GEPA-Curator: Curated Learning for Self Improving Agents
Filtering what agents learn from produced reported gains of up to 20.7% over GEPA and 21.2% over ACE.
GEPA-Curator filters tasks likely to give prompt optimizers noisy or unreliable feedback, then retrieves past successful trajectories as additional context. The authors argue that failed runs can reflect harness limitations, randomness or noisy judges, making them poor learning material. Across GoBrowse, ALFWorld, AppWorld and τ-bench, the framework matched or exceeded both baselines’ mean performance using Gemini 3.1 Pro and Claude Opus 5 as optimizers.
ArXiv · AI/CL/LG's note
GEPA-Curator filters tasks likely to give prompt optimizers noisy or unreliable feedback, then retrieves past successful trajectories as additional context. The authors argue that failed runs can reflect harness limitations, randomness or noisy judges, making them poor learning material. Across GoBrowse, ALFWorld, AppWorld and τ-bench, the framework matched or exceeded both baselines’ mean performance using Gemini 3.1 Pro and Claude Opus 5 as optimizers.
ArXiv · AI/CL/LG's note
score 5