Writing study: World-In-World: World Models in a Closed-Loop World

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Jiahan Zhang, Muqing Jiang, Nanru Dai, Taiming Lu, Arda Uzunoglu, Shunchi Zhang, Yana Wei, Jiahao Wang, Vishal Patel, Paul Liang, Daniel Khashabi, Cheng Peng, Rama Chellappa, Tianmin Shu, Alan Yuille, Yilun Du, Jieneng Chen · 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

40 physical pages

Before references / appendix

≈ 10.14 page-equivalents

Main figures / tables

7 / 6

Attached appendix figures / tables

11 / 5

Attached appendix span

≈ 22.83 page-equivalents

Before references / appendix: ≈ 25.35%References / intervening material: ≈ 17.58%Attached appendix: ≈ 57.07%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	≈ 1.14 pp.	2.85%
2 WORLD-IN-WORLD: A CLOSED-LOOP INTERFACE FOR VISUAL WORLD	3	≈ 3.38 pp.	8.45%
3 EVALUATION RESULTS AND ANALYSIS	6	≈ 3.47 pp.	8.68%
4 DISCUSSION AND FUTURE DIRECTIONS	10	≈ 0.65 pp.	1.63%
5 CONCLUSION	10	≈ 0.39 pp.	0.97%
REFERENCES	11	≈ 7.03 pp.	17.58%
Appendix	18	≈ 1.15 pp.	2.88%
A RELATED WORK	19	≈ 0.5 pp.	1.25%
B EMBODIED TASK DETAILS	19	≈ 6.41 pp.	16.02%
C POST-TRAINING RECIPE FOR EMBODIED WORLD MODELS	26	≈ 1.48 pp.	3.7%
D POST-TRAINING DATASET CONSTRUCTION	27	≈ 2.04 pp.	5.1%
E VISUALIZING WORLD MODEL PREDICTIONS	29	≈ 4.36 pp.	10.9%
F PROMPT TEMPLATES USED IN WORLD-IN-WORLD	34	≈ 6 pp.	15%
G USE OF LANGUAGE MODELS	40	≈ 0.89 pp.	2.23%

All detected figure and table locations

Main text

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

Attached appendix

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

Argument and organization

1. Introduction

   Identifies a mismatch between visual-quality metrics and embodied utility.

2. Closed-loop interface

   Defines shared planning, action translation, tasks and adaptation before comparisons.

3. Evaluation / Discussion

   Connects benchmark outcomes to controllability, data, compute and unresolved failures.

4. Appendices A–G

   Documents policy-specific evaluation details, data construction, examples and exact prompts.

Problem, claim and contribution

The introduction asks whether beautiful predictions help an agent succeed, shifting the evaluation target from image quality to downstream utility. A scatterplot makes that mismatch concrete before the interface is introduced.

§1 and Figure 2, p.2 · Main paper

Argument and section order

The method separates proposal, simulation and revision, then distinguishes common interface from task-specific decisions. This lets one mathematical framework cover both action selection and joint action–answer refinement without pretending their outputs are identical.

§§2.1–2.3, pp.3–6; Appendix B.5, PDF pp.24–25 · Main paper

Figures, captions and visual explanation

Figure 3 uses a visual boundary between imagined and actual interactions, with four numbered components and a return arrow. It communicates the closed loop more effectively than a one-way model pipeline would.

Figure 3, p.3 · Main paper

Tables, comparisons and evidence

The main tables retain control modality, input type, policy baseline and post-training markers. These columns reveal meaningful confounders and make it possible to compare a model against its own unadapted version instead of ranking all systems as interchangeable.

Tables 1–3, p.7 · Main paper

Appendix and reproducibility

Full prompts disclose that simulated observations are approximate and require real-observation verification before some answers. Good/bad rollout grids show the action-following failures motivating that instruction, although the grids do not name every generating model.

Appendix E, Figures 13–16, PDF pp.30–31; Appendix F.3, PDF p.38 · Main paper

Clarity and writing conventions

The discussion organizes limitations by practical failure mode: generalization, long horizons, contact dynamics, policy strength and cost. Each is connected to a measured weakness rather than a generic statement that more research is needed.

§4, p.10 · Main paper

What to reuse—and what not to infer

Transferable lesson: Define the downstream decision a benchmark should improve, visualize the evaluation loop, and expose control/input differences beside scores rather than hiding them in implementation details.

Boundary: The framework evaluates simulated environments, not deployed real robots. ‘Consistently improves’ overstates the manipulation table, where several unadapted models tie or worsen success. Data-scaling plots are finite trends, not a fitted universal scaling law. Input panoramas, privileged depth and policy differences qualify fairness claims; statistical significance is not established by the displayed tables.

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.1–10, §§1–5; references pp.11–17

Appendix / supplement

PDF pp.18–40 (appendix printed pages 1–23), contents and Appendices A–G, including complete task protocols, adaptation data, rollout examples and prompts

Visually inspected PDF pages

3, 7, 30, 31

• official ICLR 2026 proceedings PDF · 40 pages