PaperlayerView exact revision on arXiv

Source-linked AI summary

GEPA-Curator: Curated Learning for Self Improving Agents

Zack Yu, Nicholas Lee, Minjae Lee, Lakshya Agarwal, Luca Manolache, Wonjun Kang, Matei Zaharia, Michael W. Mahoney, Kurt Keutzer, Sehoon Kim, Amir Gholami

arXiv:2610.37651v1cs.CL

TL;DR

Reflective prompt optimization can mistake failures caused by harness, environment, execution, or evaluation noise for prompt weaknesses, making task selection important. GEPA-Curator curates tasks and retrieves prior successful trajectories as context; it matches or exceeds baseline mean performance across evaluated benchmarks, with gains up to 20.7% over GEPA and 21.2% over ACE.

  • Problem

    Agentic failures can stem from prompt weaknesses or from harness, environment, execution, and evaluation noise, so failed rollouts may provide unreliable learning signals.

    Source: §1 · p. 1

  • Method

    GEPA-Curator selects tasks likely to provide useful optimization signals and selectively supplies prior successful trajectories as context for failed tasks.

    Source: §1 · p. 2

  • Results

    GEPA-Curator matches or exceeds baseline mean performance, with gains of up to 20.7% over GEPA and 21.2% over ACE across evaluated agentic benchmarks.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The results suggest that curation benefits extend across optimizer models and benchmark settings.

    Source: §1 · p. 2, §4.2 · p. 6

    The method is generic rather than domain-specific, and optimized agents in high-risk environments require appropriate human oversight, monitoring, and controls.

    Source: ETHICS STATEMENT · p. 11

Abstract

from arXiv · show

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.

1 INTRODUCTION

Agentic failures can reflect prompt weaknesses or noisy execution and evaluation conditions, making them unreliable signals for reflective prompt optimization. GEPA-Curator filters tasks and retrieves prior successes; evaluations show gains over existing optimizers across agentic benchmarks.

1 INTRODUCTION

  • Agentic failures may arise from unavailable tools, harness limitations, execution or evaluation noise, or ambiguous and infeasible tasks, not just prompt weaknesses.Treating such failures as prompt deficiencies can lead to unnecessary or harmful updates.Source: §1 · p. 1
  • Validation-based gating can accept bad mutations or reject good ones under noisy aggregate scores, and it acts only after a mutation is generated.The authors motivate adding contextual information to improve reflection and mutation quality.Source: §1 · p. 2
  • The Curator selects tasks likely to provide clean optimization signal and retrieves past successful trajectories as context for failed tasks.These two roles are intended to help the Reflector distinguish current failures from prior successful behavior.Source: §1 · p. 2
  • GEPA-Curator matches or exceeds baseline mean performance, with gains of up to 20.7% over GEPA and 21.2% over ACE.The evaluation covers four WebArena websites, ALFWorld, AppWorld, and τ-bench.Source: §1 · p. 2

2 RELATED WORK

Related work spans tool-using agents and benchmarks, reflective prompt optimizers, and curriculum learning. GEPA-Curator focuses on which trajectories provide valid reflection signals and what historical evidence should accompany them.

LLM Agents

  • Tool calling and reasoning-tool interaction support agents that act in external environments, with benchmarks evaluating tool use across varied settings.WebArena is described as an early web benchmark covering realistic sites such as GitLab, Reddit, and shopping portals.Source: §2 · p. 2

Prompt Optimization

  • Reflective optimizers revise textual system components using natural-language feedback, while GEPA applies reflection to system-level trajectories within evolutionary search.ACE, SkillOpt, and Meta-Harness adapt agent context, skills, or harness code, respectively; GEPA-Curator asks which trajectories are valid signals and what historical evidence to provide.Source: §2 · pp. 2–3

Curriculum Learning

  • Curriculum learning studies how selecting and ordering training data can affect optimization, and prior LLM work reports that poorer-quality data can negatively affect alignment.LIMA showed that a small amount of high-quality data can be as effective as a large-scale training set.Source: §2 · p. 3

3 GEPA-CURATOR METHODOLOGY

GEPA-Curator runs as an outer harness that uses per-task histories to select minibatch tasks and optionally supply trusted successful trajectories. Task skips are temporary, and retrieval favors successes supported by consistent evaluations.

