ProverBot9000: Neural Networks for Proof Assistance

We introduce ProverBot9000, a state-of-the-art tool for proof automation and assistance. ProverBot9000 examines partially finished Coq proofs and proposes tactics to make progress on the proof. It generates these tactics using a neural network-based language model of Ltac. ProverBot9000 is trained on human-generated proofs so it suggests tactics that human experts are likely to use in a given proof state. Furthermore, it can be fine-tuned for a specific domain (e.g. distributed systems or compilers) simply by adding completed proofs in that domain to its training set. We evaluate ProverBot9000 on proofs of peephole optimization correctness in a verified C compiler.