What is chain of thought?
The step by step reasoning a model writes before its final answer. Many expert AI training tasks ask you to write or grade this reasoning.
Chain of thought is the written reasoning that leads to an answer: the steps of a proof, a calculation, or an argument. Models that show their reasoning tend to do better on hard problems, so AI labs collect examples of good reasoning and feedback on bad reasoning.
Where it shows up in the work
- Writing a full, step by step solution to a hard problem, as training data for SFT
- Grading a model's reasoning step by step, and marking the first step that goes wrong
- Writing a hard problem that the model gets wrong, together with the correct solution
This is a large share of the work for domain experts in maths, science, coding, law and finance.
What graders look for
A correct final answer with a broken step in the middle is still wrong reasoning, and most projects want that flagged. Good reasoning states its assumptions, does not skip the step that matters, and can be checked line by line.
Example
A model gets the right answer to a probability question, but step three adds two probabilities that are not independent. The grader marks step three as the first error, even though the final number is correct.
Looking for this kind of work? Science & Math roles