I believe the main ingredient is Lean, which is a formal language resembling a programming language. Math proofs written in Lean can be verified deterministically with a computer, which really helps mitigate the hallucination problems of LLMs.

100% this. Also, looking back at an earlier example that was actually written up in more detail, AlphaGeometry 1 got 28/30 problems, but entirely stripping out the LLM from the system, the symbolic logic proportion alone could get 14/30, and replacing the LLM with different heuristic methods could get 18/30 and 21/30 (for different methods).

Even if math research works out perfectly well (which is a still big if), it’s not going to pay the bills. They would need to find a use case in the real world, where hallucinations can cause serious damage and cannot be formally prevented. And they have certainly tried. Math will not change the fact that all of this will collapse.

The boosters and LLM companies still believe LLMs get their current level of performance by generalizing and not just memorizing facts (and maybe a wide shallow pool of weak heuristics). So they are hoping by pushing the LLM performance up in some narrow domain they can churn out synthetic data for, they will see some large general improvements in LLM performance.