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

cs.CLOct 27, 2026

GEPA-Curator: Curated Learning for Self Improving Agents

Authors: Zack Yu, Nicholas Lee, Minjae Lee, Lakshya Agarwal, Luca Manolache, Wonjun Kang, Matei Zaharia, Michael W. Mahoney,

Organizations: UC Berkeley · Furiosa AI · ICSI · LBNL ·

Abstract

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 $τ$ -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.

Figures & tables

Figure 1: GEPA-Curator system diagram. The prompt optimizer takes in a minibatch of tasks at each iteration and invokes the actor agent to execute them and performs reflective prompt optimization based on the resulting trajectories. The Curator augments the prompt optimizer by acting as an outer-loop process which traverses through the training dataset and constructs each minibatch to pass into the prompt optimizer. The Curator reads past execution trajectories for each task sampled from the training dataset and decides to run the task if doing so is likely to provide useful optimization signal (Section 3.1 ). The Curator can also pass prior successful trajectories for failed tasks as context to the prompt optimizer as a trusted reference example (Section 3.2 ). Once the curated training batch of selected tasks and successful context is constructed, it is passed to the prompt optimizer for execution and optimization. The resulting trajectories and evaluation outcomes are then written back to the task history for use in subsequent curation rounds.

Optimizer Model	Method	GitLab	Shopping	Shopping Admin	Reddit	Overall
Gemini 3.1 Pro	Seed	$78.2$ $\pm$ 2.5	$73.5$ $\pm$ 11.8	$65.2$ $\pm$ 6.2	$72.6$ $\pm$ 2.2	$72.4$ $\pm$ 2.4
	GEPA	$82.1$ $\pm$ 2.7	$83.8$ $\pm$ 2.5	$71.5$ $\pm$ 2.0	$77.6$ $\pm$ 2.3	$78.7$ $\pm$ 0.6
	GEPA-Curator	$\mathbf{84.6}$ $\pm$ 1.5	$\mathbf{85.8}$ $\pm$ 3.5	$\mathbf{73.1}$ $\pm$ 1.9	$\mathbf{85.0}$ $\pm$ 1.8	$\mathbf{82.1}$ $\pm$ 1.7
Claude Opus 5	Seed	$79.6$ $\pm$ 2.3	$73.9$ $\pm$ 2.5	$67.7$ $\pm$ 2.9	$71.3$ $\pm$ 1.0	$73.1$ $\pm$ 1.8
	GEPA	$82.5$ $\pm$ 1.8	$80.5$ $\pm$ 3.2	$72.9$ $\pm$ 2.4	$81.9$ $\pm$ 1.5	$79.5$ $\pm$ 1.2
	GEPA-Curator	$\mathbf{83.1}$ $\pm$ 3.9	$\mathbf{86.5}$ $\pm$ 3.9	$\mathbf{80.3}$ $\pm$ 2.9	$\mathbf{88.2}$ $\pm$ 1.6	$\mathbf{84.5}$ $\pm$ 1.0

Table 1: Comparison of the success rate (% with std) on 4 different domains (GitLab, Shopping, Shopping Admin, and Reddit) of the GoBrowse benchmark on the WebArena environment. We evaluate using seed prompt, as well as the prompt optimized by GEPA, and GEPA-Curator over held-out test set. GEPA-Curator exceeds the success rate of all baseline methods across the evaluated settings.

Optimizer Model	Method	ALFWorld	AppWorld Test-N	AppWorld Test-C	$\tau$ -bench
–	Seed	$66.9$ $\pm$ 2.4	$8.9$ $\pm$ 2.8	$5.6$ $\pm$ 0.5	$52.0$ $\pm$ 4.2
Gemini 3.1 Pro	GEPA	$82.4$ $\pm$ 2.9	$38.0$ $\pm$ 2.3	$27.8$ $\pm$ 2.1	$75.4$ $\pm$ 5.2
	ACE	$90.1$ $\pm$ 1.3	$55.7$ $\pm$ 1.8	$34.5$ $\pm$ 1.4	$80.6$ $\pm$ 6.6
	GEPA-Curator	$\mathbf{91.8}$ $\pm$ 1.4	$\mathbf{58.7}$ $\pm$ 1.9	$\mathbf{46.2}$ $\pm$ 2.1	$\mathbf{84.0}$ $\pm$ 4.8
Claude Opus 5	GEPA	$95.8$ $\pm$ 1.4	$52.9$ $\pm$ 5.7	$32.5$ $\pm$ 2.7	$\mathbf{82.9}$ $\pm$ 2.0
	ACE	$89.7$ $\pm$ 2.7	$61.2$ $\pm$ 1.7	$33.4$ $\pm$ 0.8	$61.7$ $\pm$ 5.9

