Writing study: AutoEP: LLMs-Driven Automation of Hyperparameter Evolution for Metaheuristic Algorithms

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Zhenxing Xu, Yizhe Zhang, Weidong Bao, Hao Wang, Ming Chen, Haoran Ye, Wenzheng Jiang, Hui Yan, Ji Wang · 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

19 physical pages

Before references / appendix

≈ 10.2 page-equivalents

Main figures / tables

6 / 4

Attached appendix figures / tables

9 / 4

Attached appendix span

≈ 5.84 page-equivalents

Before references / appendix: ≈ 53.68%References / intervening material: ≈ 15.58%Attached appendix: ≈ 30.74%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	≈ 1.11 pp.	5.84%
2 RELATED WORK	2	≈ 0.81 pp.	4.26%
3 AUTOEP	3	≈ 3.08 pp.	16.21%
4 EXPERIMENT	6	≈ 3.87 pp.	20.37%
5 DISCUSSION AND CONCLUSION	10	≈ 0.61 pp.	3.21%
ACKNOWLEDGMENTS	11	≈ 0.09 pp.	0.47%
REFERENCES	11	≈ 2.96 pp.	15.58%
A LLM USAGE	14	≈ 0.07 pp.	0.37%
B DETAILED EXPERIMENTAL RESULTS	14	≈ 1.3 pp.	6.84%
C PROMPT	15	≈ 3.04 pp.	16%
D DETAILS OF THE EXPERIMENTAL SETUP	18	≈ 1.43 pp.	7.53%

All detected figure and table locations

Main text

• Figure 1 · p. 2
• Figure 2 · p. 5
• Figure 3 · p. 6
• Table 1 · p. 7
• Table 2 · p. 8
• Table 3 · p. 8
• Figure 4 · p. 9
• Figure 5 · p. 9
• Table 4 · p. 9
• Figure 6 · p. 10

Attached appendix

• Table 5 · p. 14
• Table 6 · p. 15
• Table 7 · p. 15
• Figure 7 · p. 15
• Figure 10 · p. 15
• Figure 8 · p. 16
• Figure 9 · p. 16
• Figure 11 · p. 17
• Figure 12 · p. 17
• Figure 13 · p. 17
• Figure 14 · p. 18
• Figure 15 · p. 18
• Table 8 · p. 19

Argument and organization

1. Training-cost motivation → measured search state

   Connects a practical automation gap to five quantitative input features.

2. Sense → reason → act

   Decomposes control into one-time strategy, state diagnosis and numerical adjustment.

3. Solver benchmarks → component and frequency ablations

   Tests the controller alongside its overhead and context-history choices.

Problem, claim and contribution

The method deliberately leaves the numerical optimizer in place and uses language models as hyperparameter supervisors; this distinguishes the work from generating candidate solutions or new operators.

Section 2, p. 3 · Main paper

Argument and section order

Feature definitions precede the three controller roles, and a concrete GA prompt-response example explains how qualitative exploration decisions become parameter values.

Sections 3.1–3.3 and Figure 3, pp. 3–6 · Main paper

Figures, captions and visual explanation

The overview links the experience pool to a central solver–controller loop, with a magnified reasoning panel. Its caption is much less informative than the accompanying prose, so the diagram is not fully self-contained.

Figure 2, p. 5 · Main paper

Tables, comparisons and evidence

Result tables pair optimality gaps with elapsed time, exposing quality–cost trade-offs. The component ablation removes ELA, reasoning decomposition and both, while the history-window table tests context length rather than only model size.

Tables 1–4, pp. 7–9 · Main paper

Appendix and reproducibility

Complete GA, PSO and ACO prompts reveal manually supplied feature-to-strategy rules and allowed parameter ranges. This is crucial context for understanding what the training-free controller is given.

Appendix C, pp. 15–18 · Main paper

Clarity and writing conventions

Several interpretive sentences call visual correlation proof of understanding or historical-context use; these would be stronger as hypotheses supported by controlled intervention, not certainty.

Section 4.6, pp. 9–10 · Main paper

What to reuse—and what not to infer

Transferable lesson: For an agentic controller, show the measurable state, role-specific prompts and feedback loop, then report task quality beside end-to-end overhead.

Boundary: Do not copy the blanket efficiency claims: 30 ms in Section 4.2 conflicts with 0.3 seconds in Appendix D, and local-deployment rhetoric differs from the statement that all LLMs used APIs. The same dsj1000 configuration also reports 3.58% in Tables 1–2 but 1.58% in Table 3.

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, Sections 1–5, complete features, controller roles, benchmarks, ablations and discussion.

Appendix / supplement

Pages 14–19, Appendices A–D, all additional experiments, actual prompts and experimental settings.

Visually inspected PDF pages

5, 8, 9

• ICLR 2026 official proceedings PDF, retrieved 2026-09-19 · 19 pages