Writing study: 3ViewSense: Spatial and Mental Perspective Reasoning from Orthographic Views in Vision-Language Models

ICML overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICML 2026 oral writing corpus

Shaoxiong Zhan, Yanlin Lai, Zheng Liu, Lin Hai, Shen Li, Xiaodong Cai, Zijian Lin, Wen Huang, Hai-Tao Zheng · Official ICML 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: Training, objectives & optimization (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

24 physical pages

Before references / appendix

≈ 8.78 page-equivalents

Main figures / tables

5 / 5

Attached appendix figures / tables

9 / 6

Attached appendix span

≈ 12.91 page-equivalents

Before references / appendix: ≈ 36.58%References / intervening material: ≈ 9.63%Attached appendix: ≈ 53.79%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	≈ 1.37 pp.	5.71%
2. Related Work	2	≈ 0.61 pp.	2.54%
3. Methodology	3	≈ 3.38 pp.	14.08%
4. Experiments	6	≈ 0.55 pp.	2.29%
5. Results & Analysis	7	≈ 1.98 pp.	8.25%
6. Limitation and Future Work	9	≈ 0.17 pp.	0.71%
7. Conclusion	9	≈ 0.24 pp.	1%
Impact Statement	9	≈ 0.12 pp.	0.5%
References	9	≈ 2.31 pp.	9.63%
A. Preliminary Analysis	12	≈ 0.73 pp.	3.04%
B. Additional Method Details	12	≈ 0.38 pp.	1.58%
C. Additional Experiments Details	13	≈ 11.81 pp.	49.21%

All detected figure and table locations

Main text

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

Attached appendix

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

Argument and organization

1. Problem and contribution

   A three-view hint diagnostic motivates separating view induction from answer reasoning rather than treating spatial failure as one undifferentiated capacity deficit.

2. Argument sequence

   Dataset and task definitions precede two supervised stages and RL refinement; evaluation then separates in-domain fitting, out-of-domain transfer and external spatial benchmarks.

3. Evidence and verification

   Matched-input/instance/hyperparameter Direct QA controls distinguish answer supervision from reasoning targets, while stage ablations and initialization curves diagnose different parts of the training sequence.

Problem, claim and contribution

A three-view hint diagnostic motivates separating view induction from answer reasoning rather than treating spatial failure as one undifferentiated capacity deficit.

Figure 1 and contribution list, p. 2 · Main paper

Argument and section order

Dataset and task definitions precede two supervised stages and RL refinement; evaluation then separates in-domain fitting, out-of-domain transfer and external spatial benchmarks.

Sections 3-5, pp. 3-8 · Main paper

Figures, captions and visual explanation

The two-color training diagram assigns different intermediate outputs to the two stages; the result page explicitly retains red relative drops and a direct-answer control that wins in-domain but loses transfer.

Figure 3, p. 5; Tables 2, 4 and 5, p. 8; both PNGs inspected · Main paper

Tables, comparisons and evidence

Matched-input/instance/hyperparameter Direct QA controls distinguish answer supervision from reasoning targets, while stage ablations and initialization curves diagnose different parts of the training sequence.

Section 5.3, pp. 8-9 · Main paper

Appendix and reproducibility

The appendix exposes both geometric uniqueness conditions and teacher prompts that receive the answer while generating first-person reasoning, making trace provenance independently inspectable.

Appendix A.1, p. 12; Figure 9, p. 20 · Main paper

Clarity and writing conventions

Limitations distinguish geometric projections from support, affordances and dynamics; however a probe's decodable information should not be equated with a complete causal localization of failure.

Section 6, p. 9; Appendix C.3-C.4, pp. 14-15 · Main paper

What to reuse—and what not to infer

Transferable lesson: Use a diagnostic representation intervention to motivate training, then separate target-task fitting from transfer and expose teacher-trace construction.

Boundary: Three views are not universally sufficient, teacher traces are answer-conditioned, and probe accuracy or readable reasoning does not prove the deployed model uses the claimed internal mechanism.

Apply this lesson using the argument and evidence worksheet.

Inspection record and sources

Checked

2026-10-07

Reading status

Writing-focused manuscript scan + selected visuals inspected

Main text

Writing-focused distributed manuscript scan, not exhaustive technical/proof validation. Sections 1-7, pp. 1-9: diagnostic motivation, task construction, staged representation/reasoning training, ID/OOD transfer, ablations and limitations.

Appendix / supplement

Writing-focused appendix scan; code, data and external assets not executed. Appendices A-C and prompt/example plates, pp. 12-24: uniqueness conditions, generation and training data, sampling, visual probes, qualitative responses and full prompting templates.

Visually inspected PDF pages

5, 8

This is a writing-focused scan of the manuscript and the stated supplementary scope: framing, organization, visual explanation, evidence presentation and wording. It is not a line-by-line proof audit, reproduction of results or exhaustive scientific review.

• Published ICML 2026 proceedings PDF, PMLR volume 306; retrieved 2026-10-07. Public OpenReview decision notes were not independently reverified. · 24 pages