Writing study: SpatioLM: Towards General Physical Spatial Intelligence in Vision-Language Models

ICML overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICML 2026 oral writing corpus

Jing Wu, Jianhua Wu, Jiayi Guan, Jiahong Chen, Jinghui Lu, Hangjun Ye, Bingzhao Gao, Long Chen · 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: Architecture & representation (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

27 physical pages

Before references / appendix

≈ 9.16 page-equivalents

Main figures / tables

7 / 8

Attached appendix figures / tables

6 / 8

Attached appendix span

≈ 13.91 page-equivalents

Before references / appendix: ≈ 33.93%References / intervening material: ≈ 14.56%Attached appendix: ≈ 51.52%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	≈ 0.72 pp.	2.67%
2. Related Work	2	≈ 0.59 pp.	2.19%
3. Method	3	≈ 3.01 pp.	11.15%
4. Experiments	6	≈ 3.06 pp.	11.33%
5. Conclusion	9	≈ 0.41 pp.	1.52%
Impact Statement	9	≈ 0.21 pp.	0.78%
Acknowledgments	10	≈ 0.11 pp.	0.41%
References	10	≈ 3.93 pp.	14.56%
A. Related Work	14	≈ 1.38 pp.	5.11%
B. Dataset Setup and Benchmarks	15	≈ 4.32 pp.	16%
C. Additional Experiments	19	≈ 8.22 pp.	30.44%

All detected figure and table locations

Main text

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

Attached appendix

• Table 9 · p. 15
• Figure 8 · p. 16
• Figure 9 · p. 17
• Table 10 · p. 18
• Table 11 · p. 19
• Table 12 · p. 20
• Table 13 · p. 20
• Table 14 · p. 21
• Table 15 · p. 21
• Table 16 · p. 22
• Figure 10 · p. 24
• Figure 11 · p. 25
• Figure 12 · p. 26
• Figure 13 · p. 27

Argument and organization

1. Problem and contribution

   A three-paradigm comparison distinguishes explicit geometric inputs, extra encoders and the proposed side module; RGB-only inference does not mean teacher-free geometric training.

2. Argument sequence

   Notation and frozen components precede alternating attention and supervision, followed by depth, spatial understanding, retained general ability and two action-model adaptations.

3. Evidence and verification

   Matched base-model action recipes, block-count and attention controls, loss-weight sweeps and a comparable-parameter LoRA baseline address different alternative explanations.

Problem, claim and contribution

A three-paradigm comparison distinguishes explicit geometric inputs, extra encoders and the proposed side module; RGB-only inference does not mean teacher-free geometric training.

Figure 2 and Section 3, p. 3; Section 3.3, p. 5 · Main paper

Argument and section order

Notation and frozen components precede alternating attention and supervision, followed by depth, spatial understanding, retained general ability and two action-model adaptations.

Sections 3-4, pp. 3-8 · Main paper

Figures, captions and visual explanation

Task-example panels sit beside protocol-rich tables; the manipulation table explicitly marks robot pretraining and discrete/continuous actions, while component ablations do not hide nonmonotonic relative-depth scores.

Figure 4 and Table 1, p. 5; Tables 4-5, p. 8; both PNGs inspected · Main paper

Tables, comparisons and evidence

Matched base-model action recipes, block-count and attention controls, loss-weight sweeps and a comparable-parameter LoRA baseline address different alternative explanations.

Sections 4.4-4.5, pp. 8-9; Appendix C.1, pp. 19-20 · Main paper

Appendix and reproducibility

Detailed tables unpack average general-capability retention into task-specific degradation and shared route-planning failures, explaining what aggregate averages conceal.

Appendix C.3-C.4 and Tables 14-16, pp. 21-22 · Main paper

Clarity and writing conventions

The limitations explicitly acknowledge pseudo-label dependence, a geometry-versus-motion trade-off and a need for further safety-critical verification despite broad physical-intelligence language.

Limitations, p. 9 · Main paper

What to reuse—and what not to infer

Transferable lesson: Pair a preservation claim with per-task losses and failure overlap, and make training-time external supervision distinct from inference-time inputs.

Boundary: Frozen backbone weights do not guarantee unchanged end-to-end behavior; RGB-only inference still benefits from geometric teacher supervision, and simulation success is not safety-critical deployment validation.

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 1-5, pp. 1-9: RGB-only inference motivation, frozen-backbone side module, geometric supervision, spatial/general/action evaluations, component controls and limitations.

Appendix / supplement

Writing-focused appendix scan; code, data and external assets not executed. Appendices A-C, pp. 14-27: related-work distinctions, training data and benchmark construction, LoRA comparison, task-granularity discussion, general-capability degradation and qualitative examples.

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. · 27 pages