Pythagoras-Prover introduces a compute-conscious family of Lean theorem provers for formal mathematics.
The project combines curriculum training on Lean-verified data with several model sizes, including a proof-of-concept diffusion-based prover that iteratively refines formal proofs during inference.
Formal proving is an important testbed for reasoning models because success is machine-checkable. More efficient provers could make rigorous verification cheaper for research and software-adjacent workflows.