Lightman et al.: process-supervised reward models and PRM800K

What changed

Compared step-level and outcome-only supervision for mathematical reward models, reporting stronger solution selection with process feedback and active-learning gains, and releasing the PRM800K human-label resource.

Scientific scope

Compared step-level and outcome-only supervision for mathematical reward models, reporting stronger solution selection with process feedback and active-learning gains, and releasing the PRM800K human-label resource. Results concern the stated MATH subset, base model, feedback budget and sampling/selection procedure, not single-sample 78% accuracy or universal superiority. Process supervision builds on Uesato et al. rather than originating here.

Evidence

The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. Results concern the stated MATH subset, base model, feedback budget and sampling/selection procedure, not single-sample 78% accuracy or universal superiority. Process supervision builds on Uesato et al. rather than originating here.

Contributors

• Source-listed authors and technical contributors · SCIENCE: Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities.

• Let's Verify Step by Step
• Lightman et al.: process-supervised reward models and PRM800K — company research account

Provisional WRS assessment, edition 2026-10-16.1. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.

Full scientific record · Browse contribution grades · Request a correction