Table 2: Comparison of the success rate (% with std) on additional agentic benchmarks: ALFWorld, AppWorld, and

\tau

-bench. We evaluate using seed prompt, as well as the prompt optimized by GEPA, ACE, and GEPA-Curator over held-out test set. GEPA-Curator exceeds the success rate of all baseline methods across the evaluated settings with only one tie.

Method	Success (%)
Seed	$78.2$ $\pm$ 2.5
GEPA	$82.8$ $\pm$ 1.6
w/o Pareto selection	$82.5$ $\pm$ 1.9
GEPA-Curator (full)	$\mathbf{88.6}$ $\pm$ 2.0
Task curation only	$84.2$ $\pm$ 2.9
Prior success retrieval only	$84.7$ $\pm$ 3.1

Table 3: Ablation results (% with std) on 200 GitLab tasks.

Figure 2: GEPA validation curves when Optimizable tasks are mixed with Easy , Impossible , or Flaky tasks. Curves are evaluated on the training set every four iterations and report tasks solved within each 16-task group; labels mark initial, peak, and final performance.

Method	Success (%)
GEPA	66.2
Numerical curriculum	62.5
Task curation only	67.4
GEPA-Curator	70.2

Table 4: Success rate (%) on the dataset composition with Flaky tasks, averaged over five evaluations of each optimized prompt. Task curation only omits prior success retrieval, whereas GEPA-Curator uses both components.

Figure 3: Diagnoses made by the Curator across the 1,581 GoBrowse training tasks and the primary causes assigned to impossible and evaluator-flaky tasks. Eight tasks received both impossible and flaky diagnoses because their diagnoses changed across iterations. The overview counts these overlaps as impossible, leaving 57 unique flaky-only tasks, while panel (c) includes 65 flaky diagnoses.

Table 8

Method	GitLab	Shopping	Shopping Admin	Reddit
GEPA-Curator	$\mathbf{84.6}$ $\pm$ 1.5	$\mathbf{85.8}$ $\pm$ 3.5	$73.1$ $\pm$ 1.9	$85.0$ $\pm$ 1.8
Efficient GEPA-Curator	$83.8$ $\pm$ 2.4	$85.1$ $\pm$ 1.4	$\mathbf{73.3}$ $\pm$ 1.6	$\mathbf{86.8}$ $\pm$ 2.8
Method	ALFWorld	AppWorld Test-N	AppWorld Test-C	$\tau$ -bench
GEPA-Curator	$\mathbf{91.8}$ $\pm$ 1.4	$58.7$ $\pm$ 1.9	$46.2$ $\pm$ 2.1	$\mathbf{84.0}$ $\pm$ 4.8
Efficient GEPA-Curator	$90.0$ $\pm$ 1.0	$\mathbf{67.1}$ $\pm$ 3.9	$\mathbf{56.3}$ $\pm$ 1.5	$\mathbf{84.0}$ $\pm$ 3.3

Table 7: Comparison between GEPA-Curator and Efficient GEPA-Curator of the success rate (% with std) on 4 different domains (GitLab, Shopping, Shopping Admin, and Reddit) of the GoBrowse benchmark on the WebArena environment and ALFWorld, AppWorld Test-N, AppWorld Test-C, and

\tau

-bench.

Method	GitLab	Shopping	Shopping Admin	Reddit	Overall	Overall Acc.
GEPA	144.01	156.10	152.05	141.89	594.05	78.7
GEPA-Curator	210.64	186.46	201.27	178.51	776.88	82.3

Table 8: Estimated total optimization costs in USD. GEPA-Curator uses the efficiency-optimized version. AppWorld Test-N and Test-C share one optimization run. Overall Acc. (%) averages the four WebArena websites or the four non-WebArena evaluation settings, counting AppWorld Test-N and Test-C separately. The component breakdown is in Appendix D .

