llms work best on optimization problems. especially when wrapped in a proof system.