Writing study: RealPDEBench: A Benchmark for Complex Physical Systems with Real-World Data

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Peiyan Hu, Haodong Feng, Hongyuan Liu, Tongtong Yan, Wenhao Deng, Tianrun Gao, Rong Zheng, Haoren Zheng, Chenglei Yu, Chuanrui Wang, Kaiwen Li, Zhi-Ming Ma, Dezhi Zhou, Xingcai Lu, Dixia Fan, Tailin Wu · 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: Datasets, benchmarks & measurement resources (primary) · Empirical analysis, diagnosis & auditing.

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

46 physical pages

Before references / appendix

≈ 10.46 page-equivalents

Main figures / tables

6 / 1

Attached appendix figures / tables

15 / 8

Attached appendix span

≈ 28.89 page-equivalents

Before references / appendix: ≈ 22.74%References / intervening material: ≈ 14.46%Attached appendix: ≈ 62.8%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	2	≈ 1 pp.	2.17%
2 RELATED WORK	3	≈ 0.5 pp.	1.09%
3 REALPDEBENCH: BRIDGING REAL-WORLD AND SIMULATED DATA	3	≈ 3.78 pp.	8.22%
4 EXPERIMENTS	7	≈ 3.41 pp.	7.41%
5 CONCLUSION	10	≈ 0.43 pp.	0.93%
ETHICS STATEMENT	11	≈ 0.11 pp.	0.24%
REPRODUCIBILITY STATEMENT	11	≈ 0.12 pp.	0.26%
REFERENCES	11	≈ 6.65 pp.	14.46%
A OTHER EXPERIMENT RESULTS	18	≈ 8 pp.	17.39%
B DATASET DETAILS	26	≈ 5.75 pp.	12.5%
C CODE FRAMEWORK	31	≈ 2.5 pp.	5.43%
D EXPERIMENT DETAILS	34	≈ 0.3 pp.	0.65%
E BASELINE MODELS	34	≈ 2.45 pp.	5.33%
F LIMITATIONS AND FUTURE WORKS	37	≈ 0.19 pp.	0.41%
G VISUALIZATION	37	≈ 0.08 pp.	0.17%
H THE USE OF LARGE LANGUAGE MODELS (LLMS)	37	≈ 9.62 pp.	20.91%

All detected figure and table locations

Main text

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

Attached appendix

• Table 2 · p. 18
• Table 3 · p. 20
• Table 4 · p. 21
• Table 5 · p. 22
• Figure 7 · p. 23
• Figure 8 · p. 24
• Table 6 · p. 25
• Table 7 · p. 26
• Figure 9 · p. 28
• Figure 10 · p. 28
• Table 8 · p. 30
• Table 9 · p. 30
• Figure 11 · p. 31
• Figure 12 · p. 32
• Figure 13 · p. 38
• Figure 14 · p. 39
• Figure 15 · p. 40
• Figure 16 · p. 41
• Figure 17 · p. 42
• Figure 18 · p. 43
• Figure 19 · p. 44
• Figure 20 · p. 45
• Figure 21 · p. 46

Argument and organization

1. Data gap and related benchmarks

   Explain why simulated-only validation misses the intended deployment setting.

2. Tasks, datasets, metrics, baselines

   Present the four benchmark contributions as separate reproducible contracts.

3. Experimental questions

   Organize findings around domain gap, simulation pretraining, model trade-offs, and long-horizon behavior.

4. Acquisition and implementation appendix

   Connect physical measurement, preprocessing, modular code, and extended diagnostics.

Problem, claim and contribution

The contribution list is divided into data, tasks, metrics, and baselines, and the rest of the paper follows that map. This suits a benchmark contribution better than imitating a new-model narrative.

Section 1, page 2; Section 3, pages 3–7. · Main paper

Argument and section order

All three training regimes share real-world validation/test data; the text explains their different data access explicitly. Subsequent sections ask whether simulation transfers, whether pretraining helps, and how long-horizon errors grow.

Section 3.1, pages 3–4; Sections 4.2–4.5, pages 7–10. · Main paper

Figures, captions and visual explanation

The first-page montage aligns measured and simulated fields across five scenarios. Stacked channel images make differing observability visible before the reader encounters equations or model scores.

Figure 1, page 1. · Main paper

The quadrant trade-off plot packs four datasets into one visual but gives each quadrant different axis scales and omits distant points. Small multiples with explicit scales would be easier to compare accurately.

Figure 4 and its caption, page 9. · Main paper

Tables, comparisons and evidence

Pixel-level errors are supplemented by spectral, periodicity, energy, and velocity-profile metrics. The autoregressive curves expose degradation that a single-step table would miss.

Section 3.3, pages 5–6; Figure 3, page 9; Figure 5, page 10. · Main paper

Appendix and reproducibility

Acquisition hardware, boundary conditions, calibration checks, raw-data availability, dataset class arguments, and baseline adaptations receive dedicated subsections. This is a useful model for reporting provenance beyond a download link.

Appendices B–E, pages 26–36. · Main paper

What to reuse—and what not to infer

Transferable lesson: For a real-data benchmark, make data provenance and shared evaluation splits first-class contributions, and pair ordinary prediction errors with domain-specific diagnostics and long-horizon tests.

Boundary: The claim of consistent fine-tuning improvement has row-level exceptions in Table 1. Update Ratio counts fine-tuning updates, not total pretraining cost. Paired trajectories share physical parameters, not exact framewise alignment; combustion uses an extra real-data-trained modality surrogate. Appendix B reports inconsistent acquisition rates in places, so reproduce from released configurations rather than assuming every prose number agrees.

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 1–11, Sections 1–5 and ethics/reproducibility statements; references excluded from analytical reading.

Appendix / supplement

Pages 18–46, Appendices A–H in full, including every extended table, acquisition/calibration description, framework listing, and visualization caption; not all gallery images visually audited.

Visually inspected PDF pages

1, 9

• Official ICLR 2026 proceedings conference PDF; PDF creation metadata: Sat Feb 7 19:57:56 2026 EST · 46 pages