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Website	Training tasks	Impossible	Flaky
GitLab	$432$	$6\;(1.4\%)$	$14\;(3.2\%)$
Shopping	$351$	$5\;(1.4\%)$	$16\;(4.6\%)$
Shopping Admin	$449$	$4\;(0.9\%)$	$17\;(3.8\%)$
Reddit	$349$	$3\;(0.9\%)$	$18\;(5.2\%)$
Overall	$\mathbf{1{,}581}$	$\mathbf{18\;(1.1\%)}$	$\mathbf{65\;(4.1\%)}$

Appendix

Table 9: Tasks that the Curator deemed impossible or flaky at any point during optimization. Counts in the two diagnosis columns are not mutually exclusive.

Primary cause	Tasks	Share
Harness/environment limitation	$12$	$66.7\%$
Faulty or underspecified task	$4$	$22.2\%$
Speculative diagnosis	$1$	$5.6\%$
Reversed after revisit	$1$	$5.6\%$
Total	$\mathbf{18}$	$\mathbf{100\%}$

Appendix

Table 10: Primary causes of tasks that the Curator deemed impossible during optimization. Diagnoses are counted if they occurred at any point in the task history.

Primary cause	Tasks	Share
Harness/environment limitation	$19$	$29.2\%$
Faulty or underspecified task	$30$	$46.2\%$
Speculative diagnosis	$7$	$10.8\%$
Reversed after revisit	$5$	$7.7\%$
Reliable success with occasional false negatives	$4$	$6.2\%$
Total	$\mathbf{65}$	$\mathbf{100\%}$

Appendix

Table 11: Primary causes of tasks that the Curator deemed to have a flaky or broken evaluator. False negatives denote cases where reliable success behavior was sometimes rejected under harsher or shifting evaluation criteria.

Curator diagnosis	Early	Middle	Late
Mastered	$3.13$	$5.18$	$5.81$
Impossible	$5.50$	$7.67$	$8.75$
Flaky	$1.33$	$5.00$	$3.92$

Appendix

Table 12: Mean length of the completed skip streak immediately preceding a task rerun, grouped by rerun phase. Open streaks and reruns in partial final epochs are excluded.

	Early			Middle			Late
Website	Paired	No hist.	Abst.	Paired	No hist.	Abst.	Paired	No hist.	Abst.
GitLab	$5/144$	$111/144$	$28/144$	$174/296$	$51/296$	$71/296$	$88/132$	$18/132$	$26/132$
Shopping	$22/169$	$109/169$	$38/169$	$206/273$	$55/273$	$12/273$	$58/94$	$31/94$	$5/94$
Shopping Admin	$5/235$	$211/235$	$19/235$	$32/387$	$249/387$	$106/387$	$86/375$	$192/375$	$97/375$
Reddit	$18/180$	$119/180$	$43/180$	$194/267$	$29/267$	$44/267$	$167/193$	$6/193$	$20/193$
Overall	$\mathbf{6.9\%}$	$\mathbf{75.5\%}$	$\mathbf{17.6\%}$	$\mathbf{49.6\%}$	$\mathbf{31.4\%}$	$\mathbf{19.1\%}$	$\mathbf{50.3\%}$	$\mathbf{31.1\%}$	$\mathbf{18.6\%}$

Appendix

Table 13: Historical-success pairing among failed reflection instances. Paired indicates that a historical success was selected and supplied to reflection; No hist. (No history) indicates that the task had no prior successful trajectory; and Abst. (Abstention) indicates that a non-empty success pool existed but the Curator selected no anchor. Training iterations are divided into approximately equal thirds separately for each website.

Website	Eligible	Relevant selection	Meaningful abstention	Either
GitLab	$155$	$3\;(1.9\%)$	$10\;(6.5\%)$	$12\;(7.7\%)$
Shopping	$144$	$5\;(3.5\%)$	$9\;(6.3\%)$	$13\;(9.0\%)$
Shopping Admin	$102$	$4\;(3.9\%)$	$24\;(23.5\%)$	$27\;(26.5\%)$
Reddit	$153$	$5\;(3.3\%)$	$7\;(4.6\%)$	$12\;(7.8\%)$
Overall	$\mathbf{554}$	$\mathbf{17\;(3.1\%)}$	$\mathbf{50\;(9.0\%)}$	$\mathbf{64\;(11.6\%)}$

Appendix

Table 14: Task-level analysis of the Curator’s use of historical success states. Eligible tasks are those for which a failed rollout entered reflection while the task had a non-empty historical success pool. Either is the deduplicated union of relevant selection and meaningful abstention.

