Writing study: GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Lakshya A Agrawal, Shangyin Tan, Dilara Soylu, Noah Ziems, Rishi Khare, Krista Opsahl-Ong, Arnav Singhvi, Herumb Shandilya, Michael J Ryan, Meng Jiang, Christopher Potts, Koushik Sen, Alex Dimakis, Ion Stoica, Dan Klein, Matei Zaharia, Omar Khattab · Official ICLR 2026 oral record

Observed writing choices, not acceptance rules. This study examines the inspected manuscript version. It does not establish why the paper was accepted or that its presentation choices should be copied unchanged.

Figure, table and section ledger

Editorial contribution type: Learning & optimization algorithms (primary) · Inference, algorithms & systems efficiency.

Layout extraction, not a required paper template. Counts are distinct caption labels; a multi-panel figure counts once. Section spans are estimated from heading positions, including floats and whitespace. Method and limitations.

Canonical PDF

87 physical pages

Before references / appendix

≈ 9.29 page-equivalents

Main figures / tables

4 / 3

Attached appendix figures / tables

23 / 1

Attached appendix span

≈ 70.89 page-equivalents

Before references / appendix: ≈ 10.68%References / intervening material: ≈ 7.84%Attached appendix: ≈ 81.48%Estimated layout envelopes, not text/artwork area. Unresolved appendix boundaries remain within the undivided post-body material.

Section-by-section lengths and source pages

The title/abstract/front matter precede the first numbered heading. A span ends at the next detected top-level heading, not at the last sentence; end-page whitespace is included. References and statements are listed separately. These are not measurements of text-only or artwork area.

Top-level heading	Starts on PDF page	Estimated span	Share of entire PDF
1 INTRODUCTION	1	≈ 0.89 pp.	1.02%
2 PROBLEM STATEMENT	2	≈ 1.99 pp.	2.29%
3 GEPA: REFLECTIVE PROMPT EVOLUTION	4	≈ 1.48 pp.	1.7%
4 EVALUATION	6	≈ 2.57 pp.	2.95%
5 EXTENDED APPLICATIONS OF GEPA	8	≈ 0.53 pp.	0.61%
6 RELATED WORK	9	≈ 0.49 pp.	0.56%
7 CONCLUSION	9	≈ 0.32 pp.	0.37%
ACKNOWLEDGMENTS	10	≈ 0.18 pp.	0.21%
REFERENCES	10	≈ 6.82 pp.	7.84%
A APPENDIX OUTLINE	17	≈ 0.32 pp.	0.37%
B USAGE OF LARGE LANGUAGE MODELS	17	≈ 0.12 pp.	0.14%
C GEPA’S REFLECTION AND PROMPT UPDATE META PROMPT	17	≈ 1.31 pp.	1.51%
D GEPA ALGORITHM AND METHODOLOGY DETAILS	18	≈ 0.99 pp.	1.14%
E GEPA FOR INFERENCE-TIME SEARCH (CONTD.)	19	≈ 0.82 pp.	0.94%
F GEPA FOR ADVERSARIAL PROMPT SEARCH (CONTD.)	20	≈ 1.77 pp.	2.03%
G EVALUATION SETUP (CONTD.)	22	≈ 2.72 pp.	3.13%
H RESULTS AND ANALYSIS (CONTD.)	25	≈ 1.64 pp.	1.89%
I PERFORMANCE VS. BUDGET (ROLLOUTS) CURVES	26	≈ 0.08 pp.	0.09%
J GENERALIZATION GAP	26	≈ 1.59 pp.	1.83%
K COST VS. PERFORMANCE ANALYSIS FOR OPTIMIZED SYSTEMS	28	≈ 1.26 pp.	1.45%
L GEPA SEARCH TREES	29	≈ 0.09 pp.	0.1%
M VISUALIZING THE ITERATIVE REFINEMENT ACHIEVED BY GEPA	29	≈ 8.86 pp.	10.18%
N EXAMPLES OF BEST PROMPTS FOR EVERY BENCHMARK	38	≈ 43.14 pp.	49.59%
O GEPA GENERATED PROMPTS FOR KERNEL GENERATION	81	≈ 5.71 pp.	6.56%
P NUMBER OF REFLECTION LM CALLS MADE BY GEPA DURING	87	≈ 0.47 pp.	0.54%

