A verifier reward increases code proof verification from 66.1% to 93.5%
An audit rejects each result that changes the specification.
Claimed, not confirmed
Researchers used a verifier, a tool that checks code against a correct specification, as the reward to train code models. On DafnyBench, held-out verification increased from 66.1% to 93.5%. On SWE-Proof, it increased from 56.3% to 100%. The reward gave more accurate results than a pass-or-fail reward.
Sources
Posted
