I was recently asked to review this article by PetarV et al. before publication. Note that what I write below is my personal opinion and is in no way related to my work at Google :)
What they do in the paper is take a heuristic problem from a short contest (mostly Hash Code qualification rounds, but also AHC 039 to test on a contest that happened after the LLM training), implement a greedy solution in Python which chooses the next decision to make using a scoring function, and then use an LLM + evolutionary algorithm to come up with a better scoring function. I think it's a pretty cool separation of responsibilities between the LLM and the human, and I think the results are made even more impressive by the fact that they did not use simulated annealing, and only used local search in one input of one of the tasks (see the footnote on page 3), relying just on the greedy in all other cases, so there seems to be more headroom for this approach.
I've tried to search for other attempts to use LLMs for this type of contests, but could not easily find one. Surely this has been tried before, maybe somebody has more links?
