Why it matters

Verifier quality bottlenecks inference scaling. Understanding shapes design.

Advertisement

The architecture

Verifier: predicts correctness / value.

Larger = better quality.

Calibration matters for search.

Verifier scalingVerifier modelsize + qualityCalibrationconfidence accurateSearch / BoNquality liftProcess reward models > outcome reward models for reasoning; growing research
Verifier scaling.
Advertisement

How it works end to end

Process reward model (PRM) vs outcome (ORM).

PRM: step-by-step feedback.

Better calibration + gain.