FunSearch: Mathematical Discoveries from Program Search with Large Language Models
Harvard CMSA via YouTube
Overview
Watch a 27-minute lecture from Harvard CMSA's Mathematics and Machine Learning Closing Workshop where DeepMind researcher Matej Balog presents groundbreaking research on FunSearch, a novel approach combining large language models with systematic evaluation for mathematical discovery. Learn how this evolutionary procedure overcomes traditional LLM limitations like confabulation by searching for programmatic solutions rather than direct answers. Explore practical applications in extremal combinatorics, specifically the cap set problem, where FunSearch discovered new constructions surpassing previous records. Understand how this methodology extends to algorithmic challenges like online bin packing, producing interpretable programs that outperform existing baselines. Discover the potential for creating productive feedback loops between domain experts and AI systems, while gaining insights into deploying these solutions in real-world scenarios.
Syllabus
Matej Balog | FunSearch: Mathematical discoveries from program search with large language models
Taught by
Harvard CMSA