Generalizing autoformalization plus synthetic expert-iteration proofs to data-scarce domains

Determine whether the approach that autoformalizes large-scale informal problems into formal theorems and generates synthetic proofs via expert iteration can be generalized to research mathematics domains that lack abundant human-written problems.

Background

The paper highlights AlphaProof’s strategy combining autoformalization of one million IMO-like problems into formal theorems with synthetic proof generation through expert iteration.

The authors note that while effective in data-rich settings, extending this pipeline to research-level domains—where few human-written problems exist—remains unresolved and may necessitate automated conjecturing.