FunSearch

Published in Nature on December 14, 2023, FunSearch (short for searching in the function space) was the first method to produce genuinely new scientific knowledge using a large language model. It pairs a pre-trained LLM, which proposes candidate programs in code, with an automated evaluator that rejects hallucinations; an evolutionary loop then breeds the winners. Applied to the longstanding cap set problem in extremal combinatorics it found new constructions, and on online bin packing it discovered heuristics better than human-designed ones.

DeepMind open-sourced the FunSearch implementation under Apache 2.0 (github.com/google-deepmind/funsearch), so anyone can run the method with their own LLM and evaluator. The honest limitation is that the LLM itself never understands the problem — it only mutates priority functions while the evolutionary wrapper does the steering — so FunSearch is a discovery amplifier for well-specified search problems, not a general mathematician.