AI Trend Research

← AI glossary

AI glossary · Basic · Language models

CoT

CoT (Chain-of-Thought) is an approach in which an AI model does not answer immediately but reaches the answer by writing out its intermediate reasoning step by step.

About this term
Level
Basic · Technical terms you often meet in AI news
Added
2026-10-08
In plain words It is similar to writing out your working step by step on a math test instead of doing it in your head and writing only the answer. However, there is no guarantee that the reasoning the model writes is always the same as the process that actually determined the answer inside the model.

How it works

An LLM writes text one token at a time, choosing the next token based on what it has already written. So if it writes out intermediate steps first, that content becomes a stepping stone for solving the next step. Initially, CoT was elicited by adding a sentence such as 'Let's think step by step' to the prompt, or by showing examples that include worked solutions. Today's reasoning models are trained through post-training such as reinforcement learning to write long CoT on their own.

Direct answering vs. CoTQuestion → Direct answer; Question → Step 1 reasoning; Step 1 reasoning → Step 2 reasoning(Continue writing); Step 2 reasoning → Final answerDirect answeringCoTQuestionDirect answerProne to mistakeson complex problemsQuestionStep 1 reasoningStep 2 reasoningFinal answerContinue writing
Direct answering vs. CoT The top shows going straight from the question to the answer; the bottom shows the CoT approach, which writes intermediate reasoning in sequence before reaching the answer. The number of reasoning steps is illustrative.

Why it matters

It often improves accuracy on problems that require multiple steps, such as math problems, logic puzzles and coding. Because a large problem is broken into smaller pieces, it reduces the mistakes that tend to occur when answering in one go. It is a representative method of test-time compute, which improves performance by spending more computation at answer time, and it is also the foundation of reasoning models. Another advantage is that people can read the reasoning and check where it went wrong.

Things to know

The longer the reasoning, the more tokens it uses, increasing cost and response time. For simple questions, the benefit is small or almost nonexistent. Plausible-looking reasoning can contain incorrect steps. Also, there is no guarantee that the written reasoning faithfully reflects the model's actual internal judgment, which makes this a topic of interpretability research. Some services show only a summary instead of the raw CoT.

Example

Chatbots that use reasoning models often show a 'thinking' indicator along with a collapsible view of the reasoning process, which is a typical example of CoT being made visible. With general models, you can get a similar effect by asking, 'Explain it step by step.'

Related terms

Reasoning model Test-time compute Prompt Reinforcement learning RLVR Interpretability