SpecEvo: Speculative Evolution with Large Language Models for Cost-Efficient Scientific Discovery

SpecEvo is a multi-LLM framework that coordinates lightweight and frontier models for cost-efficient scientific discovery through continual consolidation of search evidence. Lightweight models act as parallel Speculators generating and refining candidate programs and as an Advisor distilling accumulated successes, failures, and saturated directions into persistent guidance. A frontier-model Navigator periodically inspects the evolving population and intervenes through synthesis, surgical refinement, or reframing based on search stagnation.