Verifier transfer coefficient #
The verifier transfer coefficient is the ratio of held-out true success to training-verifier acceptance, used to measure whether verifier-guided optimization transfers beyond the verifier that produced the reward.
Definition #
If a verifier accepts 80 percent of trajectories and an external audit finds that only 56 percent are true successes, the transfer coefficient is 0.70. The coefficient is not an accuracy score. It is a transfer score: how much of the local verifier gain survives contact with a held-out or grounded evaluator.
Why this matters #
The same local acceptance rate can hide very different systems. One model may generalize cleanly from the verifier. Another may exploit the verifier and collapse under transfer. The coefficient makes that difference visible.
Production signal #
Track the coefficient per task class and per verifier version. A falling coefficient after RLVR, best-of-N selection, or judge-model tuning is an early warning that measured reward is decoupling from true success.