Gabriel Poesia

Proof Artifact Co-training for Theorem Proving with Language Models (@ ICLR 2022)

Jesse Michael Han, Jason Rute, Yuhuai Wu, Edward W. Ayers, Stanislas Polu


This paper describes a simple data augmentation scheme for learning language models for theorem proving. The main challenge is that there is limited data for theorem proving when compared with other domains: you can get millions of Python repositories from Github, for instance, but much fewer sources for Lean proofs (the theorem proving language the authors use). Their idea is to use proof objects to create a number of auxiliary tasks, which all benefit from the uniformity of training language models for different tasks. Given the tasks all require learning relevant structures of proofs and of the underlying language, this co-training scheme improves downstream performance.