All terms
Chain-of-Thought
The step-by-step reasoning an AI model shows before arriving at a final answer.
Some evaluation and annotation tasks specifically involve reviewing a model's reasoning steps, not just its final answer - checking whether the logic holds up even if the final answer happens to be right, or catching a flawed reasoning step that produced a right answer by accident.
Example
A model gets the right final numeric answer to a word problem, but an evaluator notices it made two arithmetic errors that happened to cancel out - and flags the reasoning as unreliable despite the correct result.