Self improving language agents learn better by choosing tasks carefully
GEPA-Curator: Curated Learning for Self Improving Agents
Computation and Language
Summary
Large language models can act as agents that try to complete multiple connected tasks, but sometimes they fail in ways that don't help them learn. The authors show that not all failure experiences are useful for improving these agents, so their system, GEPA-Curator, picks which tasks and past successes to use for learning. By selecting better examples, the agents get smarter and perform tasks more successfully. They tested this approach on several language-based tasks and found it improves results by up to about 20%.
What this means in practice
- •For ai application developers: Improve multi-turn language agent performance by filtering ineffective training tasks and adding past successful examples for better prompt tuning.
- •For automation system engineers: Build more reliable autonomous systems by curating agent experiences to avoid misleading feedback during self-improvement.
Authors
Zack Yu, Nicholas Lee, Minjae Lee, Lakshya Agarwal, Luca Manolache, Wonjun Kang, Matei Zaharia, Michael W. Mahoney, Kurt Keutzer, Sehoon Kim, Amir Gholami
Abstract
Large Language Models (LLMs) are increasingly used in multi-turn agentic workloads that require sustained interaction with complex environments. Existing reflective prompt optimizers such as GEPA learn from failed rollouts, but agentic failures can also come from prompt weaknesses, harness limitations, stochastic execution, or noisy judges. We argue that such failures may not provide useful learning signal, and may even be harmful; and thus it is essential to curate which tasks are provided to the prompt optimizer for better performance. We introduce GEPA-Curator, a framework that augments the prompt optimizer to select which agent experiences should shape prompt optimization. In particular, GEPA-Curator consists of two components: task curation, which filters out tasks that are likely to provide noisy or unreliable learning signals, and prior success retrieval, which selectively retrieves past successful trajectories for a task and provides them as additional context to the prompt optimizer. We evaluate GEPA-Curator on GoBrowse, ALFWorld, AppWorld, and $τ$-bench, using both Gemini 3.1 Pro and Claude Opus 5 as optimizer models. GEPA-Curator matches or exceeds the mean performance of both baselines, with gains of up to 20.7% over GEPA and 21.2% over ACE.