EvoTune

EvoTune ("Algorithm Discovery With LLMs: Evolutionary Search Meets Reinforcement Learning") combines FunSearch-style evolutionary program search with RL fine-tuning of the searcher LLM itself, so the model improves at proposing better programs over the course of the search — discovering algorithms faster than search alone. From Caglar Gulcehre's CLAIRE lab (co-authors include Fields Medalist Maryna Viazovska); ~140★.