3 months ago, I submitted a hack on using Monte Carlo Tree Search (MCTS) with an LLM guided by a program verifier: https://news.ycombinator.com/item?id=38235407
I just wanted to send this update, because we now have a paper that explains the hack in detail.
We also show the use of verifiers with ChatGPT (implemented as actions within GPTs), and our method makes open models like Phind Code Llama competitive with ChatGPT augmented with verifier feedback and step by step instructions.
Happy to answer any questions!
0