Writing study: Modality-Balanced Collaborative Distillation for Multi-Modal Domain Generalization

AAAI overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← AAAI 2026 oral writing corpus

Xiaohan Wang, Zhangtao Cheng, Ting Zhong, Leiting Chen, Fan Zhou · Official AAAI 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

9 physical pages

Before references / appendix

≈ 7.05 page-equivalents

Main figures / tables

6 / 5

Attached appendix figures / tables

0 / 0

Attached appendix span

No boundary detected; not assumed zero

Before references / appendix: ≈ 78.33%References / intervening material: ≈ 21.67%Estimated layout envelopes, not text/artwork area. Unresolved appendix boundaries remain within the undivided post-body material.

Extraction notes: No attached appendix boundary detected. Zero means no recognized caption, not independently verified absence.

Source-bound correction: Checked 2026-10-25 against the authored writing scan and physical-page Poppler boxes. These are unnumbered main headings, distinct from the smaller named method/analysis subsections. Coordinates use left/right column plus within-page y. Captions match Figures 1–6 and Tables 1–5; no attached appendix in the nine-page publisher PDF.

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
Introduction	1	≈ 1.07 pp.	11.89%
Related Work	2	≈ 0.5 pp.	5.56%
Methodology	2	≈ 2 pp.	22.22%
Experiments	4	≈ 2.89 pp.	32.11%
Conclusion	7	≈ 0.11 pp.	1.22%
Acknowledgments	7	≈ 0.13 pp.	1.44%
References	8	≈ 1.95 pp.	21.67%

All detected figure and table locations

Main text

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

Attached appendix

No recognized caption labels.

Argument and organization

1. Introduction → Related Work

   Diagnose modality imbalance as a limitation of directly transferring weight averaging, then locate the intervention among generalization and multimodal optimization approaches.

2. Methodology

   Define source/target domains, then introduce adaptive dropout, gradient consistency and EMA-guided collaborative distillation as separate mechanisms before the combined loss.

3. Experiments → Further Analysis → Conclusion

   Separate cross-domain and in-domain comparisons, test component variants, then use perturbation and representation diagnostics to examine the proposed explanation.

Problem, claim and contribution

The opening connects a familiar technique, weight averaging, to a specific failure under unequal modality learning speeds. Its contributions follow diagnosis → intervention → evaluation, giving the reader a reason for each component rather than presenting a module list without a problem.

Abstract and Introduction, physical pp.1–2; Figure 1 · Main paper

Argument and section order

The method names the three interventions before deriving their operations; the final weighted objective arrives only after the component definitions. Experimental headings distinguish multi-source, single-source and in-domain tests, so these are not silently pooled into one generalization claim.

Methodology pp.2–4; Experiments pp.4–6; Eq.9 · Main paper

Figures, captions and visual explanation

Figure 2 uses a top-level student/EMA-teacher view and a lower zoom into dominance scoring and dropout. Encoders, features, heads and losses have a legend, while arrow routing exposes the supervisory relationship. Figure 3 answers a different question through perturbation-response plots; it is not another architecture diagram.

Figure 2 physical p.3; Figure 3 physical p.6 · Main paper

Tables, comparisons and evidence

Tables 1–2 name the source-to-target direction and modality access and report per-split mean and standard deviation over three seeds. Table 4 isolates EMA and selected added components; Table 5 compares fixed and adaptive dropout. Retain non-wins: the HAC video+audio average in Table 1 is 70.66 for MBCD versus 71.25 for CMRF.

Tables 1–2 physical p.5; Tables 4–5 physical p.6 · Main paper

Appendix and reproducibility

The publisher version supplies the central losses and interpretation in the main body and a code link on its first page, then ends with references rather than an attached appendix. A future drafting plan should place exact hyperparameters, model-selection rules and full implementation details in accessible support; that recommendation is not a claim that such support was inspected here.

Code link p.1; Eqs.5–9 p.4; references pp.8–9 · Main paper

Clarity and writing conventions

The strongest reusable convention is the repeated mapping from imbalance to a named corrective operation. Do not imitate the paper’s broad superiority language without qualifications: average gains do not establish every-split wins, and a t-SNE picture or local loss-flatness diagnostic does not independently prove the asserted mechanism.

Introduction p.2; Table 1 p.5; Further Analysis and Conclusion pp.6–7 · Main paper

What to reuse—and what not to infer

Transferable lesson: Tie every distillation component to a stated failure, distinguish the teacher’s update from its supervision, and keep cross-domain comparison conditions and component controls adjacent to the numbers.

Boundary: This study evaluates writing and selected visuals, not the validity of the derivation or reproducibility of the code. The two action-recognition benchmarks do not establish pathology transfer, universal domain robustness or causal proof from embedding plots.

Apply this lesson using the argument and evidence worksheet.

Inspection record and sources

Checked

2026-10-25

Reading status

Writing-focused manuscript scan + selected visuals inspected

Main text

Writing-oriented scan of the introduction, method and experiments on physical pp.1–7; references/end inspected for structure, not citation verification or independent scientific reproduction.

Appendix / supplement

The inspected nine-page publisher PDF ends with references and contains no attached appendix. External code, independent supplementary assets and implementation details there were not inspected.

Visually inspected PDF pages

3, 6

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.

• Official AAAI-26 published proceedings PDF retrieved 2026-10-25, printed pp.26535–26543 · 9 pages