Chong Zhao, Ruyi Yang, Changming Liu, Alice Zhao
This paper investigates the instability of GEPA's merge operator across models, showing that performance variations stem from configuration sensitivity rather than inherent flaws. By testing nine merge configurations, the authors demonstrate that optimal settings can improve both GPT-4.1-mini and Qwen3-8B. A behavior-based probe using diagnostic insights reduces large merge failures by abstaining on high-risk tasks.
Understanding merge instability enables safer, more reliable prompt optimization in real-world LLM applications. The findings guide practitioners in tuning GEPA configurations and inform future design of adaptive, self-diagnosing prompt optimizers. The open-source probe offers a practical safeguard against catastrophic merge failures.
Thank you to arXiv for use of its open access interoperability. arxiv.org. aibrief.ai is not affiliated with or endorsed by arXiv.