HyperTree Proof Search for Neural Theorem Proving
This work addresses the problem of improving automated theorem proving for mathematical reasoning, representing a strong specific gain in performance.
The authors tackled automated theorem proving by introducing HyperTree Proof Search (HTPS), an online training procedure for a transformer-based prover, which achieved 82.6% accuracy on a held-out set of Metamath theorems, up from 56.5%, and improved accuracy on the Lean-based miniF2F-curriculum dataset from 31% to 42%.
We propose an online training procedure for a transformer-based automated theorem prover. Our approach leverages a new search algorithm, HyperTree Proof Search (HTPS), inspired by the recent success of AlphaZero. Our model learns from previous proof searches through online training, allowing it to generalize to domains far from the training distribution. We report detailed ablations of our pipeline's main components by studying performance on three environments of increasing complexity. In particular, we show that with HTPS alone, a model trained on annotated proofs manages to prove 65.4% of a held-out set of Metamath theorems, significantly outperforming the previous state of the art of 56.5% by GPT-f. Online training on these unproved theorems increases accuracy to 82.6%. With a similar computational budget, we improve the state of the art on the Lean-based miniF2F-curriculum dataset from 31% to 42% proving accuracy.