arXiv cs.CL· Chong Zhao, Ruyi Yang, Changming Liu, Alice Zhao·· 12 小时前AI 评分31
GEPA 的 merge 何时有效？一项机制层面的诊断
When Does GEPA Merge Help? A Mechanism-Level Diagnostic
AI 导读
研究揭示 GEPA 的 merge 算子不稳定源于配置敏感而非算子本身缺陷：在五个数据集、两个模型上的九种 merge 时机与算子配置扫描显示，配置得当时 gpt-4.1-mini 与 Qwen3-8B 均可从 merge 中获益。
正文
Abstract:GEPA is among the strongest reflective prompt optimizers for language-model programs, but its system-aware merge operator is unstable across models: merge improves gpt-4.1-mini on average while degrading Qwen3-8B, an asymmetry noted but left unexplained by the GEPA paper. We investigate this instability across five datasets and two models. Our nine-configuration sweep over merge timing and merge operator shows that merge can improve both models when configured appropriately. This suggests that merge instability is better explained by configuration sensitivity than by an inherent flaw in the merge operator. We then test whether prompt-level features predict GEPA's merge-accept decisions and find that they do not, ruling out static-feature gating. As a proof of concept, a behavior-based probe built from these diagnostics avoids large merge failures in our multi-seed runs by abstaining on high-risk tasks, although it does not consistently outperform NoMerge. Code is available as GEPA-MergeProbe at this https URL .
| Comments: | 12 pages, 5 figures. Code: this https URL |
| Subjects: | Computation and Language (cs.CL) |
| Cite as: | arXiv:2610.34027 [cs.CL] |
| (or arXiv:2610.34027v1 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2610.34027 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Ruyi Yang [view email]
[v1]
Sun, 25 Oct 2026 03:32:01 UTC (1,650 KB)
来源：arXiv cs.CL · arxiv.org