Writing study: Latent Particle World Models: Self-supervised Object-centric Stochastic Dynamics Modeling

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Tal Daniel, Carl Qi, Dan Haramati, Amir Zadeh, Chuan Li, Aviv Tamar, Deepak Pathak, David Held · 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: Architecture & representation design (primary) · Learning & optimization algorithms.

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

49 physical pages

Before references / appendix

≈ 10.46 page-equivalents

Main figures / tables

4 / 3

Attached appendix figures / tables

13 / 10

Attached appendix span

≈ 28.89 page-equivalents

Before references / appendix: ≈ 21.35%References / intervening material: ≈ 19.69%Attached appendix: ≈ 58.96%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	≈ 1.27 pp.	2.59%
2 RELATED WORK	3	≈ 1.21 pp.	2.47%
3 BACKGROUND	4	≈ 0.72 pp.	1.47%
4 LATENT PARTICLE WORLD MODELS (LPWM)	5	≈ 2.8 pp.	5.71%
5 EXPERIMENTS	8	≈ 2.32 pp.	4.73%
6 CONCLUSION	10	≈ 0.68 pp.	1.39%
7 ETHICS STATEMENT	11	≈ 0.14 pp.	0.29%
8 REPRODUCIBILITY STATEMENT	11	≈ 0.12 pp.	0.24%
ACKNOWLEDGMENTS	11	≈ 0.1 pp.	0.2%
REFERENCES	11	≈ 9.65 pp.	19.69%
A APPENDIX	21	≈ 28.89 pp.	58.96%

All detected figure and table locations

Main text

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

Attached appendix

• Figure 5 · p. 22
• Figure 6 · p. 22
• Figure 7 · p. 23
• Figure 8 · p. 24
• Figure 9 · p. 25
• Figure 10 · p. 28
• Figure 11 · p. 30
• Figure 12 · p. 30
• Figure 13 · p. 33
• Figure 14 · p. 34
• Figure 15 · p. 37
• Table 4 · p. 41
• Table 5 · p. 44
• Table 6 · p. 45
• Table 7 · p. 45
• Table 8 · p. 46
• Table 9 · p. 46
• Figure 16 · p. 47
• Table 10 · p. 47
• Table 11 · p. 47
• Table 12 · p. 48
• Figure 17 · p. 49
• Table 13 · p. 49

Argument and organization

1. Motivation, related work and background

   Motivates object-centric tokens, distinguishes neighboring methods in a capability table and defines the inherited particle representation.

2. Four-module world model

   Introduces encoder, decoder, context and dynamics with special emphasis on the new per-particle latent-action mechanism.

3. Prediction, ablation and policy experiments

   Separates video modeling from post-hoc imitation learning, followed by explicit applicability limitations.

Problem, claim and contribution

The introduction contrasts semantic language units with fixed image patches through a small visual example, then positions the contribution as an extension of DLP/DDLP rather than hiding its lineage.

Figure 2 and Sections 1–2, pp. 2–4 · Main paper

Argument and section order

Background notation comes before the four named modules. Main-text summaries point to matching appendix subsections, where modifications to inherited components are called out explicitly.

Sections 3–4, pp. 4–7; Appendix A.4, pp. 24–36 · Main paper

Figures, captions and visual explanation

The architecture diagram uses consistent particle colors and separate prior/posterior paths, then zooms into context conditioning. Its caption explicitly distinguishes training-time inverse dynamics from inference-time policy sampling.

Figure 3, p. 7 · Main paper

Tables, comparisons and evidence

The results table places training, conditioning and prediction horizons beside each dataset and explains which metrics use posterior latent actions versus prior sampling, preventing unlike protocols from being silently conflated.

Table 2, p. 9; Table 10, p. 47 · Main paper

Appendix and reproducibility

Small code listings bridge mathematical definitions to implementable spatial transforms, compositing and policy mapping; full policy results expose weak tasks that the main table omits.

Figures 6–7, 12, 15; Table 13, pp. 22–49 · Main paper

What to reuse—and what not to infer

Transferable lesson: For a modular method, align notation, architecture colors, subsection names and implementation snippets, and state what information each evaluation mode may access.

Boundary: Do not present posterior-latent-action reconstruction as unconstrained future prediction. The prose overstates PandaPush dominance: LPWM is below EC Diffusion Policy on three cubes, and its OGBench overall score is below HIQL.

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: abstract and Sections 1–8 through reproducibility and acknowledgments; references excluded from analytical reading.

Appendix / supplement

Pages 21–49: Appendix A.1–A.10.4, including mathematical definitions, implementation pseudocode, all dataset/architecture tables and full policy results.

Visually inspected PDF pages

7, 9

• Official ICLR 2026 proceedings conference PDF; PDF creation metadata: Wed Feb 25 10:31:27 2026 EST · 49 pages