Writing study: On the Generalization Capacities of MLLMs for Spatial Intelligence

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Gongjie Zhang, Wenhao Li, Quanhao Qian, Jiuniu Wang, Deli Zhao, Shijian Lu, Ran Xu · 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) · Empirical analysis, diagnosis & auditing.

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

15 physical pages

Before references / appendix

≈ 9.11 page-equivalents

Main figures / tables

7 / 5

Attached appendix figures / tables

0 / 1

Attached appendix span

≈ 2.89 page-equivalents

Before references / appendix: ≈ 60.73%References / intervening material: ≈ 20%Attached appendix: ≈ 19.27%Estimated layout envelopes, not text/artwork area. Unresolved appendix boundaries remain within the undivided post-body material.

Source-bound correction: Caption audit 2026-09-19, DE5ZJtR4bg: rendered physical PDF page 6 confirms Figure 5, the camera-aware geometric augmentation diagram. The bbox line merges preceding body text with the bare caption label and places caption text in following lines; the parser therefore missed the real caption. Position uses caption yMin 698.501357 / page height 792 plus page index 5.

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	≈ 1.15 pp.	7.67%
2 RELATED WORK	2	≈ 0.94 pp.	6.27%
3 BRITTLENESS OF CAMERA-AGNOSTIC MLLMS IN SPATIAL REASONING	3	≈ 1.95 pp.	13%
4 CAMERA-AWARE MLLM FRAMEWORK	5	≈ 1.85 pp.	12.33%
5 EXPERIMENTS	7	≈ 2.37 pp.	15.8%
6 CONCLUSION	9	≈ 0.32 pp.	2.13%
REFERENCES	10	≈ 3 pp.	20%
A DETAILED PROOF OF GEOMETRIC AMBIGUITY	13	≈ 1.28 pp.	8.53%
B IMPLEMENTATION DETAILS	14	≈ 0.72 pp.	4.8%
C LLM USAGE DISCLOSURE	15	≈ 0.89 pp.	5.93%

All detected figure and table locations

Main text

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

Attached appendix

• Table 6 · p. 14

Argument and organization

1. Camera-dependent ambiguity and robustness diagnosis

   Connects a physical identifiability argument to controlled image-rescaling failures.

2. Camera-aware representation

   Introduces ray embeddings, intrinsics-consistent augmentation and a pretrained geometry prior.

3. Evaluation and analysis

   Separates broad spatial benchmarks from controlled camera-shift ablations.

Problem, claim and contribution

The focal-length/object-size/depth relationship supplies an unusually concrete opening failure mechanism. The appendix restricts the geometric argument more carefully than the broad introductory claim about RGB-only spatial intelligence.

Sections 1–3; Appendix A, pages 13–14. · Main paper

Figures, captions and visual explanation

The architecture expands a compact overview into three colored contributions, marking frozen and trainable components and showing the additive fusion. The neighboring augmentation figure pairs image changes with updated intrinsics, making the experimental control visible.

Figures 4–5, page 6. · Main paper

Tables, comparisons and evidence

A small component table makes six configurations easy to compare; however, its 20-class protocol differs from the 31-class diagnostic and the table is not a complete factorial design.

Table 1, page 3; Table 5, page 9. · Main paper

The broad benchmark tables retain outcomes where the proposed method is below the base model, which is valuable evidence even though the surrounding state-of-the-art language is stronger than those comparisons support.

Tables 2–4, pages 8–9. · Main paper

Appendix and reproducibility

The supplement records camera transformations, frozen components and training hardware, but does not fully specify the purported geometry-distillation objective. The method description instead describes injection of a pretrained geometric prior.

Section 4.3; Appendix B, page 14. · Main paper

Clarity and writing conventions

Qualitative zoom and unseen-camera examples follow the controlled ablation, making them illustrations rather than the only evidence for robustness.

Figure 7, page 9. · Main paper

What to reuse—and what not to infer

Transferable lesson: Start from a measurable physical ambiguity, show a controlled failure, and make each proposed component correspond to that failure in the architecture and ablation.

Boundary: Known intrinsics do not alone remove absolute-scale ambiguity. Camera-aware conditioning and an auxiliary geometry model are not automatically teacher-free knowledge distillation; broad spatial reasoning claims require more than synthetic zoom robustness.

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 1–9, Sections 1–6; references on pages 10–12.

Appendix / supplement

Pages 13–15, Appendices A–C: geometric analysis, implementation and language-model disclosure.

Visually inspected PDF pages

6, 9

• ICLR 2026 final proceedings · 15 pages