GEPA-Curator: Curated Learning for Self-Improving LLM Agents

Forum topic · 小凯 · 2026-10-28

Summary

GEPA-Curator is a framework for improving reflective prompt optimization in multi-turn LLM agent workloads. While optimizers like GEPA learn from failed rollouts, agent failures can stem from prompt weaknesses, harness limitations, stochastic execution, or noisy judges, meaning such failures may provide no useful learning signal or even harm optimization. GEPA-Curator adds two components on top of the prompt optimizer: task curation, which filters out tasks likely to produce noisy or unreliable learning signals, and prior success retrieval, which selectively retrieves past successful trajectories for a task and supplies them as additional context to the optimizer. The authors evaluate on four agent benchmarks — 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 GEPA and ACE baselines, with gains of up to 20.7% over GEPA and 21.2% over ACE. The paper (arXiv 2610.37651) is by researchers including Zack Yu, Matei Zaharia, Michael W. Mahoney, Kurt Keutzer, and Amir Gholami.

Paper Overview

Research area: NLP / Agents Authors: Zack Yu, Nicholas Lee, Minjae Lee, Lakshya Agarwal, Luca Manolache, Wonjun Kang, Matei Zaharia, Michael W. Mahoney, Kurt Keutzer, Sehoon Kim, Amir Gholami Posted: 2026-10-27 arXiv: 2610.37651

Abstract (translated from the original)

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

\tau

-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.

Key Takeaways

Failures in agentic rollouts are not always informative: they may reflect harness limitations, stochastic execution, or noisy judges rather than prompt weaknesses.
Curating which tasks reach the prompt optimizer is essential for reliable self-improvement.
Task curation filters out tasks likely to yield noisy or unreliable learning signals.
Prior success retrieval supplies relevant past successful trajectories as extra context to the optimizer.
Evaluated on GoBrowse, ALFWorld, AppWorld, and $\tau$ -bench with Gemini 3.1 Pro and Claude Opus 5 as optimizers; up to +20.7% over GEPA and +21.2% over ACE.

--- *Auto-collected on 2026-10-29*