Han et al 2021 - Proof Artifact Co-training for Theorem Proving with Language Models Generates the next tactic using an LLM. Uses a best-first search algorithm: maintains a priority queue of search nodes, of tactic state (partial proof) and metadata, sorted by LLM heuristic score. For Lean. Introduces the LeanStep dataset to to predict the next tactic give a goal state, built from mathlib. Uses an interesting PACT pretraining dataset (sec 3.2).