Make the proof from LLMs with the same constraints as human mathematicians have:
The proof must be as succinct as possible so that it's reviewable. It's ok for the same LLM to publish repeatedly, but it's not Ok to drop a bulk 600kb Lean proof.
The proof must build on other works that are similarly bite sized and published before (so that it's incrementally reviewed)
The proof must be published as soon as fully formed, and not in bulk like current LLMs proofs.
The proof must be as succinct as possible so that it's reviewable. It's ok for the same LLM to publish repeatedly, but it's not Ok to drop a bulk 600kb Lean proof.
The proof must build on other works that are similarly bite sized and published before (so that it's incrementally reviewed)
The proof must be published as soon as fully formed, and not in bulk like current LLMs proofs.