Kevin on Lemmata

That's true. I guess you could build something entirely different as well. Like design some subsystem that tries to generate "interesting intermediate results", treat that as a single step in a graph, and run MCTS on that model. So an entirely separate architecture for either "prove this result using tree search on the underlying prover" or "use an LLM to generate code and hope it compiles". Perhaps in some sense this is equivalent to training a reasoning LLM that can call out to the prover in substeps.