GEPA-Curator filters agent experiences before prompt optimization
Original title: GEPA-Curator: Curated Learning for Self Improving Agents
On four agent benchmarks it gains up to 20.7% over GEPA and 21.2% over ACE with Gemini 3.1 Pro and Claude Opus 5.
Researchers introduce GEPA-Curator, which chooses which agent experiences a reflective prompt optimizer should learn from. Task curation drops tasks likely to yield noisy signals, and prior success retrieval adds past successful trajectories as context. The paper says it matches or exceeds GEPA and ACE on GoBrowse, ALFWorld, AppWorld, and tau-bench.
- Agent failures can come from prompts, harness limits, stochastic execution, or noisy judges.
- Task curation filters experiences that may give unreliable learning signals.
- Prior success retrieval supplies successful trajectories as extra optimizer context.
- The optimizer models evaluated are Gemini 3.1 Pro and Claude Opus 5.
- AI models
- Gemini 3.1 ProClaude Opus 5
From the original article
1 min read at arxiv.orgLarge 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 · arxiv.org
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.