Writing study: BioX-Bridge: Model Bridging for Unsupervised Cross-Modal Knowledge Transfer across Biosignals

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Chenqi Li, Yu Liu, Timothy Denison, Tingting Zhu · 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) · Learning & optimization algorithms.

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

27 physical pages

Before references / appendix

≈ 9.6 page-equivalents

Main figures / tables

4 / 3

Attached appendix figures / tables

10 / 9

Attached appendix span

≈ 13.89 page-equivalents

Before references / appendix: ≈ 35.56%References / intervening material: ≈ 13%Attached appendix: ≈ 51.44%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	≈ 1.58 pp.	5.85%
2 RELATED WORKS	3	≈ 1.26 pp.	4.67%
3 METHODS	4	≈ 2.8 pp.	10.37%
4 EXPERIMENTS	7	≈ 2.78 pp.	10.3%
5 CONCLUSION	10	≈ 0.36 pp.	1.33%
ACKNOWLEDGEMENTS	10	≈ 0.14 pp.	0.52%
REFERENCES	10	≈ 3.5 pp.	12.96%
APPENDIX	14	≈ 0.53 pp.	1.96%
A METHOD DETAILS: LINEAR CKA	14	≈ 0.47 pp.	1.74%
B LIMITATIONS AND FUTURE WORK	15	≈ 0.32 pp.	1.19%
C ADDITIONAL DISCUSSION ON RELATED WORKS AND BASELINES	15	≈ 0.84 pp.	3.11%
D EXPERIMENT DETAILS	16	≈ 11.73 pp.	43.44%

All detected figure and table locations

Main text

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

Attached appendix

• Figure A1 · p. 17
• Table A1 · p. 18
• Table A2 · p. 18
• Table A3 · p. 19
• Table A4 · p. 19
• Table A5 · p. 19
• Table A6 · p. 20
• Table A7 · p. 20
• Figure A2 · p. 21
• Figure A3 · p. 22
• Figure A4 · p. 22
• Figure A5 · p. 23
• Figure A6 · p. 24
• Figure A7 · p. 24
• Figure A8 · p. 24
• Figure A9 · p. 25
• Figure A10 · p. 25
• Table A8 · p. 25
• Table A9 · p. 26

Argument and organization

1. Motivation and competing transfer routes

   Contrast signal translation, full student distillation and bridging.

2. Bridge position, architecture and training

   Decompose transfer into layer selection, low-rank prototype projection and final-representation alignment.

3. Directional transfer and ablations

   Evaluate six transfer directions before unpacking parameter and position choices.

Problem, claim and contribution

The introduction makes the deployment motivation tangible through sensors with different cost and comfort, then narrows the obstacle to adapting pretrained models with limited local training resources.

Sections 1–2, pages 1–3 · Main paper

Argument and section order

Layer selection is split into discriminative input probing and output CKA matching before introducing the prototype factorization. Algorithm 1 then reassembles these choices into a reproducible sequence.

Sections 3.3–3.5, pages 5–7 · Main paper

Figures, captions and visual explanation

Figure 2 separates training from inference and uses frozen/trainable icons plus a bridge inset. This is crucial because the old model's tail remains at inference, a distinction easily lost in the headline parameter reduction.

Figure 2, page 4 · Main paper

Tables, comparisons and evidence

Table 1 groups both transfer directions within each dataset and explicitly lists input modality, preventing comparison of teacher and transferred scores as if they used identical information.

Table 1 and Section 4.2, pages 8–9 · Main paper

Appendix and reproducibility

The appendix adds a contents map, preprocessing, subject/sample splits, five-seed deviations, layer heatmaps and a reshuffled-subject experiment. These qualify variability hidden by the compact main table.

Appendices D.1–D.8, pages 16–26 · Main paper

Clarity and writing conventions

The discussion acknowledges weaker modalities and metric tradeoffs, but calling the teacher an absolute oracle is too strong; some transferred balanced accuracies exceed it. The split diagram's subject-independent caption also needs the FOG exception stated.

Section 4.1, page 7; Section 4.2, pages 8–9; Figure A1 and D.1.2, pages 16–17 · Main paper

What to reuse—and what not to infer

Transferable lesson: Use distinct train/inference panels and a direction-aware result table whenever transfer changes both the model path and the information available to it.

Boundary: Trainable-parameter reductions do not prove lower inference cost. Paired biosignals, pretrained models and task-specific teacher labels are prerequisites; these results do not establish WSI or cross-magnification effectiveness.

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 1–10, Sections 1–5 and ethics/reproducibility

Appendix / supplement

Pages 14–27, Appendices A–D.8, all extra results and appendix references

Visually inspected PDF pages

4, 8, 19

• ICLR 2026 final proceedings · 27 pages