ABEL: Sample Efficient Online Reinforcement Learning for Neural Theorem Proving

Fabian Gloeckle, Jannis Limperg, Gabriel Synnaeve, Amaury Hayat

Keywords: theorem proving, reinforcement learning, automated theorem proving, MCTS, reasoning, Lean

TL;DR: Online reinforcement learning for Lean gives near-SOTA with almost no data.

Abstract: We propose a scalable and efficient reinforcement learning framework as a strong baseline for theorem proving with limited data. This baseline reaches performances comparable to the current state-of-the-art in theorem proving, while only training on a few hundred examples. This a first step toward an efficient and easily reproducible combination of autoformalization, synthetic data generation and reinforcement learning, which could unlock significant advancements in neural theorem proving.

Concurrent Submissions: N/A

Submission Number: 106