TurboEvolve: Towards Fast and Robust LLM-Driven Program Evolution
Summary: arXiv:2604.18607v1 Announce Type: cross
Abstract
LLM-driven program evolution has shown great promise in discovering high-quality programs. However, its inherent costs and run-to-run variance often impede reliable progress. To address these challenges, we introduce TurboEvolve, a multi-island evolutionary framework designed to enhance sample efficiency and robustness while operating within fixed evaluation budgets.
Key Features of TurboEvolve
- Multi-Island Evolutionary Framework: TurboEvolve utilizes a multi-island architecture that allows for parallel exploration of solutions, enabling better performance while managing computational resources effectively.
- Verbalized Sampling: This innovative approach prompts the LLM to produce K diverse candidates with explicit self-assigned sampling weights, facilitating a more nuanced exploration of the solution space.
- Dynamic Online Scheduler: The online scheduler intelligently adapts the number of candidates K based on the evolution process’s progress, expanding exploration during stagnation and optimizing resources during steady improvement.
- Seed-Pool Injection: To leverage existing solution pools, TurboEvolve introduces a method of clustering seeds and distributing them across islands with controlled perturbations and elitist preservation, balancing diversity and refinement.
Performance and Results
TurboEvolve has demonstrated consistent improvements across various program-optimization benchmarks. The framework not only achieves stronger performance at lower evaluation budgets but also enhances best-known solutions for several tasks. This dual advantage makes TurboEvolve a compelling choice for researchers and developers looking to push the boundaries of LLM-driven program evolution.
Conclusion
In summary, TurboEvolve represents a significant advancement in the realm of LLM-driven program evolution. By focusing on sample efficiency and robustness, this innovative framework addresses the critical challenges of cost and variability, providing a reliable pathway for discovering high-quality programs. As the field continues to evolve, TurboEvolve stands out as a promising solution that could reshape the future of program optimization.