GEPA-Curator
An agent self-improvement method that adds task filtering and successful-experience retrieval modules on top of reflective prompt optimizers such as GEPA. It aims to filter noisy learning signals and use successful trajectories as positive references, thereby improving the quality of prompt optimization.
Timeline (last 90 days)
GEPA-Curator adds two modules to the original prompt optimizer GEPA: task curation and prior success retrieval.
GEPA-Curator's task curation component determines, before optimization, which tasks are likely to provide noisy, unreliable learning signals and excludes them from the optimization process.
The paper evaluated GEPA-Curator on four agent benchmarks: GoBrowse, ALFWorld, AppWorld, and τ-bench.
The optimizer models used in the GEPA-Curator experiments are Gemini 3.1 Pro and Claude Opus 5.
GEPA-Curator matches or exceeds both baselines, GEPA and ACE, in average performance.
Compared with GEPA, GEPA-Curator achieves performance improvements of up to 20.7%.
Compared with ACE, GEPA-Curator achieves performance improvements of up to 21.2%.
The benefits of GEPA-Curator hold across two different optimizer models and multiple heterogeneous environments, indicating that its gains do not depend on a particular model or a single task type.
All Facts (8)
GEPA-Curator adds two modules to the original prompt optimizer GEPA: task curation and prior success retrieval.
50%UnverifiedGEPA-Curator's task curation component determines, before optimization, which tasks are likely to provide noisy, unreliable learning signals and excludes them from the optimization process.
50%UnverifiedThe paper evaluated GEPA-Curator on four agent benchmarks: GoBrowse, ALFWorld, AppWorld, and τ-bench.
50%UnverifiedThe optimizer models used in the GEPA-Curator experiments are Gemini 3.1 Pro and Claude Opus 5.
50%UnverifiedGEPA-Curator matches or exceeds both baselines, GEPA and ACE, in average performance.
50%UnverifiedCompared with GEPA, GEPA-Curator achieves performance improvements of up to 20.7%.
50%UnverifiedCompared with ACE, GEPA-Curator achieves performance improvements of up to 21.2%.
50%UnverifiedThe benefits of GEPA-Curator hold across two different optimizer models and multiple heterogeneous environments, indicating that its gains do not depend on a particular model or a single task type.
50%