Title: Evolution, Abstraction, and Discovery with Large Language Models
Abstract:
Large language models (LLMs) have shown tremendous promise in code generation. This talk demonstrates how these capabilities can be a powerful tool for scientific and mathematical discovery. The central idea is that scientific hypotheses and mathematical conjectures can be naturally represented as programs, which do not follow standard idioms but are novel and surprising. I present LaSR, a method using LLMs to guide an evolutionary search over programs, discovering new programs using a mix of standard evolutionary steps and LLM-guided steps. Experimental results show that LLM-guided search outperforms state-of-the-art baselines in tasks like automated discovery of physics equations and automatic mathematical proof.