- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper presents GEPA-Curator, a framework designed to improve reflective prompt optimization by curating the experiences used for learning. The authors argue that failures in agentic environments do not always indicate prompt deficiencies, as they may instead arise from environment limitations, faulty tasks, or noisy evaluations. Learning from such failures can lead to unnecessary or harmful prompt mutations.
To address this issue, GEPA-Curator introduces two components. Task Curation selects tasks that are likely to provide useful optimization signals based on their execution histories, while Prior Success Retrieval supplies reliable historical successful trajectories to the Reflector as additional context for diagnosing current failures.
Experiments on GoBrowse, ALFWorld, AppWorld, and τ-bench demonstrate consistent improvements over GEPA and ACE across two optimizer models. The authors further extend the approach to Meta-Harness optimization and report additional improvements.
Critical strengths
-
The paper identifies an important and practical problem in reflective prompt optimization. The observation that agent failures may originate from environment limitations or evaluation noise rather than prompt weaknesses is well motivated.
-
Both task curation and prior success retrieval can be implemented as an additional layer without fundamentally changing the underlying optimization process. The use of historical successful trajectories is particularly intuitive and potentially useful for improving reflection quality.
-
The experimental results are promising across multiple agentic benchmarks. The ablation study also suggests that the two proposed components are complementary.
Critical weaknesses
-
The reliance on strong LLMs raises concerns about cost and the actual source of improvement. The paper uses powerful models like Gemini and Claude for task filtering and prior success retrieval. This introduces additional inference costs and makes it unclear whether the gains come from the proposed curation strategy or simply from the strong models’ reasoning abilities. More detailed ablations, including weaker Curator models and cost-matched baselines, would help clarify this.
-
The quality of task-selection decisions is not sufficiently evaluated. The curator aims to identify tasks that provide useful learning signals, but the paper mainly reports improvements in final performance. It remains unclear how often useful tasks are skipped or noisy tasks are retained. More importantly, the paper does not show whether the selected tasks actually lead to better prompt updates.
-
The ablation studies are limited. The main ablation is conducted only on GitLab, making it unclear whether both components are equally effective across other benchmarks. Moreover, the Meta-Harness experiments use only one optimization run per setting, which makes the reported improvements less convincing.
Questions for authors
-
How much additional inference cost does the Curator introduce? Could the authors report the additional inference costs? How would GEPA-Curator compare with GEPA under the same total inference budget, rather than just the same rollout budget?
-
Why filter tasks instead of filtering unreliable failures? Since the main concern is that the Reflector may learn from misleading failures, why not let it identify and ignore environment-related or evaluation-related failures directly after execution? How would such a failure-aware approach compare with the proposed task-level curation?
5 comments
Response to Review
2. Cost Analysis
How much additional inference cost does the Curator introduce? Could the authors report the additional inference costs?
The motivation of this work is to establish whether an autonomous agent can curate its own learning experiences. We therefore prioritized demonstrating the effectiveness of autonomous curation, and did not initially optimize the frequency of Curator calls or the amount of historical context processed by each call. These implementation choices leave room to improve efficiency.
We introduced three optimization methods that can be applied to GEPA-Curator without score degradation:
- Decide the skip interval once. The Curator specifies how many epochs to skip a task, avoiding repeated Curator calls during that interval.
- Compact older histories. Keep complete trajectories for recent execution units and summarize older entries using their epochs, outcomes, and previous Curator assessments. We retain the latest four execution units per task on WebArena and the latest one on the non-WebArena benchmarks.
- Reusing successful reference decisions. Preserve an optional full successful trajectory across history compaction and supply it to the Reflector when a fresh execution fails. The Curator can replace this reference with a better success from recent executions.
| Method | GitLab | Shopping | Shopping Admin | |
|---|---|---|---|---|
| GEPA-Curator | 84.6 ± 1.5 | 85.8 ± 3.5 | 73.1 ± 1.9 | 85.0 ± 1.8 |
| + Optimizations | 83.8 ± 2.4 | 85.1 ± 1.4 | 73.3 ± 1.6 | 86.8 ± 2.8 |
| Method | ALFWorld | AppWorld Test-N | AppWorld Test-C | τ-bench |
|---|---|---|---|---|
| GEPA-Curator | 91.8 ± 1.4 | 58.7 ± 1.9 | 46.2 ± 2.1 | 84.0 ± 4.8 |
| + Optimizations | 90.0 ± 1.0 | 67.1 ± 3.9 | 56.3 ± 1.5 | 84.0 ± 3.3 |
As we can see, the performance of GEPA-Curator using these optimizations is on-par with the previous results, so for the final paper, we will use make these part of the default method.
Below, we compare the costs between GEPA, GEPA-Curator, and ACE averaged across Webarena and non-Webarena benchmarks. We report the estimated monetary inference costs accounting for all of the computational expense for the input and output tokens consumed by all of the actor rollouts, reflection, and curation. This also captures the additional inference overhead from the Curator calls and potential longer reflection context lengths.
| Benchmark group | Method | Estimated optimization cost (USD) | Average test success (%) |
|---|---|---|---|
| WebArena | GEPA | 594.05 | 78.7 |
| WebArena | GEPA-Curator | 776.88 | 82.3 |
| Non-WebArena | GEPA | 132.65 | 55.9 |
| Non-WebArena | ACE | 768.69 | 65.2 |
| Non-WebArena | GEPA-Curator | 173.98 | 74.4 |
Compared with GEPA, GEPA-Curator improves average success by 3.6% on WebArena and 18.5% on the non-WebArena settings, at approximately 1.31× the cost in each group. Compared with ACE, it achieves 9.2% higher average success on the non-WebArena settings with 4.42x less cost than ACE (\768.69).
