Writing study: MomaGraph: State-Aware Unified Scene Graphs with Vision-Language Models for Embodied Task Planning

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Yuanchen Ju, Yongyuan Liang, Yen-Jen Wang, Gireesh Nandiraju, Yuanliang Ju, Seungjae (Jay) Lee, Qiao Gu, Elvis Hsieh, Furong Huang, Koushil Sreenath · 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) · Datasets, benchmarks & measurement resources.

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

26 physical pages

Before references / appendix

≈ 10.34 page-equivalents

Main figures / tables

5 / 3

Attached appendix figures / tables

14 / 3

Attached appendix span

≈ 10.89 page-equivalents

Before references / appendix: ≈ 39.77%References / intervening material: ≈ 18.35%Attached appendix: ≈ 41.88%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	≈ 1.18 pp.	4.54%
2 RELATED WORKS	3	≈ 0.5 pp.	1.92%
3 PRELIMINARY FINDINGS AND MOTIVATION EXPERIMENTS	3	≈ 1.42 pp.	5.46%
4 METHOD	5	≈ 1.55 pp.	5.96%
5 DATASET AND BENCHMARK	6	≈ 1.88 pp.	7.23%
6 EXPERIMENTS	8	≈ 2.07 pp.	7.96%
7 CONCLUSION	10	≈ 0.4 pp.	1.54%
8 ACKNOWLEDGEMENTS	11	≈ 0.24 pp.	0.92%
REFERENCES	11	≈ 4.76 pp.	18.31%
A APPENDIX	16	≈ 5.14 pp.	19.77%
B ADDITIONAL ABLATION STUDIES	21	≈ 5.76 pp.	22.15%

All detected figure and table locations

Main text

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

Attached appendix

• Figure 6 · p. 16
• Figure 7 · p. 17
• Figure 8 · p. 17
• Figure 9 · p. 18
• Figure 10 · p. 18
• Figure 11 · p. 19
• Figure 12 · p. 19
• Table 4 · p. 20
• Figure 13 · p. 20
• Table 5 · p. 21
• Table 6 · p. 21
• Figure 14 · p. 22
• Figure 15 · p. 23
• Figure 16 · p. 23
• Figure 17 · p. 24
• Figure 18 · p. 25
• Figure 19 · p. 26

Argument and organization

1. Motivating experiments

   Direct-plan and single-relation failures introduce the need for a task-specific unified scene representation before method details.

2. Representation, data, benchmark, deployment

   Separate named artifacts support graph-generation training, structured evaluation, and downstream robot demonstrations.

Problem, claim and contribution

Three limitations—single relation type, static state, and missing task relevance—map directly to the proposed representation's design choices.

Page 2, introduction. · Main paper

Argument and section order

Question-form headings and boxed preliminary findings make the motivating experiments serve as explicit design arguments.

Pages 3–5, section 3. · Main paper

Figures, captions and visual explanation

A central household scene is surrounded by task-coloured instruction panels, object crops, and small graphs, grounding abstract relations in recognizable actions.

Page 1, Figure 1. · Main paper

Tables, comparisons and evidence

The main table pairs without/with-graph columns within each difficulty tier and separates closed/open models, exposing both intervention effects and baseline scale.

Page 9, Table 2. · Main paper

Appendix and reproducibility

A concrete JSON annotation precedes dataset statistics; later sections report training configuration, alternative supervision, reward sensitivity, and stage-specific robot failures.

Pages 16–22, appendices A–B. · Main paper

Clarity and writing conventions

The deployment description explicitly separates the high-level planning contribution from teleoperation-derived task-specific primitive skills.

Page 10, section 6.3. · Main paper

What to reuse—and what not to infer

Transferable lesson: Use motivating controlled comparisons to justify each representation choice, then keep the representation, dataset, model, benchmark, and execution layer terminologically distinct.

Boundary: MCQ accuracy is not autonomous task-completion rate. The quantitative robot experiment has only 10 trials and task-specific primitive skills. The 60.2→71.6 gain is 11.4 percentage points, despite the source calling it a relative improvement.

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 1–8; references pages 11–15.

Appendix / supplement

Pages 16–26, appendices A–B: collection/annotation, training configuration, benchmark design, SFT/ICL and reward-weight ablations, robot evaluation and qualitative cases.

Visually inspected PDF pages

1, 9

• ICLR 2026 final proceedings · 26 pages