OpenEvolve: Teaching LLMs to Discover Algorithms Through Evolution (algorithmicsuperintelligence.ai)

OpenEvolve is a groundbreaking open-source evolutionary coding agent designed to enable large language models (LLMs) to autonomously discover algorithms through a quality-diversity search framework. By leveraging LLMs within an evolutionary context, OpenEvolve transcends traditional approaches that rely on manual heuristics or exhaustive searches. Instead, it generates candidate programs using LLM-guided edits, evaluates them based on user-defined metrics, and organizes the successful outcomes with MAP-Elites to promote diversity in algorithmic solutions. This framework is particularly significant for the AI/ML community as it opens avenues for automated algorithm discovery across various domains, such as systems optimization and scientific research. Key innovations of OpenEvolve include an island-based evolutionary model enabling parallel exploration, cascade evaluation to streamline the testing of generated candidates, and a double-selection strategy that encourages both high-fitness and diverse inspirations in the candidate generation process. Notable achievements using OpenEvolve showcase its effectiveness, including significant speedups in JAX compilation, the discovery of optimized algorithms for GPU kernels, and even the evolution of prompts that enhance natural language processing tasks. By supporting complex configurations and parallel evaluations, OpenEvolve not only facilitates extensive exploration of algorithm spaces but also emphasizes the role of LLMs in creative problem-solving within machine learning.