Interesting Nature paper from Google DeepMind on using LLMs for solving mathematical problems that are easy to evaluate but difficult to solve:

https://www.nature.com/articles/s41586-023-06924-6

Their procedure, which they call FunSearch, applies an islands model genetic algorithm using the Codey LLM on the cap set problem and online bin packing.

This seems in line with my personal theory that we should view LLMs similarly to MCMC methods, i.e., they are closer to random number generators and should be optimized for sampling.