Abstraction and Evolution with Large Language Models

Watch a 51-minute lecture from the IPAM's Naturalistic Approaches to Artificial Intelligence Workshop where Swarat Chaudhuri from the University of Texas at Austin explores innovative approaches to genetic programming enhanced by large language models (LLMs). Discover how traditional evolutionary approaches can be augmented by using zero-shot LLM queries to identify and evolve abstract concepts from high-performing programs. Learn about a novel methodology that combines standard evolutionary steps with LLM-guided operations, creating an iterative process of concept discovery and program evolution. Examine practical applications through case studies in symbolic regression and descriptor-based image classification, where this concept-guided search demonstrates superior performance compared to existing state-of-the-art methods.