SeaEvo: Advancing Algorithm Discovery with Strategy Space Evolution

A team of researchers led by Sichun Luo from multiple institutions (including Lei Li, Junlan Feng, and Qi Liu) has introduced SeaEvo, a new framework that advances automated algorithm discovery by treating natural-language strategy descriptions as persistent, first-class evolutionary state rather than transient prompt context. The system addresses a key limitation of LLM-guided evolutionary search: traditional methods track progress through executable programs and scalar fitness, often failing to distinguish syntactically different implementations of the same idea, preserve lower-fitness but strategically promising directions, or detect when an entire strategy family has saturated.

SeaEvo augments each candidate program with an explicit natural-language strategy description and leverages it through three core mechanisms. First, Strategy Articulation transforms mutation into a diagnose-direct-implement process, improving how new programs are generated. Second, Stratified Experience Retrieval organizes the evolutionary archive into strategy clusters and selects inspirations based on behavioral complementarity rather than just fitness. Third, Strategic Landscape Navigation periodically summarizes effective, saturated, and underexplored strategy families to guide future mutations. Across benchmarks spanning mathematical algorithm discovery, systems optimization, and agent-scaffold tasks, SeaEvo improved underlying evolutionary backbones, with particularly large gains of 21% relative improvement on open-ended system optimization tasks. The results suggest that persistent strategy representations provide a practical mechanism for improving the robustness and efficiency of LLM-guided evolutionary search, pointing toward compound AI systems that accumulate algorithmic knowledge over time.