noun · AI technique
Chain of thought
also called step-by-step reasoning
Asking the AI to show its work. Writing the steps out first often leads to a better final answer than jumping straight to it.
Prompting or training a model to write out intermediate reasoning steps before its final answer.
The model writes intermediate steps as ordinary tokens before the answer. Because each step becomes input for the next token, the model gets more computation and a place to keep partial results, which helps on multi-step problems.
Real-life example
You ask a model a word problem about train speeds and it blurts out a wrong number. You add "work through it step by step" and it writes out the distances and the times first. The final number is now right. The steps were not decoration, they were where the work happened.
Playground
Ask the same question with and without “think step by step” and compare the replies.
Pick a question
The prompt it receives
A notebook and a pen cost $1.10 together. The notebook costs $1.00 more than the pen. How much is the pen?
The model’s reply
Press “Ask the model” to see its reply.
Illustrative: scripted answers chosen to show the pattern. A real model can get any of these right or wrong, and its written steps are not a guaranteed record of how it got there.
Where the analogy breaks
-
The written steps are not a guaranteed record of the real reason. The explanation can sound sensible and still differ from what drove the answer.
-
It helps most on multi-step problems. On a simple question it just adds length without improving anything.
-
It costs tokens and time. Every step is generated text that you wait for and may pay for.