“FunSearch works by pairing a pre-trained LLM, whose goal is to provide creative solutions in the form of computer code, with an automated "evaluator", which guards against hallucinations and incorrect ideas.”
DeepMind found the largest cap sets in 20 years and better bin-packing heuristics. PaLM 2 does the guessing. A hand-written evaluator throws out the junk, and an evolutionary loop keeps the winners. The model is a mutation engine, and the evaluator is the part that knows any math. That works only for problems where a human can already score the answer.