as working memory and give it more computation before it commits to an answer. The early papers found the benefit mostly in large models; small ones wrote fluent but wrong chains.

Built into reasoning models

Since 2024, reasoning models are trained with reinforcement learning to produce long chains by themselves, often hidden from the user. Telling them to think step by step adds little.

Faithfulness

A chain can be convincing without being the true cause of the answer. In 2025, Anthropic found that when given a hint, Claude 3.7 Sonnet mentioned using it only 25% of the time, and DeepSeek-R1 39%.

Named

January 2022, Jason Wei and colleagues at Google, "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models"

Definition

"a series of intermediate reasoning steps"

Famous phrase

"Let's think step by step" (Kojima et al., May 2022)

Abbreviation

CoT

Questions people ask

What is chain-of-thought prompting?

Asking a model to show its working before the answer, either by including worked examples in the prompt or by adding a line like "think step by step". It helps most on maths, logic and multi-step questions.

Is chain of thought the same as a reasoning model?

No. Chain of thought is the technique of writing out steps. A reasoning model is a model trained to do this automatically and well, usually at length and often hidden. Reasoning models are chain of thought, built in.

Does "think step by step" still work?

With older or smaller chat models, it can still help on multi-step problems. Reasoning models already think before answering, so the phrase adds little, and OpenAI advises against it, noting it can sometimes even hurt.

Can we trust the steps a model shows?

Partly. They are often useful for spotting errors. But studies show the written chain does not always reflect what actually drove the answer, so the steps are a helpful trace, not a proof.