ASI-LIB-038 technical evolutionary search

ImprovEvolve: Basin-Hopping Meets LLM-Guided Evolutionary Search

Alexey Kravatskiy, Valentin Khrulkov, Ivan Oseledets

ImprovEvolve method diagram
Figure via ar5iv rendering of arXiv:2602.10233

ImprovEvolve is a useful refinement of the AlphaEvolve pattern. Instead of asking an LLM to evolve one monolithic optimizer, it evolves specialized operators for initialization, local improvement, and perturbation.

Results to watch

The paper reports new state-of-the-art packings for multiple hexagon-in-hexagon cases, a stronger lower bound for the second autocorrelation inequality, and improvements for many spherical-code instances.

ASI relevance

The big lesson is cognitive load management. If we want LLM-guided search to scale, we may need decomposed operator libraries rather than heroic single-shot program synthesis.