Writing study: Visual Planning: Let's Think Only with Images

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Yi Xu, Chengzu Li, Han Zhou, Xingchen Wan, Caiqi Zhang, Anna Korhonen, Ivan Vulić · 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: Learning & optimization algorithms (primary) · Architecture & representation design.

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

34 physical pages

Before references / appendix

≈ 9.44 page-equivalents

Main figures / tables

6 / 2

Attached appendix figures / tables

10 / 11

Attached appendix span

≈ 18.89 page-equivalents

Before references / appendix: ≈ 27.76%References / intervening material: ≈ 16.68%Attached appendix: ≈ 55.56%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	1	≈ 1.76 pp.	5.18%
2 VISUAL PLANNING VIA REINFORCEMENT LEARNING	3	≈ 2.18 pp.	6.41%
3 EXPERIMENTS AND RESULTS	5	≈ 2.85 pp.	8.38%
4 DISCUSSIONS AND ANALYSIS	8	≈ 1.02 pp.	3%
5 RELATED WORK	9	≈ 0.68 pp.	2%
6 CONCLUSION	10	≈ 0.21 pp.	0.62%
ACKNOWLEDGEMENTS	10	≈ 0.12 pp.	0.35%
REFERENCES	10	≈ 5.66 pp.	16.65%
A THE USE OF LARGE LANGUAGE MODELS	16	≈ 0.08 pp.	0.24%
B ETHICS STATEMENT	16	≈ 0.09 pp.	0.26%
C REPRODUCIBILITY STATEMENT	16	≈ 0.13 pp.	0.38%
D LIMITATIONS AND FUTURE WORK	16	≈ 0.77 pp.	2.26%
E IMPLEMENTATION DETAILS	17	≈ 3.64 pp.	10.71%
F RESULTS	20	≈ 12.28 pp.	36.12%
G PROMPTING TEMPLATES	33	≈ 1.89 pp.	5.56%

All detected figure and table locations

Main text

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

Attached appendix

• Table 3 · p. 18
• Table 4 · p. 18
• Table 5 · p. 19
• Table 6 · p. 20
• Figure 7 · p. 21
• Table 7 · p. 21
• Table 8 · p. 22
• Table 9 · p. 22
• Figure 8 · p. 23
• Figure 9 · p. 24
• Figure 10 · p. 25
• Figure 11 · p. 26
• Figure 12 · p. 27
• Table 10 · p. 27
• Figure 13 · p. 28
• Figure 14 · p. 29
• Figure 15 · p. 30
• Figure 16 · p. 31
• Table 11 · p. 31
• Table 12 · p. 32
• Table 13 · p. 32

Argument and organization

1. A different reasoning medium

   Contrasts textual rationales with generated visual-state trajectories.

2. Initialization then reinforcement learning

   Separates valid exploratory image generation from goal-directed planning.

3. Performance, failure analysis and generalization

   Explains improvements through grounding errors and exploration rather than a single score.

Problem, claim and contribution

The first comparison diagram makes the proposed change of representation explicit: actions emerge from image-state transitions instead of being verbalized. It frames visual planning as a supplement, not a universal replacement for language.

Figure 1 and introduction, pages 1–3 · Main paper

Argument and section order

The method separates a random-valid-transition warm-up from goal-directed GRPO. A later entropy-versus-invalidity analysis specifically tests why the first stage is needed.

Section 2.2 pages 3–5; Section 4 and Figure 6 pages 8–9 · Main paper

Figures, captions and visual explanation

Figure 2 follows the complete training loop from generated candidates through decoded image pairs, parsed actions and rewards back to a policy update. Colors distinguish optimal, nonoptimal and invalid outcomes.

Figure 2, page 4 · Main paper

Tables, comparisons and evidence

Table 1 exposes input/output modality and distinguishes post-trained baselines. The appendix strengthens interpretation with two-stage supervised controls, invalid-failure ratios and larger-grid tests.

Table 1 page 6; Tables 6, 8–10 pages 20–27 · Main paper

Appendix and reproducibility

The state parser and BFS-based progress signal are described concretely, including grayscale matching and object pick/drop rules. Prompt templates and sample counts make the controlled setting reproducible.

Appendix E pages 17–20; Appendix G pages 33–34 · Main paper

Clarity and writing conventions

Broad generalization language should be balanced by the larger-grid numbers: improvement is relative, while absolute exact-match success can remain below one percent for MiniBehavior.

Table 9, page 22; Appendix F.4 · Main paper

What to reuse—and what not to infer

Transferable lesson: When proposing a new reasoning representation, visualize the actual information flow and test the proposed explanation through an initialization ablation, not only a leaderboard comparison.

Boundary: The planner uses image-only learned outputs, but its training reward depends on a symbolic, task-specific parser and BFS oracle. Different pretraining backbones and image-token costs prevent treating the result as an isolated causal proof that images are intrinsically better than text.

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 1–10, Sections 1–6

Appendix / supplement

Pages 16–34, Appendices A–G including full prompts, reward parsing, dataset splits and additional results

Visually inspected PDF pages

4, 6

• ICLR 2026 final proceedings · 34 pages