3 GEPA-CURATOR METHODOLOGY

  • GEPA-Curator maintains per-task trajectories and evaluation outcomes, using them to curate tasks and retrieve past successes as metadata without changing the optimizer.The Curator and optimizer form an outer and inner workflow, with new outcomes returned to task history.Source: §3 · p. 3, Figure 1 · p. 4

3.1 TASK CURATION

  • The Curator initially prioritizes underexplored tasks, then may skip tasks with stable success, external causes of failure, or inconsistent evaluations.Skipping is temporary: tasks are reconsidered each epoch because solved tasks may regress and previously difficult tasks may become solvable as the prompt changes.Source: §3.1 · p. 3

3.2 PRIOR SUCCESS RETRIEVAL

  • For failed tasks, the Curator selects a prior successful trajectory as a reference only when evaluation history supports it, and otherwise abstains.The reference gives the Reflector a concrete comparison for diagnosing failure rather than requiring it to guess the fix.Source: §3.2 · p. 4
  • Task curation and prior-success retrieval jointly aim to provide cleaner training data and grounded context for prompt mutations.The intended signal comes from removing noisy or redundant tasks and supplying reliable successful trajectories.Source: §3.2 · p. 4

4 EMPIRICAL EVALUATION

The evaluation compares GEPA-Curator with GEPA and ACE across WebArena and additional agentic benchmarks, then examines its components in GitLab ablations.

4.1 EXPERIMENTAL SETUP

  • Experiments used a shared gpt-oss-120b actor, 16-task minibatches, and five repeated runs; evaluation covered four GoBrowse websites, ALFWorld, AppWorld, and τ-bench.The authors describe held-out test splits for GoBrowse and additional benchmark-specific training, validation, and test procedures.Source: §4.1 · p. 5

4.2 MAIN RESULTS

  • 3.4% and 5.0% aggregate held-out test improvements over GEPA across WebArena domains for Gemini 3.1 Pro and Claude Opus 5, respectively.GEPA-Curator's gains reached 7.4% for both optimizer models.Source: §4.2 · p. 5, Table 1 · p. 6
  • 14.3% over GEPA and 5.0% over ACE were Gemini 3.1 Pro's average gains across four additional agentic test sets; Claude Opus 5's averages were 8.1% and 12.6%, respectively.With Claude Opus 5, GEPA-Curator improved over GEPA on ALFWorld and both AppWorld test sets, while matching mean success on τ-bench.Source: Table 2 · p. 6, §4.2 · p. 6

4.3 ABLATIONS

  • 88.6% was the strongest GitLab ablation result, achieved by combining task selection and success selection.Task selection alone beat GEPA by 1.4%, and success selection alone by 1.9%.Source: Table 3 · p. 6, §4.3 · pp. 6–7
  • Figure 2 tracks training-set solved tasks within 16-task groups every four iterations as Optimizable tasks are mixed with Easy, Impossible, or Flaky tasks.Labels mark initial, peak, and final performance.Source: Figure 2 · p. 7

5 DISCUSSION

Across the motivating experiments and extensions, curation improves optimization outcomes, while the task-selection analysis identifies noisy or unreliable tasks that the Curator filters. The section also reports comparable held-out performance for an efficiency-optimized variant at substantially lower cost than ACE on non-WebArena benchmarks.

