15 upvotes
Paper argues human workflows still outpace raw LLMs in theorem proving
What is h1: A Taste of the Bitter Lesson in Automated Theorem Proving?
The authors review recent automated theorem proving systems and find that simply making language models bigger doesn’t close the performance gap. They show that adding planning, problem decomposition, and specialized orchestration improves proof success.
Who it is for
Anyone building AI for formal mathematics should note that structured pipelines still beat raw scaling, and can incorporate these techniques into their own provers.
How to try it
open https://huggingface.co/papers/2610.35962