Griffon
Griffon is the combination of a novel transformer architecture and a Coq plugin that allows user to ask for useful lemma suggestions while trying to prove a theorem. The model is trained on the CoqGym dataset[1]. The encoder is based on the Code Transformer architecture[2]. The decoder uses a two level attention, first looking at the titles of each hypothesis in the context to compute an attention distribution over statements before attending to tokens in each statement.
References
[1] Kaiyu Yang, Jia Deng Learning to Prove Theorems via Interacting with Proof Assistants CoRR,abs/1905.09381, 2019
[1] Daniel Zügner, Tobias Kirschstein, Michele Catasta, Jure Leskovec, and Stephan Günnemann Language-agnostic representation learning of source code from structure and context. International Conferenceon Learning Representations (ICLR), 2021