5.1 MOTIVATING EXPERIMENTS

  • 12/16 peak accuracy: GEPA reaches near-perfect accuracy with Optimizable plus Easy tasks, but peaks at 12/16 when Impossible or Flaky tasks are added.The authors interpret these results as evidence that Impossible and Flaky task signals can contaminate prompts and hinder optimization.Source: §5.1 · p. 7
  • 70.2% success: GEPA-Curator exceeds task curation alone (67.4%), GEPA (66.2%), and numerical curriculum learning (62.5%) on the flaky-task composition.The Table 4 success rates average five evaluations of each optimized prompt; task curation alone omits prior success retrieval.Source: Table 4 · p. 7, §5.1 · pp. 7–8

5.2 ANALYSIS OF TASK SELECTION

  • 75 of 1,581 GoBrowse training tasks (4.7%) were diagnosed as impossible or flaky at least once; ambiguous or faulty descriptions were the most common flaky-task cause.Harness or environment limitations accounted for most impossible diagnoses and a substantial share of flaky diagnoses.Source: §5.2 · p. 8, Figure 3 · p. 8

5.3 META-HARNESS-CURATOR

  • GEPA-Curator extends curation to Meta-Harness, where the optimizer modifies agent harness code; the authors report modest held-out performance improvements.The Curator selects a parent harness and training tasks, and may provide prior successful executions as references when fresh parent executions fail.Source: §5.3 · pp. 8–9
  • On Meta-Harness, curation raises Math-RAG test accuracy from 45.33% to 46.33% and mean reward from 45.69% to 47.26%, while text-classification accuracy rises from 48.06% to 51.04%.Classification gains are concentrated in LawBench and Symptom2Disease, while USPTO accuracy is slightly lower.Source: §5.3 · p. 9, Table 5 · p. 9, Table 6 · p. 9

5.4 EFFICIENCY OPTIMIZATION AND COST ANALYSIS

  • +3.6% on WebArena and +18.5% on non-WebArena benchmarks: Efficient GEPA-Curator exceeds GEPA with a 1.31× increase in total optimization cost.On non-WebArena benchmarks it costs $173.98, 4.42× less than ACE while achieving +9.2% higher average accuracy; its held-out performance remains comparable to GEPA-Curator across benchmarks.Source: §5.4 · pp. 10–11, Table 8 · p. 10, Table 7 · p. 10

6 CONCLUSION

The paper concludes that GEPA-Curator augments prompt optimization by selecting tasks and retrieving prior successes to shape reflection. It reports improvements over GEPA and ACE across several benchmarks, while noting that high-risk deployments require oversight and controls.

6 CONCLUSION

  • GEPA-Curator selects and filters tasks for reliable training signals and can provide prior successful trajectories as context for reflections.The conclusion reports gains of up to 21.2% over GEPA and ACE on WebArena, ALFWorld, AppWorld, and τ-bench, and an extension to harness-code optimization.Source: §6 · p. 11
  • The paper uses LLMs to curate data, retrieve successful trajectories, and propose prompt improvements; standard LLM tools also assisted with writing.Models, datasets, hyperparameters, and prompts are documented in Appendix B and Section 4.1.Source: AI USE STATEMENT · p. 11, REPRODUCIBILITY STATEMENT · p. 11
  • The method is generic prompt improvement for LLM agents, not a system designed for a specific domain.The authors advise evaluating optimized agents in their intended environments and using appropriate human oversight, monitoring, and controls.Source: ETHICS STATEMENT · p. 11

A.1 SELECTION OF DATASET COMPOSITION

The appendix explains how preliminary runs and manual inspection formed the motivating task categories, then illustrates negative transfer from impossible-task feedback. It also describes a numerical curriculum that removes persistently failing tasks from sampling.