We separate rollout, reflection, and curation costs below:
| Benchmark group | Method | Rollouts (USD) | Reflection (USD) | Curation (USD) | Total (USD) |
|---|---|---|---|---|---|
| WebArena | GEPA | 552.77 | 41.27 | — | 594.05 |
| WebArena | GEPA-Curator | 578.47 | 98.87 | 99.54 | 776.88 |
| Non-WebArena | GEPA | 115.90 | 16.75 | — | 132.65 |
| Non-WebArena | ACE | 175.43 | 593.25 | — | 768.69 |
| Non-WebArena | GEPA-Curator | 97.81 | 31.55 | 44.62 | 173.98 |
We can see based on this breakdown that while the cost of Reflection and Curation is higher for GEPA-Curator, the majority of the cost is associated with the actor rollout cost. In these experiments, we have used gpt-oss-120b, which has extremely cheap API costs compared to normal frontier models. We expect that the relative overhead of GEPA-Curator to decrease when we use more expensive actor models, since rollout costs would begin to dominate.
Response to Review
3. Matched Costs Experiments
How would GEPA-Curator compare with GEPA under the same total inference budget, rather than just the same rollout budget?
We also run experiments with matched costs, where we stop the GEPA-Curator training strictly before cost exceeds GEPA. The experiments uses gpt-oss-120b as the actor model and Gemini-3.1-pro as all optimizer models. The results are shown in the table below:
| Benchmark | GEPA-Curator test (%) | GEPA test (%) | GEPA-Curator cost ($) | GEPA cost ($) |
|---|---|---|---|---|
| GitLab | 87.3 | 82.1 | 141.23 | 144.01 |
| Shopping | 86.0 | 83.8 | 155.22 | 156.10 |
| 85.8 | 77.6 | 141.01 | 141.89 | |
| Shopping Admin | 70.7 | 71.5 | 150.38 | 152.05 |
| ALFWorld | 94.3 | 82.4 | 54.96 | 55.66 |
| AppWorld, Test-N / Test-C | 65.5 / 51.8 | 38.0 / 27.8 | 60.14 | 61.78 |
When matching total cost rather than rollout budgets, GEPA-Curator outperforms GEPA by an average of 11.2% in success rate over the six benchmarks.
4. Curation quality and whether selected tasks yield better updates
The quality of task-selection decisions is not sufficiently evaluated. The curator aims to identify tasks that provide useful learning signals, but the paper mainly reports improvements in final performance. It remains unclear how often useful tasks are skipped or noisy tasks are retained. More importantly, the paper does not show whether the selected tasks actually lead to better prompt updates.
We independently audit the Curator's run/skip decisions using Codex to examine execution histories and assess whether failures are prompt-addressable or uninformative. We report the accuracy of the Curator's decisions under these conditions below.
| Metric | Opus 5 | Gemini 3.1 Pro |
|---|---|---|
| Correctly identifying prompt-addressable failure | 92.57% | 86.13% |
| Correctly identifying uninformative failure | 95.31% | 83.82% |
| Overall accuracy | 93.94% | 84.98% |
Both models generally select the appropriate action, with Opus 5 being better at task curation than Gemini 3.1 Pro. Opus 5 achieves an overall accuracy of 93.94% and Gemini 3.1 Pro achieved 84.98%. Details of how this metric is constructed can be found in Appendix H.
As shown in Section 5.1, retaining already mastered tasks does not affect the optimization on optimizable tasks. In the mastered + optimizable task composition, performance on optimizable tasks consistently improves and approaches saturation. Instead, the inclusion of impossible or flaky tasks is what harms optimization. These findings suggest that the primary benefit of task curation comes from filtering out impossible and flaky tasks rather than removing already-mastered tasks.
In the table below, we run an ablation using just the task curation component of GEPA-Curator and compare it to GEPA on AppWorld and Gitlab. We can see that using task curation matches or improves over GEPA on both datasets, which shows us that task curation can lead us to better prompt updates.
| Method | GitLab (%) | AppWorld Test-N (%) | AppWorld Test-C (%) |
|---|---|---|---|
| GEPA | 82.8 ± 1.6 | 38.0 ± 2.3 | 27.8 ± 2.1 |
| Task selection only | 84.2 ± 2.9 | 46.9 ± 2.1 | 27.7 ± 2.1 |
Response to Review
We appreciate the reviewer for taking the time to review our work and providing us with their valuable feedback.
1. Weak curator model experiment
Thank you for suggesting this experiment. We have run experiments using Gemini 3.5 Flash-Lite for the Actor, Reflector, and Curator. We chose this smaller model to demonstrate Curation is not restricted to strong optimizer models, and to ensure affordability.
AppWorld:
WebArena:
On AppWorld, GEPA-Curator improves over GEPA by 7.7% on Test-N and 11.0% on Test-C. On WebArena, GEPA-Curator improves the average success rate by 1.7%. Thus, even with smaller optimizer models, GEPA-Curator provides considerable gains over GEPA. This demonstrates how the benefit persists when the actor and optimizer use the same underlying model and does not require a stronger teacher model.