GEPA improves prompts for model programs by proposing changes and testing their behavior. It works with DSPy, a Python framework for programming model workflows. This extension proposes several changes together so evaluations can run concurrently.
A round selects multiple parent prompts and several changes per parent. Improved candidates receive full validation before the pool is updated.
- On math, two parents with two changes each raise held-out scores from 68.9% to 72.1% under the same 5,000-call scoring budget. Runtime falls from 7.7 to 4.1 hours.
- On HoVer, eight parents with one change each raise complete evidence-page retrieval from 49.0% to 60.0% under a 3,000-call budget. This measures retrieval, not final fact-verdict accuracy.
- The widest settings run roughly three to four times faster, using up to 512 concurrent workers for math and 64 for retrieval.
Caveat: The two-task sweep relies on substantial concurrency; its fastest settings and winning prompt configurations are not established defaults for rate-limited installations.
The open-source strategy implements parent-and-change sampling. Its default still proposes one change per round. Fewer adaptive rounds may improve generalization, but the paper does not isolate that explanation causally.