Torcheur is an automated theorem prover for Multiplicative Linear Logic (MLL) and an experimental framework for learning proof search with neural networks.
The project automatically generates random MLL statements and valid proofs to create a synthetic dataset for transformer-based models implemented in PyTorch.
- automated generation of MLL statement and proofs
- compact serialization of MLL proofs
- transformer-based architecture for autoregressive proof generation
- unit test for proof generators and serialization
- latexify a proof tree for visualization
- generate proof state, next action) dataset
- learn (state, action) proof (IN PROGRESS)
- integrate the next-step predictor into a proof search algorithm and evaluate it (IN PROGRESS)