All detected figure and table locations

Main text

• Figure 1 · p. 1
• Figure 2 · p. 3
• Figure 3 · p. 4
• Table 1 · p. 6
• Table 2 · p. 7
• Table 3 · p. 8
• Figure 4 · p. 8

Attached appendix

• Figure 5 · p. 18
• Figure 6 · p. 19
• Figure 27 · p. 20
• Figure 7 · p. 21
• Figure 8 · p. 21
• Figure 9 · p. 25
• Figure 10 · p. 26
• Figure 11 · p. 27
• Figure 12 · p. 27
• Figure 13 · p. 27
• Figure 14 · p. 27
• Figure 15 · p. 28
• Figure 16 · p. 28
• Figure 17 · p. 29
• Figure 18 · p. 29
• Figure 19 · p. 30
• Figure 20 · p. 30
• Figure 21 · p. 30
• Figure 22 · p. 30
• Figure 23 · p. 31
• Figure 24 · p. 31
• Figure 25 · p. 31
• Figure 26 · p. 32
• Table 4 · p. 87

Argument and organization

1. Problem formulation and GEPA

   Defines compound-system parameters and rollout budget before explaining reflection, candidate selection and optional merge.

2. Evaluation: numbered observations

   Turns the result tables into six interpretable claims, with secondary analyses deferred to Appendix H.

3. Extended applications and appendix

   Separates inference-time search and adversarial search from held-out task adaptation; retains full instructions and search histories.

Problem, claim and contribution

The title says the method can outperform RL, while the formulation explicitly permits prompt-space versus weight-space optimization; this helps readers understand what is actually being compared.

Title; Section 2, pp. 3–4 · Main paper

Figures, captions and visual explanation

Figure 3 combines a concrete per-task score matrix, candidate lineage and mutation/merge branches. The acceptance loop explains both exploration and expensive validation rather than merely drawing model boxes.

Figure 3, p. 4 · Main paper

Tables, comparisons and evidence

Table 1 includes benchmark-specific scores and rollout budgets, and its bolding preserves the AIME exception. Numbered observations guide interpretation without requiring the reader to infer every claim.

Section 4, Table 1, p. 6 · Main paper

Appendix and reproducibility

The appendix opens with a roadmap and provides executable-level algorithm choices, split sizes, optimizer settings, costs, complete GEPA prompts and search trees. MIPRO examples are explicitly abridged, a useful distinction from a complete artifact.

Appendices A, C–D, G, L–P, pp. 17–87 · Main paper

Clarity and writing conventions

Some aggregate figures cover four tasks although the main tables cover six; the cost section also alternates between one-third of the size and one-third shorter. These are avoidable scope and wording ambiguities.

Tables 1–2, pp. 6–7; Figure 9, p. 25; Appendix K, p. 28 · Main paper

What to reuse—and what not to infer

Transferable lesson: Pair each headline advantage with its comparison unit, an exception-preserving table and a worked artifact showing what the optimizer actually changed.

Boundary: Rollouts are not equal to wall-clock time, tokens or monetary cost. Keep four-task plots distinct from six-task aggregates, and do not generalize GEPA+Merge improvements to Qwen3, where several results decline.

Apply this lesson using the argument and evidence worksheet.

Inspection record and sources

Checked

2026-09-19

Reading status

Full text + selected visuals inspected

Main text

Pages 1–10: abstract and Sections 1–7, acknowledgments; references on pp. 11–16 excluded from analytical reading.

Appendix / supplement

Pages 17–87: Appendices A–P, including algorithms, setup, all prompt examples and the image-only NPU prompt on p. 82.

Visually inspected PDF pages

4, 6, 82

• Official ICLR 2026 proceedings conference PDF; PDF creation metadata: Mon May 18 22:28:15 2026 EDT · 87 pages