I can't find a link to an actual paper, that just seems to be a blog post. But from what I gather the problems were manually translated to Lean 4, and then the program is doing some kind of tree search. I'm assuming they are leveraging the proof checker to provide feedback to the model.