Rather than rewarding only a final answer, the project collected human judgments on intermediate reasoning steps and trained process reward models from those labels. PRM800K exposes the annotations, labeling instructions and scored model samples used in the study.
OpenAI reported that process supervision significantly outperformed outcome supervision for training reward models on MATH solutions; its process-supervised model solved 78% of a representative held-out MATH subset, and the project released PRM800K with 800,000 step-level correctness labels.
Step-level verification became a durable component of later mathematical reasoning systems, providing training and selection signals that diagnose where a chain of reasoning first goes wrong rather than judging only its endpoint.
Significance rationale: The work made step-level human feedback a concrete and scalable research object for mathematical reasoning, directly comparing process and outcome supervision and releasing PRM800K as a large process-supervision dataset that influenced later verifier and reward-model work.
—
OpenAI released the research paper and official description comparing process and outcome supervision on MATH.
The public PRM800K repository exposes 800,000 step-level labels, labeling instructions, data splits and evaluation samples. E3 records inspectability and does not imply independent verification of the reported model results.
No corrections recorded.
