Study traces GEPA merge instability to configuration, not the operator
Original title: When Does GEPA Merge Help? A Mechanism-Level Diagnostic
A nine-configuration sweep across five datasets finds merge can help both gpt-4.1-mini and Qwen3-8B.
GEPA's system-aware merge operator was reported to improve gpt-4.1-mini on average while degrading Qwen3-8B. A sweep over merge timing and operator shows both models can improve when merge is configured appropriately. Prompt-level features do not predict merge-accept decisions, ruling out static-feature gating. A behavior-based probe avoids large merge failures by abstaining on high-risk tasks but does not consistently beat NoMerge.
- The probe is released as GEPA-MergeProbe on GitHub.
- Tests cover five datasets and two models.
- The original GEPA paper noted the asymmetry but left it unexplained.
- Products
- GEPA-MergeProbe
- AI models
- gpt-4.1-miniQwen3-8B
From the original article
1 min read at arxiv.orgGEPA 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.
By Chong Zhao · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.