Writing study: Binding Multiple Modalities via Multimodal Wasserstein Barycenter

NeurIPS overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← NeurIPS 2026 oral writing corpus

Xiaole Tang, Jiayi Xu, Xiang Gu, Yan Yang, Jian Sun · Official NeurIPS 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

35 physical pages

Before references / appendix

≈ 9.57 page-equivalents

Main figures / tables

5 / 6

Attached appendix figures / tables

3 / 18

Attached appendix span

≈ 14 page-equivalents

Before references / appendix: ≈ 27.34%References / intervening material: ≈ 32.66%Attached appendix: ≈ 40%Estimated layout envelopes, not text/artwork area. Unresolved appendix boundaries remain within the undivided post-body material.

Source-bound correction: Top-level heading and reference/appendix boundaries checked against source PDF text. Bibliography titles/author initials, numbered contribution/property bullets, diagram steps, equations and TOC entries are not substantive headings. Section spans are page-coordinate estimates, not ink area or prose-only length. This is an author manuscript mirror, not a verified accepted-version template. Source layout uses physical single-column y/page-height, including equation-heavy pages.

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	≈ 1.66 pp.	4.74%
2 Related Works	3	≈ 0.38 pp.	1.09%
3 Method	3	≈ 3.28 pp.	9.37%
4 Experiments	7	≈ 3.19 pp.	9.11%
5 Conclusion	10	≈ 0.16 pp.	0.46%
Acknowledgments and Disclosure of Funding	10	≈ 0.12 pp.	0.34%
References	10	≈ 4.53 pp.	12.94%
A Theoretical Results	15	≈ 4.38 pp.	12.51%
B Experimental Details and More Results	19	≈ 9.62 pp.	27.49%
NeurIPS Paper Checklist	29	≈ 6.91 pp.	19.74%

All detected figure and table locations

Main text

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

Attached appendix

• Figure 6 · p. 19
• Table 7 · p. 20
• Table 8 · p. 21
• Table 9 · p. 22
• Figure 7 · p. 22
• Table 10 · p. 23
• Table 11 · p. 23
• Table 12 · p. 23
• Table 13 · p. 24
• Table 14 · p. 24
• Table 15 · p. 24
• Table 16 · p. 24
• Table 17 · p. 25
• Table 18 · p. 25
• Table 19 · p. 25
• Table 20 · p. 26
• Table 21 · p. 26
• Table 22 · p. 26
• Table 23 · p. 27
• Table 24 · p. 27
• Figure 8 · p. 28

Argument and organization

1. Problem and contribution

   Identifies the asymmetry created by choosing one observed modality as the alignment anchor, motivating a learned shared barycenter rather than simply adding more pairwise losses.

2. Argument sequence

   Introduces Wasserstein and simplex-volume preliminaries before separating barycenter optimization from volumetric alignment, then tests retrieval directions, missing modalities and individual loss components.

3. Evidence and verification

   Main tables distinguish task and retrieval direction; ablations separate the barycenter and volume terms. Appendix timing distinguishes similarity computation from whole forward/backward training overhead.

Problem, claim and contribution

Identifies the asymmetry created by choosing one observed modality as the alignment anchor, motivating a learned shared barycenter rather than simply adding more pairwise losses.

Introduction and Figures 1–2, pp. 1–2 · Main paper

Argument and section order

Introduces Wasserstein and simplex-volume preliminaries before separating barycenter optimization from volumetric alignment, then tests retrieval directions, missing modalities and individual loss components.

Sections 3–4, pp. 3–10 · Main paper

Figures, captions and visual explanation

Figure 3 uses two named regions for the two optimization processes; modality colors persist through encoder tokens, distribution diagrams, barycenter mapping and a geometric simplex, with losses placed beside their inputs. The result tables group methods by available modalities and label text-to-video versus video-to-text retrieval separately. Smaller tables distinguish missing audio during training from missing video during inference; colored deltas name VAST as the reference instead of implying every row has identical information access.

Figure 3, p. 3 (render inspected); Tables 2–4, p. 8 (additional rendered inspection) · Main paper

Tables, comparisons and evidence

Main tables distinguish task and retrieval direction; ablations separate the barycenter and volume terms. Appendix timing distinguishes similarity computation from whole forward/backward training overhead.

Tables 2–6, pp. 8–10; Tables 9 and 19, pp. 22 and 25 · Main paper

Appendix and reproducibility

Proofs are separate from pseudocode and empirical controls, including initializer, batch size, modality count, training overhead and generation/retrieval examples.

Appendices A–B, pp. 15–28 · Main paper

Clarity and writing conventions

The two-stage vocabulary is consistent across overview, mathematics and experiments; distribution plots illustrate geometric motivation but should not be described as proving semantic balance.

Section 3, pp. 3–6; Figures 2 and 4, pp. 2 and 6 · Main paper

What to reuse—and what not to infer

Transferable lesson: For a multi-loss alignment method, draw where each objective acts, ablate terms separately and distinguish local operation timing from end-to-end cost.

Boundary: A t-SNE projection is an illustration, not a guarantee of balanced semantics. The public preprint is not verified to be the accepted revision.

Apply this lesson using the argument and evidence worksheet.

Inspection record and sources

Checked

2026-10-02

Reading status

Writing-focused manuscript scan + selected visuals inspected

Main text

Writing-oriented scan across manuscript framing, section sequence, figure/table captions, selected argument/evaluation prose and conclusion; not exhaustive technical or proof verification. Sections 1–5, pp. 1–10; references pp. 11–14. Source is the exact public author version listed below, not a verified accepted/camera-ready revision.

Appendix / supplement

Writing-oriented appendix/documentation scan: Appendix A proofs pp. 15–18; Appendix B algorithms, protocols and further analyses pp. 19–28; NeurIPS checklist pp. 29–35. Source code, data, model outputs and external video assets were not executed or independently validated.

Visually inspected PDF pages

3, 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.

• Public author arXiv manuscript 2609.33800v1; not verified identical to the accepted or camera-ready OpenReview revision · 35 pages