Process supervision gives feedback on individual written reasoning steps; outcome supervision gives feedback on the final result. OpenAI's 2023-05-31 study compared learned solution scorers, called reward models, on the MATH problem dataset. For each problem, each scorer ranked generated candidate solutions; evaluation measured whether its selected answer was correct. The process-supervised scorer outperformed the compared outcome-supervised scorer under this setup. Generalization beyond mathematics was explicitly unknown. Analytical boundary: scoring written steps is not a direct measurement of a reader's comprehension or of the model's internal computation.

Evidence status: demonstrated_in_mathematical_testbed.