A.1 SELECTION OF DATASET COMPOSITION

  • Easy tasks were selected for consistently high seed-prompt success, while Optimizable tasks had low seed-prompt success but consistent success under a preliminary optimized prompt.Impossible tasks had the lowest success across rollouts; Flaky tasks had roughly 50% success with both seed and optimized prompts, followed by manual inspection for inconsistent judgments on similar end states.Source: §A.1 · p. 16

A.2 A CONCRETE EXAMPLE OF NEGATIVE TRANSFER

  • An exhaustive-search mutation accepted with a minibatch score increase from 8 to 9 caused a solvable WCAG task to exhaust all 30 actions without reaching its guide.The mutation replaced a short repository-level search path with broad exploration; the example shows how difficult-task feedback can overgeneralize and mask a regression on a solvable task.Source: §A.2 · p. 16

A.3 IMPLEMENTATION OF NUMERICAL CURRICULUM

  • The numerical curriculum increments a task score on failure, reduces it by k on success down to zero, and removes the task once its score exceeds threshold b.This permits occasional failures but removes tasks that repeatedly fail from subsequent sampling.Source: §A.3 · p. 17

B IMPLEMENTATION DETAILS

The implementation appendix specifies model deployments, Curator decision criteria, seed-agent instructions, and GoBrowse judge prompts. The Curator uses accumulated evidence to decide whether tasks merit another rollout and whether a historical success is trustworthy enough to retrieve.

B.1 MODEL DEPLOYMENT

  • GoBrowse deployments used gpt-oss-120B and Qwen3-8B on separate H100 and H200 machines; the authors note the H200 runs may have an execution-speed advantage and a 0.7% higher seed score.Gemini 3.1 Pro used the hosted API with temperature 1.0 and high thinking; Claude Opus 5 used its API with temperature 1.0 and xhigh reasoning.Source: §B.1 · p. 17

B.2 CURATOR PROMPTS

  • The Curator runs under uncertainty, prioritizes unseen or underexplored tasks, and skips only when history strongly indicates little diagnostic or optimization value; skipped tasks are reconsidered later.Mastery requires multiple stable clean successes, prompt sensitivity is indicated by before/after improvement, and operational impossibility or absent optimization signal requires evidence beyond a single claim or unrelated failed mutations.Source: §B.2 · p. 18
  • A retrieved success must be correct and reproducible, supported by at least one other independent clean success; it is shown to the Reflector only if the current task is run and fails.The Curator jointly makes task-selection and success-selection decisions in one LLM call.Source: §B.2 · p. 18

B.3 SEED PROMPTS

  • Seed prompts prescribe environment-specific actions: GoBrowse agents inspect DOM state and verify changes, ALFWorld agents use admissible commands, and AppWorld agents call APIs through Python.AppWorld instructions include verifying changes and completing tasks through the supervisor API; the general agent prompt says to follow environment policy and avoid inventing tool results.Source: §B.3 · pp. 19–20

B.4 JUDGE PROMPTS

  • The GoBrowse judge evaluates information seeking, navigation, and content modification from user intent, action history, final page state, and agent response where relevant.The appendix also identifies tables listing primary causes of impossible and flaky/broken-evaluator diagnoses, and tasks diagnosed as impossible or flaky; diagnosis counts may overlap.Source: §B.4 · p. 20, Table 9 · p. 21, Table 10 · p. 21

C.1 ANALYSIS OF TASK SELECTION

GEPA-Curator identifies impossible or flaky tasks that may yield unreliable optimization signals, then adjusts how often they are sampled rather than retiring them permanently.

C.1 ANALYSIS OF TASK SELECTION

  • Most impossible diagnoses involved harness or environment limits, while faulty tasks omitted necessary information or specified infeasible or unsatisfiable objectives.Speculative diagnoses and those later overturned by evidence were also retained.Source: §C.1 · p. 21
  • Among flaky evaluations, faulty or underspecified tasks were the largest group, followed by harness or environment limits; four reproducibly successful behaviors received occasional false negatives.Those false negatives came from evaluators applying harsher or shifting criteria.Source: §C.1 · p. 21
  • 75 of 1,581 training tasks (4.7%) were diagnosed as impossible or flaky at least once; withholding them keeps potentially misleading outcomes from influencing those optimization steps.The diagnosis categories overlap on eight tasks, with 18 impossible and 65 flaky diagnoses.Source: §C.1 · pp. 21–22
  • Skipped tasks are reconsidered in later epochs, with impossible tasks deferred longest, flaky tasks revisited more often, and mastered tasks periodically checked for regressions.Mean skip streaks for impossible tasks increase from 5.50 to 8.75 epochs; accumulated histories generally lengthen skip streaks.Source: §C.1 · p. 22, Table 12 · p. 22
  • Under uniform sampling, a 16-task minibatch has about a 54% chance of containing at least one task diagnosed as impossible or flaky.The passage estimates this using the 4.7% diagnosed-task share; such tasks also fail more often than the rest of the training set.Source: §C.1 · p. 22

