Direct discovery · Mathematics · Computer science
FunSearch uses language models and program search for mathematical discovery
New constructions for the cap set problem and new algorithms for studied online bin-packing distributions.
Summary
FunSearch combined a frozen language model, a scoring evaluator, and evolutionary search over programs. The paper reports new mathematical and algorithmic outputs in settings where candidate programs can be executed and compared.
AI role
Generated programs that were scored by an evaluator and improved through evolutionary search.
Narrative role
This is a useful precursor milestone for AI discovery because it demonstrates a verifiable loop: generate, execute, score, and iterate.
Caveat
The approach fits domains with executable evaluators; it is not general proof of autonomous scientific discovery.