The article also outlines future improvements, including supporting other verifiers like Coq and other LLM infrastructures like Ollama. It also plans to design a steerable interaction for feedback to the LLM and a reinforcement learning scheme where the LLM learns from trials. The project credits its montecarlo library to ImparaAI/monte-carlo-tree-search and its inspiration to the paper "Planning with Large Language Models for Code Generation" (ICLR 2023).
Key takeaways:
- The prototype synthesizes verified code with an LLM using Monte Carlo Tree Search (MCTS) to explore the space of possible generation of a verified program.
- The project has been tested on the HAL machine and relies on GPU access. It uses Dafny and logs for example runs can be found in the log directory.
- Future improvements include supporting other verifiers like Coq, other LLM infrastructures like Ollama, designing a steerable interaction for feedback, and a reinforcement learning scheme for the LLM to learn from trial.
- The montecarlo library used is adapted from ImparaAI/monte-carlo-tree-search and the inspiration comes from the paper _Planning with Large Language Models for Code Generation_ (ICLR 2023).