Writing study: Holi-Spatial: Evolving Video Streams into Holistic 3D Spatial Intelligence

ICML overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICML 2026 oral writing corpus

Yuanyuan Gao, Hao Li, Yifei Liu, Xinhao Ji, Yuning Gong, Yuanjun Liao, Fangfu Liu, Manyuan Zhang, Yuchen Yang, Dan Xu, Xue Yang, Huaxi Huang, Hongjie Zhang, Ziwei Liu, Xiao Sun, Dingwen Zhang, Zhihang Zhong · 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: Datasets, benchmarks & measurement (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

17 physical pages

Before references / appendix

≈ 9.13 page-equivalents

Main figures / tables

10 / 5

Attached appendix figures / tables

6 / 0

Attached appendix span

≈ 4.91 page-equivalents

Before references / appendix: ≈ 53.71%References / intervening material: ≈ 17.41%Attached appendix: ≈ 28.88%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	≈ 0.99 pp.	5.82%
2. Related Work	3	≈ 0.71 pp.	4.18%
3. Method	4	≈ 1.32 pp.	7.76%
4. Dataset	5	≈ 1.4 pp.	8.24%
5. Experiment	6	≈ 2.6 pp.	15.29%
6. Conclusion	9	≈ 0.32 pp.	1.88%
Acknowledgment	9	≈ 0.05 pp.	0.29%
Impact Statement	9	≈ 0.35 pp.	2.06%
References	10	≈ 2.95 pp.	17.35%
A. Appendix	13	≈ 4.91 pp.	28.88%

All detected figure and table locations

Main text

• Figure 1 · p. 1
• Table 1 · p. 2
• Figure 2 · p. 2
• Figure 3 · p. 5
• Figure 4 · p. 5
• Figure 5 · p. 6
• Table 2 · p. 6
• Figure 6 · p. 7
• Figure 7 · p. 8
• Figure 8 · p. 8
• Table 4 · p. 8
• Table 3 · p. 8
• Figure 9 · p. 9
• Figure 10 · p. 9
• Table 5 · p. 9

Attached appendix

• Figure 11 · p. 13
• Figure 12 · p. 14
• Figure 13 · p. 15
• Figure 14 · p. 15
• Figure 15 · p. 16
• Figure 16 · p. 16

Argument and organization

1. Problem and contribution

   The introduction locates the bottleneck in limited and inconsistent spatial annotations and separates a curation pipeline from the models later fine-tuned on its outputs.

2. Argument sequence

   Three curation stages lead to a dataset-composition section before evaluation, which separates framework annotation quality, downstream finetuning and refinement ablations.

3. Evidence and verification

   Framework evaluation uses 10 manually annotated scenes from each of three datasets, distinct from the 12K-scene generated corpus; refinement ablations expose filtering's recall loss before agent recovery.

Problem, claim and contribution

The introduction locates the bottleneck in limited and inconsistent spatial annotations and separates a curation pipeline from the models later fine-tuned on its outputs.

Introduction and capability matrix, p. 2 · Main paper

Argument and section order

Three curation stages lead to a dataset-composition section before evaluation, which separates framework annotation quality, downstream finetuning and refinement ablations.

Sections 3-5, pp. 4-9 · Main paper

Figures, captions and visual explanation

The overview maps geometry, image proposals and scene refinement to named data products; qualitative comparisons align methods on the same scenes while downstream tables retain missing entries.

Figures 3-4, p. 5; Figures 7-8 and Tables 3-4, p. 8; both PNGs inspected · Main paper

Tables, comparisons and evidence

Framework evaluation uses 10 manually annotated scenes from each of three datasets, distinct from the 12K-scene generated corpus; refinement ablations expose filtering's recall loss before agent recovery.

Section 5.1, p. 6; Section 5.3 and Table 5, pp. 8-9 · Main paper

Appendix and reproducibility

Full category-memory and verifier prompts expose acceptance rules, including accepting a correct core noun despite unverifiable modifiers; the final algorithm makes keep/discard/verify branches explicit.

Figures 12-13, pp. 14-15; Algorithm 1, p. 17 · Main paper

Clarity and writing conventions

Limitations name dependence on upstream foundation models and expensive per-scene optimization; the automated-corpus claim should remain separate from the manual work creating evaluation ground truth.

Section 5.1, p. 6; Impact Statement, pp. 9-10 · Main paper

What to reuse—and what not to infer

Transferable lesson: For an automatically curated dataset, separate corpus-scale production, sampled quality assessment and downstream utility, and show the precision-recall cost of filtering.

Boundary: Four million mixed annotation items are not four million independent scenes; sampled quality and foundation-model verification do not establish human-level correctness across the entire corpus.

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 1-6, pp. 2-9: annotation-scale motivation, geometry/perception/refinement pipeline, dataset composition, evaluation, stage ablations and limitations.

Appendix / supplement

Writing-focused appendix scan; code, data and external assets not executed. Appendix A, pp. 13-17: multi-view merge rationale, verification and category prompts, qualitative QA/grounding examples and scene-refinement pseudocode.

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. · 17 pages