arXiv cs.AI· Jialin Zhang, Lakshya A Agrawal, Shangyin Tan, Donghyun Lee, Lutfi Eren Erdogan, Dan Klein, Koushik Sen, Alex Dimakis, Matei Zaharia·· 13 小时前AI 评分31
GEPA 优化器引入批量反思优化：并行候选让优化更快更好
Batching the Reflective Optimization Loop: Parallel Proposals Make GEPA Faster and Better
AI 导读
开源优化器 GEPA 被扩展为每步批量提出并评估多个候选，而非逐个处理。在两个任务上，多数批量运行耗时减半，最快约为原来的四分之一到三分之一；LiveBench-Math 测试分数从 68.9% 升至 72.1%（2×2），HoVer 从 49.0% 升至 60.0%（8×1）。批量设置还更高效地使用指标调用与美元预算。
正文 · 原文
Abstract:Reflective text optimization makes agentic systems better, but the optimization loop is itself a resource-hungry workload whose sequential structure leaves parallel evaluation capacity idle. We extend the open-source GEPA optimizer to propose and evaluate a batch of candidates on each optimization step instead of one candidate at a time. In our sweep on two tasks, most batched runs finished in half the wall-clock time or less, and the fastest in about a quarter to a third. Batched settings also achieved higher held-out test scores, from 68.9% to 72.1% on LiveBench-Math (with $2\times2$) and from 49.0% to 60.0% on HoVer (with $8\times1$). A simple runtime model explains the measured speedups under a finite worker pool, and batched settings also spend their metric-call and dollar budgets more efficiently. We frame the axes along which a reflective optimization step can scale in parallel, a first look at allocating a fixed optimization budget across them.
| Comments: | NeurIPS 2026 Workshop on Resource-Aware Agentic AI. 9 pages, 7 figures. First four authors contributed equally. Code: this https URL |
| Subjects: | Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Machine Learning (cs.LG) |
| Cite as: | arXiv:2610.36774 [cs.AI] |
| (or arXiv:2610.36774v1 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2610.36774 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Jialin Zhang [view email]
[v1]
Tue, 27 Oct 2026 07:41:40 UTC (211 KB)
来源：arXiv cs.AI · arxiv.org