Pythagoras-Prover Achieves State-of-the-Art Formal Theorem Proving With 167x Fewer Parameters Than DeepSeek

Researchers introduce Pythagoras-Prover, an open-source family of efficient Lean theorem provers that outperforms much larger models through curriculum learning and a novel data augmentation technique called Augmented Lean Formalisation (ALF). The 4B-parameter model surpasses DeepSeek-Prover-V2 (671B parameters) on the MiniF2F benchmark, while the 32B variant sets new open-source performance records at 93.0% accuracy and solves 93 of 672 Putnam problems.

Why it matters: This breakthrough demonstrates that formal theorem proving—a compute-intensive AI task critical for mathematics, verification, and program correctness—can be made dramatically more efficient through smarter training strategies, reducing barriers to practical deployment and experimentation.