Benchmark	Method	Rollout	Reflector	Curator	Total
GoBrowse GitLab	GEPA	133.20	10.81	–	144.01
GoBrowse GitLab	GEPA-Curator	161.38	24.47	24.80	210.64
GoBrowse Shopping	GEPA	147.41	8.69	–	156.10
GoBrowse Shopping	GEPA-Curator	138.30	22.78	25.38	186.46
GoBrowse Shopping Admin	GEPA	140.69	11.36	–	152.05
GoBrowse Shopping Admin	GEPA-Curator	150.58	26.10	24.60	201.27

Appendix

Table 15: Estimated optimization costs in USD, separated into rollouts, reflection, and curation. Rollout costs include WebArena judge calls. Overall rows sum the costs within each benchmark group. Totals are computed before rounding. For ACE, the Reflector column combines reflection and playbook curation. GEPA-Curator refers to the version optimized for efficiency. The Curator column covers task scheduling and success anchor updates.

	AIME 2025	IFBench	HoVer	HotpotQA
Seed	$27.3$	$36.9$	$35.3$	$42.3$
GEPA	$64.7$	$45.2$	$51.0$	$\mathbf{66.3}$
GEPA-Curator	$\mathbf{65.3}$	$\mathbf{45.9}$	$\mathbf{55.3}$	$65.0$

Appendix

Table 16: Test performance (%) on QA benchmarks from the GEPA evaluation suite.

LLM Agent Self-Improvement

Large language model agent self-improvement studies how agents use experience and feedback to improve their behavior, tools, and workflows.

See all 276 papers

Environment-Grounded Automated Prompt Optimization for LLM Game Agents
Jun 16, 2026Rean Clive Fernandes, Lukas Fehring, Theresa Eimer +2Multi-Agent LLM Systems Game-Playing Agents
Lamarr institute for ML and AI TU Dortmund University · Leibniz University Hannover · L3S Research Center
RLMOpt: Adaptive Prompt Optimization via Recursive Language Models
Aug 11, 2026Subhash Bangalore Satheesha, Nirvik Pande, Deepthi Duddempudi +1Large Language Model-Guided Optimization LLM Agents
Autonomize AI · Carnegie Mellon University
CurateEvo: Data-Curation Evolving for Agentic Post-Training
Jul 7, 2026Dingzirui Wang, Xuanliang Zhang, Keyan Xu +2Training Data Curation LLM Agent Training
Harbin Institue of Technology

Jun 16, 2026J/K move · Enter open · S save

Environment-Grounded Automated Prompt Optimization for LLM Game Agents

Rean Clive Fernandes, Lukas Fehring, Theresa Eimer, Marius Lindauer+1

Lamarr institute for ML and AI TU Dortmund University · Leibniz University Hannover · L3S Research Center

Multi-Agent LLM Systems Game-Playing Agents Automatic Prompt Optimization LLM Agents

Open paper AskPDF arXiv

LLM agents in interactive environments are highly sensitive to their prompts, yet prompt engineering remains a manual, task-specific process. We introduce an automated prompt optimization framework for LLM agents that decomposes the observation-to-action pipeline into a goal-conditioned descriptor agent and an action selection agent, and iteratively refines each module's prompt through an LLM-driven evolutionary loop guided by environment returns. We propose a behavior analyzer to attribute episode outcomes to specific prompt components, and a mutator to propose targeted revisions to the prompt, before validating them through environment rollouts. We evaluate on all five BabyAI tasks in the BALROG benchmark, comparing our pipeline against BALROG's RobustCoTAgent under both plain and guided prompt initializations. Optimization improves performance consistently across tasks and conditions, without requiring updates to the model weights. On PutNext, a multi-step coordination task where the RobustCoTAgent achieves 0% success, our framework reaches up to 72.5% success rate using the same underlying LLM with optimized prompts. These results suggest that a multi-agent framework, combined with automatic prompt optimization, enhances LLMs without the need for fine-tuning or extensive human supervision.

Abstract

LLM Agent Self-Improvement

Explore similar work

Environment-Grounded Automated Prompt Optimization for LLM Game Agents

RLMOpt: Adaptive Prompt Optimization via Recursive Language Models

CurateEvo: Data-Curation Evolving for Agentic Post-Training