Writing study: PhyWorldBench: A Comprehensive Evaluation of Physical Realism in Text-to-Video Models

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Jing Gu, Xian Liu, Yu Zeng, Ashwin Nagarajan, Fangrui Zhu, Daniel Hong, Yue Fan, Qianqi Yan, Kaiwen Zhou, Ming-Yu Liu, Xin 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: Datasets, benchmarks & measurement resources (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

35 physical pages

Before references / appendix

≈ 9.21 page-equivalents

Main figures / tables

6 / 4

Attached appendix figures / tables

15 / 13

Attached appendix span

≈ 23.28 page-equivalents

Before references / appendix: ≈ 26.31%References / intervening material: ≈ 7.17%Attached appendix: ≈ 66.51%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	≈ 0.9 pp.	2.57%
2 RELATED WORK	3	≈ 1.25 pp.	3.57%
3 P H Y W O R L D B E N C H	4	≈ 3.04 pp.	8.69%
4 VIDEO GENERATION RESULTS AND ANALYSIS	7	≈ 2.1 pp.	6%
5 CONTRIBUTIONS AND FUTURE WORK	9	≈ 0.41 pp.	1.17%
REFERENCES	10	≈ 2.5 pp.	7.14%
A LLM USAGE STATEMENT	12	≈ 0.09 pp.	0.26%
B BENCHMARK DETAILS	12	≈ 0.71 pp.	2.03%
C ANNOTATION DETAILS	13	≈ 0.32 pp.	0.91%
D DETAILED ANALYSIS ON EVALUATOR	13	≈ 1.99 pp.	5.69%
E MODEL APPLICATION DETAILS	15	≈ 0.87 pp.	2.49%
F CAP RESULTS	16	≈ 0.08 pp.	0.23%
G INSIGHTS INTO PERFORMANCE BETWEEN MODELS	16	≈ 0.7 pp.	2%
H DETAILED PERFORMANCE OVER ALL PHYSICS CATEGORIES	17	≈ 0.08 pp.	0.23%
I LIMITATIONS	17	≈ 0.15 pp.	0.43%
J CODE AND DATA AVAILABILITY	17	≈ 0.37 pp.	1.06%
K QUERY USED	18	≈ 0.56 pp.	1.6%
L QUALITATIVE ANALYSIS ON INDIVIDUAL MODEL	18	≈ 0.94 pp.	2.69%
M STATIC-SCENE RATIONALIZATION AS AN EMERGING FAILURE MODE	19	≈ 12.59 pp.	35.97%
N VIDEO LENGTH FROM MODELS	32	≈ 0.04 pp.	0.11%
O FURTHER ANALYSIS ON GENERATIVE MODELS STRUGGLING WITH	32	≈ 0.23 pp.	0.66%
P LEADERBOARD	32	≈ 0.42 pp.	1.2%
Q VIDEO QUANTITATIVE ANALYSIS	32	≈ 0.55 pp.	1.57%
R INFLUENCE OF MULTIPLE OBJECTS	33	≈ 0.09 pp.	0.26%
S DATASET STRUCTURE	33	≈ 1.59 pp.	4.54%
T DATA RELEASE	35	≈ 0.89 pp.	2.54%

All detected figure and table locations

Main text

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

Attached appendix

• Figure 7 · p. 14
• Table 5 · p. 15
• Table 6 · p. 15
• Table 7 · p. 16
• Table 8 · p. 17
• Table 9 · p. 17
• Table 10 · p. 18
• Table 11 · p. 18
• Table 12 · p. 18
• Table 13 · p. 19
• Figure 8 · p. 20
• Figure 9 · p. 21
• Figure 10 · p. 22
• Figure 11 · p. 23
• Figure 12 · p. 24
• Figure 13 · p. 25
• Figure 14 · p. 26
• Figure 15 · p. 27
• Figure 16 · p. 28
• Figure 17 · p. 29
• Figure 18 · p. 30
• Figure 19 · p. 31
• Table 14 · p. 32
• Table 15 · p. 32
• Table 16 · p. 33
• Table 17 · p. 33
• Figure 20 · p. 33
• Figure 21 · p. 34

Argument and organization

1. Benchmark construction and evaluation

   Define a hierarchical prompt inventory and separate semantic adherence from physical criteria before reporting scores.

2. Video generation results and analysis

   Separate generator rankings, prompt interventions and impossible-scene behavior.

3. Appendices A–T

   Expose annotation, model versions, prompts, failure examples and sensitivity analyses.

Problem, claim and contribution

The opening taxonomy turns a vague physical-realism question into 10 categories, 50 subcategories and three prompt styles. The arithmetic behind 1,050 prompts makes coverage inspectable, although category breadth is not proof of physical completeness.

§3.1, Figs.1–3, pp.1–4 · Main paper

Argument and section order

Human curation, LLM-assisted scenario proposals and evaluator prompting occupy separate steps. This distinguishes constructing the test from training a generator; CAP is an evaluation prompt, not a new video-generation architecture.

§§3.1–3.3, pp.3–7 · Main paper

Figures, captions and visual explanation

The worked video strips annotate objects, events and physical criteria separately, with marks directly on frames. The reader can see why semantic success and physical failure can coexist instead of inferring the distinction from a formula alone.

Fig.5, p.6 · Main paper

Tables, comparisons and evidence

Table 2 validates the automatic judge against human labels; Table 3 evaluates generators using human labels. Maintaining this separation avoids substituting judge agreement for generator quality, or treating their differently scaled scores as interchangeable.

Tables 2–3, pp.7–8; Tables 5 and 7, pp.15–16 · Main paper

Appendix and reproducibility

The appendices disclose three-rater majority annotation, compensation, exact two-stage prompts and model-specific video durations. The expanded examples support auditing operational criteria that headline averages cannot reveal.

Appendices B, K and N, pp.13, 18–20, 32 · Main paper

Clarity and writing conventions

The strongest interpretations exceed the operational measurements: impossible prompts test requested counterfactual adherence, and small false-positive changes do not by themselves establish judge neutrality. Descriptive benchmark claims are safer than claims about genuine physical understanding.

§§3.3 and 4.4, pp.7–8; Appendices I and M, pp.17, 27–31 · Main paper

What to reuse—and what not to infer

Transferable lesson: Define the operational rubric, illustrate a scored example, and validate the evaluator separately from the system being evaluated.

Boundary: The actual table covers 12 models despite a stale 10-model caption and abstract count. Anti-physics criteria are prompt-specific, not literal compliance with ordinary physics. Some illustrative criteria are overly restrictive without controlled conditions. Eight sampled frames can miss short events, durations differ across generators, and these 2024–2025 model versions are not a current frontier ranking.

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

PDF pp.1–9, §§1–5; statements and references pp.10–12

Appendix / supplement

Appendices A–T, PDF pp.12–35, including all prompt templates, category examples and additional analyses

Visually inspected PDF pages

6, 7, 8

• official ICLR 2026 proceedings PDF · 35 pages