GEPA-Curator: Curated Learning for Self Improving Agents
GEPA-Curator filters noisy tasks and retrieves successful trajectories to improve prompt optimization for agentic workloads
GEPA-Curator augments existing optimizers with task curation to exclude unreliable signals and prior success retrieval for better context. Evaluated on GoBrowse, ALFWorld, AppWorld, and τ-bench using Gemini 3.1 Pro and Claude Opus 5. Matches or exceeds baselines with gains of up to 20.7% over GEPA and 21.2% over ACE. Identifies failures from prompt weaknesses, harness limits, and noisy judges.
- Up to 20.7% performance gain
- Filters noisy or harmful learning signals
- Uses successful trajectories for additional context
- Evaluated with two frontier models
From the original article
1 min read at huggingface.co
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.
By Zack Yu, Nicholas Lee, Minjae Lee · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.