Fisika & Matematika editorial
Open AccessOA2022
Formal Mathematics Statement Curriculum Learning
Expert iteration, combining proof search with learning, dramatically outperforms proof search alone in formal mathematics. Applied to varied difficulty statements, it automatically finds a curriculum and solves high school olympiad problems, achieving state-of-the-art on miniF2F.
Stanislas Polu; Jesse Michael Han; Kunhao Zheng; Mantas Baksys; Igor Babuschkin; I. Sutskeverยท International Conference on Learning Representationsยท 2022ยท DOI 10.48550/arXiv.2202.01344
The core problem
The paper explores the application of expert iteration to language modeling for formal mathematics. The goal is to improve automated theorem proving by interleaving proof search with learning. The authors hypothesize that this approach can outperform proof search alone and can automatically discover a curriculum of increasingly difficult problems without ground-truth proofs.
Innovation
The method uses expert iteration, where a language model is trained on successful proofs found by proof search. The process alternates between proof search using the current model and training the model on the discovered proofs. This is applied to a collection of formal statements with varied difficulty. The authors also manually curate a set of problem statements for targeted improvement.
Introduction
The paper explores the application of expert iteration to language modeling for formal mathematics. The goal is to improve automated theorem proving by interleaving proof search with learning. The authors hypothesize that this approach can outperform proof search alone and can automatically discover a curriculum of increasingly difficult problems without ground-truth proofs.
Why it matters
The results demonstrate the effectiveness of expert iteration in formal mathematics. The ability to automatically generate a curriculum of increasing difficulty is a key finding, as it allows the system to learn without human-provided proofs. The state-of-the-art performance on miniF2F highlights the potential of this approach for automated theorem proving. The authors suggest that further work could explore scaling and applying the method to other formal domains.
Who should read this
Pembaca Research Digest dan praktisi rumpun ini.
Opening member contentโฆ