Writing study: Decentralized Attention Fails Centralized Signals: Rethinking Transformers for Medical Time Series

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Yu Guoqi, Juncheng Wang, Chen Yang, Jing Qin, Angelica Aviles-Rivero, Shujun Wang · 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).

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

21 physical pages

Before references / appendix

≈ 10.11 page-equivalents

Main figures / tables

5 / 5

Attached appendix figures / tables

0 / 7

Attached appendix span

≈ 5.89 page-equivalents

Before references / appendix: ≈ 48.14%References / intervening material: ≈ 23.81%Attached appendix: ≈ 28.05%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	1	≈ 1.45 pp.	6.9%
2 RELATED WORK	3	≈ 0.59 pp.	2.81%
3 PRELIMINARIES	3	≈ 0.41 pp.	1.95%
4 METHOD	4	≈ 2.16 pp.	10.29%
5 EXPERIMENTS	6	≈ 4.35 pp.	20.71%
6 CONCLUSION	10	≈ 0.19 pp.	0.9%
7 ACKNOWLEDGEMENTS	10	≈ 0.3 pp.	1.43%
REFERENCES	11	≈ 5 pp.	23.81%
A DATA AUGMENTATION BANKS	16	≈ 0.38 pp.	1.81%
B DATA PREPROCESSING	16	≈ 1.62 pp.	7.71%
C IMPLEMENTATION DETAILS	18	≈ 3.89 pp.	18.52%

All detected figure and table locations

Main text

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

Attached appendix

• Table 6 · p. 19
• Table 7 · p. 20
• Table 8 · p. 20
• Table 9 · p. 20
• Table 10 · p. 21
• Table 11 · p. 21
• Table 12 · p. 21

Argument and organization

1. Introduction / Related Work

   Frames a proposed mismatch between physiological dependencies and token interaction.

2. Preliminaries / Method

   Defines subject-independent evaluation before module equations and full architecture.

3. Experiments / Conclusion

   Combines main classification tables with efficiency, perturbations, component tests and visualization.

4. Appendices A–C

   Documents split identities, dataset-specific configuration and stronger follow-up comparisons.

Problem, claim and contribution

The opening distinguishes temporal from channel dependencies and ties its proposed module to the latter. Defining subject-independent splitting before the architecture makes the evaluation unit visible before presenting large sample counts.

§§1, 3, pp.1–3 · Main paper

Argument and section order

The module is specified first with tensor shapes and aggregation/redistribution equations; the next subsection embeds it into tunable temporal and channel branches. This separates a reusable interaction operation from the complete classifier.

§§4.1–4.2, Eqs.2–5, pp.4–6 · Main paper

Figures, captions and visual explanation

Figure 2 contrasts pairwise attention and core-token processing at the same visual scale. Figure 3 then carries colors and tensor shapes from tokenization through branch averaging and final prediction, clarifying where the module is substituted.

Figures 2–3, pp.4–5 · Main paper

Tables, comparisons and evidence

The results include five-seed standard deviations, a fixed hardware/batch-size efficiency comparison and explicit ablations of tokenization versus token interaction. The main discussion acknowledges the ADFTD aggregate exception instead of hiding that result.

§§5.1–5.3, pp.6–9 · Main paper

Appendix and reproducibility

Exact subject IDs and preprocessing rules support leakage-aware reuse. Dataset-specific branch depths and later five-fold subject-level tests reveal what a short claim of adaptive tokenization and generalization actually entails.

Appendices B, C.2, C.5, pp.16–19, 21 · Main paper

Clarity and writing conventions

The biological analogy and t-SNE discussion exceed the evidence when interpreted as proof of physiological mechanisms. No dedicated limitations section separates architectural motivation from empirically established causal structure.

§5.4, pp.9–10; Appendix C.6, p.21 · Main paper

What to reuse—and what not to infer

Transferable lesson: Present the reusable module and complete network at separate scales, attach dimensions to diagrams, and identify the independent evaluation unit before reporting sample totals.

Boundary: All-to-all attention does not mean equal attention weights. The 11.6% APAVA gain is relative change in a six-metric average, not 11.6 accuracy points. Table 12 shows MedGNN beating TeCh on some datasets. TDBrain subject counts differ between Table 1 and Appendix B.2; verify the released split before reuse.

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

PDF pp.1–10, §§1–7; references pp.11–15

Appendix / supplement

PDF pp.16–21, Appendices A–C: augmentation, preprocessing, implementation, complete ablations, subject-level cross-validation and centralization analysis

Visually inspected PDF pages

4, 5

• official ICLR 2026 proceedings PDF · 21 pages