SkyDiscover: A Flexible Framework for AI-Driven Sci. and Algorithmic Discovery (skydiscover-ai.github.io)

SkyDiscover has been introduced as a groundbreaking framework for AI-driven scientific and algorithmic discovery, addressing the limitations of existing evolutionary search frameworks used for algorithm development. Unlike predecessors like AlphaEvolve and OpenEvolve, which are tightly coupled and challenging to modify, SkyDiscover offers a modular architecture that separates its components—Context Builder, Solution Generator, Evaluator, and Solution Selector—allowing for rapid experimentation and comparison across different discovery methods. This new flexibility has resulted in the development of two state-of-the-art evolutionary algorithms, AdaEvolve and EvoX, which collectively achieved significant performance improvements on over 200 tasks, surpassing previous benchmarks and established human solutions. The significance of SkyDiscover lies not only in its superior performance—achieving a 34% median improvement over existing algorithms in programming problems—but also in its potential to redefine how algorithms are discovered and optimized in various domains, including math and systems optimization. With concrete application outcomes, such as a 41% reduction in cloud transfer costs and improved GPU load balancing, SkyDiscover marks a pivotal advancement in leveraging AI for practical, impactful discoveries. Its meta-evolving strategies, which allow on-the-fly adjustments to optimization techniques based on real-time observations, represent a significant evolution in the field of AI/ML research, paving the way for more adaptive and efficient solution generation.