arXiv：cs.AI· Yang Yang, Zining Zhong, Jindong Li, Jiemin Wu, Kaishen Yuan, Wenshuo Chen, Menglin Yang, Yutao Yue·· 19 小时前
TurboEvolve：面向快速稳健的 LLM 驱动程序进化
TurboEvolve: Towards Fast and Robust LLM-Driven Program Evolution
AI 导读
TurboEvolve 是一个多岛屿进化框架，在相同 API 成本预算下提升 LLM 驱动程序进化的样本效率与稳定性。它引入 Verbalized Sampling，在一次 LLM 调用中生成 K 个多样候选，并用在线调度器自适应调整 K；同时通过岛屿感知初始化对种子程序聚类并做受控跨簇混合。
正文
Abstract:LLM-driven program evolution can discover high-quality programs, but high API cost and run-to-run variance hinder reliable progress. We propose TurboEvolve, a multi-island evolutionary framework that improves sample efficiency and stability under matched API-cost budgets. Inspired by multiple-offspring strategies in evolutionary algorithms, TurboEvolve introduces Verbalized Sampling, which elicits K diverse candidates in one LLM call while using self-assigned probabilities only as generation signals, together with an online scheduler that adapts K to expand exploration under stagnation and reduce redundant generation during steady progress. To exploit existing solution pools, TurboEvolve further introduces island-aware initialization, which clusters seed programs into mode-specific subpopulations and applies controlled cross-cluster mixing with elitist preservation to balance diversity, sharing, and refinement. We also curate a reusable cross-task solution-pool dataset to support warm-start initialization studies and matched-resource comparisons. Across mathematical discovery and GPU-kernel optimization tasks, TurboEvolve achieves stronger cost-quality trade-offs and more stable performance than controlled baselines, while reaching competitive reported scores.
| Comments: | 12 pages, 8 figures |
| Subjects: | Neural and Evolutionary Computing (cs.NE); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2604.18607 [cs.NE] |
| (or arXiv:2604.18607v2 [cs.NE] for this version) | |
| https://doi.org/10.48550/arXiv.2604.18607 arXiv-issued DOI via DataCite |
Submission history
From: Zining Zhong [view email]
[v1]
Sun, 12 Apr 2026 12:42:09 UTC (3,393 KB)
[v2]
Mon, 12 Oct 2026 13:56:56 UTC (3,329 KB)
来源：arXiv：cs.AI · arxiv.org