Writing study: PAI-Bench: A Comprehensive Benchmark For Physical AI

CVPR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← CVPR 2026 oral writing corpus

Fengzhe Zhou, Jiannan Huang, Jialuo Li, Deva Ramanan, Humphrey Shi · Official CVPR 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 (primary).

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

15 physical pages

Before references / appendix

≈ 8.17 page-equivalents

Main figures / tables

7 / 6

Attached appendix figures / tables

0 / 0

Attached appendix span

Separate file recorded below; not in this PDF total

Before references / appendix: ≈ 54.47%References / intervening material: ≈ 45.53%Estimated layout envelopes, not text/artwork area. Unresolved appendix boundaries remain within the undivided post-body material.

Extraction notes: No attached appendix boundary detected. Zero means no recognized caption, not independently verified absence.

Source-bound correction: Physical page 5 places Figure 4 and Figure 5 side by side. Their captions were joined into one extracted line; raster inspection confirms the second label. Figure 5 is a word cloud, not another panel of Figure 4.

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	≈ 0.7 pp.	4.67%
2. Related Work	2	≈ 0.5 pp.	3.33%
3. PAI-Bench	3	≈ 2.15 pp.	14.33%
4. Experiments	5	≈ 3.22 pp.	21.47%
5. Further Analysis and Discuss	8	≈ 0.2 pp.	1.33%
6. Conclusion	8	≈ 0.22 pp.	1.47%
Acknowledgements	9	≈ 0.13 pp.	0.87%
References	9	≈ 6.83 pp.	45.53%

All detected figure and table locations

Main text

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

Attached appendix

No recognized caption labels.

Separate source files—excluded from canonical-PDF totals

A separate author version may repeat main text and references. Whole-file lengths and counts are not extra appendix material and are never added to the canonical PDF totals. Read the study's source scope for which sections received substantive reading versus layout inspection.

Additional source 1

Separate PDF (or ZIP containing it): whole-file layout of 17 physical pages; 15 detected figures and 3 tables. These may include overlapping main-text material.

Located appendix only: starts on physical p.1; ≈ 16.95 page-equivalents, 15 figure labels and 3 table labels. This interval is distinct from the whole file.

Detected heading	PDF page	Estimated span
Appendix	1	≈ 0.03 pp.
A. More Details about PAI-Bench-G	1	≈ 2.17 pp.
B. PAI-Bench-C Score Calculation Details	3	≈ 0.67 pp.
C. More Details about PAI-Bench-U	3	≈ 1.63 pp.
D. PAI-Bench Leaderboard Visualization	5	≈ 0.07 pp.
E. Discussion and Future Work	5	≈ 12.38 pp.

• Table 7 · p. 2
• Figure 8 · p. 2
• Table 8 · p. 3
• Figure 9 · p. 4
• Figure 10 · p. 4
• Table 9 · p. 5
• Figure 11 · p. 6
• Figure 12 · p. 7
• Figure 13 · p. 8
• Figure 14 · p. 9
• Figure 15 · p. 10
• Figure 16 · p. 11
• Figure 17 · p. 12
• Figure 18 · p. 13
• Figure 19 · p. 14
• Figure 20 · p. 15
• Figure 21 · p. 16
• Figure 22 · p. 17

Argument and organization

1. Benchmark §§3.1–3.3

   Define each generation, conditional-generation and understanding track with its data and scoring.

2. Experiments §§4.1–4.3 and analysis §5

   Mirror the same three tracks, then synthesize common weaknesses and alternative explanations.

Problem, claim and contribution

The benchmark distinguishes visual fidelity from physical plausibility and perception from prediction; these become separate suites and metrics rather than an undefined overall notion of intelligence.

Abstract, PDF p. 1; §3, PDF pp. 3–5 · Main paper · Additional source 2

Argument and section order

The experimental subsections mirror the benchmark-definition subsections, giving readers a stable G/C/U naming scheme from data curation to results.

§§3–4, PDF pp. 3–7 · Main paper · Additional source 2

Figures, captions and visual explanation

Page 3 pairs a benchmark-coverage matrix with domain sample-count bars; page 6 pairs detailed score tables with human–automated correlation plots and visible confidence intervals.

Table 1 and Figure 2, PDF p. 3; Tables 3–4 and Figure 6, PDF p. 6 · Main paper · Additional source 2

Tables, comparisons and evidence

The understanding table includes human and random-guess baselines. Frame-count and thinking-mode tests probe what drives scores rather than merely ranking model families.

Table 5, PDF p. 7; Figure 7 and Table 6, PDF p. 8 · Main paper · Additional source 2

Appendix and reproducibility

The supplement expands generation settings, scoring, human-study interfaces, model versions and domain examples; radar charts are explicitly normalized rather than presented as raw comparable units.

Supplement §§A–E, PDF pp. 1–17, Tables 7–9 and Figures 20–22 · Main paper · Additional source 2

Clarity and writing conventions

The paper reports exceptions to uniform evaluation: InternVL receives fewer frames, and unsupported conditional-generation settings are omitted with explanation. These details must accompany any leaderboard inference.

§§4.2–4.3, PDF p. 7 · Main paper · Additional source 2

What to reuse—and what not to infer

Transferable lesson: A multi-track benchmark should mirror its taxonomy in the results and include sample distributions, chance/human anchors and evidence that the scoring protocol measures the intended capability.

Boundary: Frame-count controls reduce some shortcuts but do not prove the benchmark is unbiased; model-judge scores and correlations are measurements under a protocol, not exhaustive physical understanding.

Apply this lesson using the argument and evidence worksheet.

Inspection record and sources

Checked

2026-09-26

Reading status

Writing-focused manuscript scan + selected visuals inspected

Main text

Manuscript-wide writing scan of PDF pp. 1–15: framing, argument sequence, section organization, figure/table captions, evaluation interpretation and limitations; reference sections identified. This is a writing-focused scan, not an exhaustive technical proof or reproduction audit. Main PDF pages 3 and 6 were rendered and visually inspected.

Appendix / supplement

Writing-focused scan of the separate supplementary PDF, pp. 1–17, including its organization, captions, protocol/evidence detail and limitation material where present; proofs and code listings were inspected for exposition, not independently verified or executed. Supplement §§A–E, PDF pp. 1–17, Tables 7–9 and Figures 20–22. Linked video/audio demonstrations and non-PDF archive contents were not assessed.

Visually inspected PDF pages

3, 6

This is a writing-focused scan of the manuscript and the stated supplementary scope: framing, organization, visual explanation, evidence presentation and wording. It is not a line-by-line proof audit, reproduction of results or exhaustive scientific review.

• Official CVPR 2026 CVF open-access accepted manuscript; retrieved 2026-09-26 · 15 pages
• Official CVPR 2026 CVF supplementary PDF; retrieved 2026-09-26 · 17 pages