I think serious math research progress should come in 1-2 years. It basically only depends on how hard informal verification is, because training data should be not a problem and if informal verification is easy you can throw RL compute at it until it improves.