GEPA-Curator: Curated Learning for Self Improving Agents

GEPA-Curator. Curated Learning for Self Improving Agents. By Zack Yu and ten others, from UC Berkeley, the International Computer Science Institute, and three other institutions. Published on the twenty-seventh of October.

Large language models are increasingly run in multi-turn agentic workloads, where they have to sustain interaction with complex environments. Reflective prompt optimizers such as GEPA improve an agent by learning from its failed rollouts. The catch is that an agentic failure is not always the prompt's fault. It can come from a weakness in the prompt, certainly, but also from limitations of the harness, from stochastic execution, or from a noisy judge. The authors argue that failures of those other kinds may provide no useful learning signal, and may even be harmful to learn from.

If some failures teach the wrong lesson, the optimizer should not be handed all of them. GEPA-Curator augments the prompt optimizer so that it selects which agent experiences get to shape prompt optimization. The claim behind it is that curating which tasks reach the optimizer is essential to performing well, not a refinement added at the end.

Curation here has two components. Task curation filters out the tasks that are likely to provide noisy or unreliable learning signals. Prior success retrieval works from the other direction: for a given task, it selectively retrieves past successful trajectories and provides them to the prompt optimizer as additional context. One component keeps misleading failures away from the optimizer, and the other puts evidence of what has already worked in front of it.

Four environments make up the evaluation: GoBrowse, ALFWorld, AppWorld, and tau-bench. Two different models serve as the optimizer, Gemini three point one Pro and Claude Opus five. Against both baselines, GEPA-Curator matches or exceeds the mean performance. Its gains reach twenty point seven percent over GEPA and twenty-one point two percent over ACE.

Nothing about the reflective step itself is replaced. The optimizer reflects as before. What changes is its diet: which tasks it sees, and which past successes sit beside them. An agent that fails because the harness broke or the judge was noisy has not revealed anything about its prompt, and keeping those failures out of the lesson is the whole of the change.