Language models are currently the best AI systems that solve general reasoning problems, including mathematical ones. So it seems more than obvious to model the problem of finding proofs as a language problem.