SpecEvo is a multi-LLM evolutionary search framework that coordinates lightweight and frontier models to cut the computational cost of scientific discovery. Lightweight models play two roles: Speculators generate and refine candidate programs in parallel; the Advisor consolidates successes, failures and saturated directions into persistent guidance.
At sparse checkpoints, a frontier-model Navigator inspects the population and, depending on stagnation, intervenes by synthesis, surgical refinement or re-framing. This “speculate‑then‑consolidate” workflow assigns routine exploration and evidence aggregation to cheap models, reserving expensive frontier reasoning for strategic moves that benefit many subsequent low-cost evaluations.
Across 176 tasks spanning four scientific and algorithmic discovery domains, SpecEvo matches or exceeds frontier-model baselines while reducing API cost by up to 80%. On equation-discovery it improves evaluation efficiency by up to 1.5× over cost-efficient baselines and beats the strongest CO-Bench baseline by 3.8%. The results demonstrate that trajectory-conditioned model coordination is an effective path to cost-efficient scientific discovery.
Review: SpecEvo’s hierarchical model division balances exploration efficiency with compute expense, offering a reproducible blueprint for deploying large language models in real scientific research.