TurboEvolve: Towards Fast and Robust LLM-Driven Program Evolution

A research team from multiple Chinese institutions has introduced TurboEvolve, a novel framework that addresses two critical limitations in LLM-driven program evolution: high computational costs and unreliable performance variance. The system adapts evolutionary algorithm principles specifically for large language models, creating a "multi-island" approach where different solution populations evolve in parallel. This architecture allows the framework to maintain diversity while efficiently exploring the solution space, overcoming the tendency of single-threaded LLM optimization to get stuck in local optima.

TurboEvolve's core innovation is its two-phase adaptive system. First, it employs "verbalized sampling," where the LLM generates multiple candidate solutions simultaneously with self-assigned confidence weights, dramatically improving sample efficiency. Second, an online scheduler dynamically adjusts the number of candidates (K) based on progress—expanding exploration during stagnation phases and reducing overhead during steady improvement periods. This intelligent resource allocation is what enables the framework to outperform traditional methods under fixed evaluation budgets.

The framework further enhances performance through "seed-pool injection," which intelligently reuses and perturbs existing high-quality solutions across different evolutionary islands. By clustering similar solutions and applying controlled variations, TurboEvolve balances exploration of new possibilities with refinement of promising approaches. The researchers demonstrated significant improvements across multiple program optimization benchmarks, where TurboEvolve not only achieved better results with fewer computational resources but actually discovered new state-of-the-art solutions for several challenging tasks.