SkyDiscover: A Flexible, Adaptive Framework for AI-Driven Scientific and Algorithmic Discovery
Shu Liu (UC Berkeley), Mert Cemri (UC Berkeley), Shubham Agarwal (UC Berkeley), Alexander Krentsel (UC Berkeley), Ashwin Naren (UC Berkeley), Qiuyang Mang (UC Berkeley), Zhifei Li (UC Berkeley), Akshat Gupta (UC Berkeley), Monishwaran Maheswaran (University of California, Berkeley), Audrey Cheng (UC Berkeley), Melissa Pan (UC Berkeley), Ethan Boneh (Stanford University), Kannan Ramchandran (University of California at Berkeley), Koushik Sen (UC Berkeley), Matei Zaharia (UC Berkeley), Alexandros G. Dimakis (UC Berkeley), Ion Stoica (UC Berkeley)
Architectural Patterns & Composition
Summary
A modular framework for AI-driven algorithmic discovery via evolutionary search, achieving strongest open-source performance across 200+ optimization tasks and matching AlphaEvolve on many.
Description
LLM-driven evolutionary search is emerging as a powerful approach for discovering algorithms and designs, but existing frameworks are difficult to reuse, extend, and compare. We present SkyDiscover, a flexible, adaptive framework for AI-driven scientific and algorithmic discovery. SkyDiscover decomposes the discovery loop into four reusable components: Context Builder, Solution Generator, Evaluator, and Solution Selector, while exposing the control logic above them as a programmable interface. This modular design enables rapid experimentation and even supports adaptive designs where AI can adapt or even optimize the optimization process itself during search. We demonstrate SkyDiscover across more than 200 optimization tasks spanning mathematical optimization, systems design, algorithmic programming, and constrained image generation. Under fixed budgets and shared models, the adaptive algorithms implemented on top of SkyDiscover achieves the strongest open-source performance compared to OpenEvolve, ShinkaEvolve, and GEPA, and matches or exceeds AlphaEvolve on many tasks. A live demo further showcases end-to-end discovery with real-time monitoring, human-in-the-loop steering, and meta-optimization of the search process.