C.2 ANALYSIS OF SUCCESS SELECTION

The Curator sometimes selects reliable past successes as reflection anchors and sometimes abstains when available successes appear untrustworthy; these decisions are evaluated across eligible tasks and reflection instances.

C.2 ANALYSIS OF SUCCESS SELECTION

  • Historical successes were paired with a rising share of failed reflection instances, from 6.9% early to approximately 50% in the middle and late phases.Among late failures, 31.1% had no historical success and 18.6% had an available pool but no selected anchor.Source: §C.2 · p. 22, Table 13 · p. 23
  • The Curator selected a reliable success state in 17 of 554 eligible tasks (3.1%) where the pool also contained at least one noisy success state.Eligibility required a failed rollout entering reflection while a historical success was available.Source: §C.2 · p. 22, Table 14 · p. 23
  • Meaningful abstention occurred for 50 tasks (9.0%) when the Curator judged available score-positive states untrustworthy, contradictory, contaminated, evaluator-lenient, or unsupported by a reproducible success mode.The count excludes output-validation failures and cases where a single success simply lacked a second corroborating rollout.Source: §C.2 · p. 22, Table 14 · p. 23
  • 64 of 554 eligible tasks (11.6%) received a semantically meaningful success-pool decision, combining relevant reliable-state selection and meaningful abstention.The categories overlap for three tasks because the Curator made judgments at different iterations as evidence accumulated.Source: §C.2 · p. 23, Table 14 · p. 23

D DETAILS OF COST ANALYSIS

The cost analysis separates estimated optimization costs by benchmark group and by rollouts, reflection, and curation.

D DETAILS OF COST ANALYSIS

  • Estimated optimization costs are separated into rollouts, reflection, and curation, with overall rows summing costs within each benchmark group.Rollout costs include WebArena judge calls.Source: Table 15 · p. 24
  • For ACE, the Reflector cost column combines reflection and playbook curation.The Curator column covers task scheduling and success-anchor updates.Source: Table 15 · p. 24
  • The reported GEPA-Curator costs refer to the version optimized for efficiency.Totals are computed before rounding.Source: Table 15 · p. 24

E QA-TYPE BENCHMARKS FROM THE GEPA PAPER

On the QA benchmarks, GEPA-Curator matches or exceeds GEPA on three of four tasks and has a higher displayed average score. The appendix also lists licenses for the evaluated tools and datasets.

E QA-TYPE BENCHMARKS FROM THE GEPA PAPER

  • The evaluation uses matched runs with a Qwen3-8B actor and Gemini 3.1 Pro Preview Reflector; QA tasks use task curation without selected-success evidence.The passage specifies training, validation, and test splits for AIME 2025, IFBench, HoVer, and HotpotQA.Source: E QA-TYPE BENCHMARKS FROM THE GEPA PAPER · p. 24
  • 57.9 versus 56.8 average benchmark score: GEPA-Curator leads GEPA by +1.1% and matches or outperforms it on three of four QA benchmarks.It gains 4.3% on HoVer and underperforms by 1.3% on HotpotQA; the passage notes a smaller improvement than on WebArena.Source: E QA-TYPE BENCHMARKS FROM THE GEPA PAPER · p. 24, Table 16 · p. 25

F LICENSES

  • The appendix lists licenses for the evaluated tools and benchmarks, including MIT, Apache, and CC BY-SA terms.It notes additional redistribution requirements for protected bundled portions of AppWorld.Source: F LICENSES · p. 25

Related papers

Loading 2610.37651v1…