An evolved verifier had 0.21 more held-out agreement on MBPP+
Without its guards, the verifier always passes, but the skills that it trains are as good.
Claimed, not confirmed
Researchers made the verifier, the check on changes to an agent, the part that changes. A verifier tells if a change to the agent is an improvement. They made it from small detectors. They selected it for agreement with a reference set of 10 items, not for the score of the agent. On MBPP+ it had 0.21 more held-out agreement than the hand-written seed. The researchers say that a task score cannot show that an evolved verifier is correct.
Sources
Posted
