Writing study: Extending Sequence Length is Not All You Need: Effective Integration of Multimodal Signals for Gene Expression Prediction

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Zhao Yang, Yi Duan, Jiwei Zhu, Ying Ba, Chuan Cao, Bing Su · 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: Learning & optimization algorithms (primary) · Empirical analysis, diagnosis & auditing.

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

22 physical pages

Before references / appendix

≈ 9.87 page-equivalents

Main figures / tables

5 / 3

Attached appendix figures / tables

1 / 13

Attached appendix span

≈ 9.15 page-equivalents

Before references / appendix: ≈ 44.86%References / intervening material: ≈ 13.55%Attached appendix: ≈ 41.59%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	≈ 1.78 pp.	8.09%
2 CURRENT METHODS DO NOT BENEFIT FROM LONG SEQUENCE INPUT	3	≈ 1.04 pp.	4.73%
3 METHOD	4	≈ 2.26 pp.	10.27%
4 EXPERIMENTS	6	≈ 2.93 pp.	13.32%
5 RELATED WORK	9	≈ 0.72 pp.	3.27%
6 CONCLUSION	10	≈ 0.17 pp.	0.77%
ETHICS STATEMENT	10	≈ 0.1 pp.	0.45%
REPRODUCIBILITY STATEMENT	10	≈ 0.14 pp.	0.64%
ACKNOWLEDGEMENTS	10	≈ 0.1 pp.	0.45%
REFERENCES	10	≈ 2.98 pp.	13.55%
A THE USE OF LARGE LANGUAGE MODELS (LLMS)	13	≈ 0.25 pp.	1.14%
B SHORTENING INPUT SEQUENCE LENGTH AT TEST TIME	14	≈ 0.6 pp.	2.73%
C EXPERIMENTAL DATA OF TABLE 1	14	≈ 0.94 pp.	4.27%
D CASE STUDY AND QUANTITATIVE EVIDENCE OF WIDESPREAD	15	≈ 1.47 pp.	6.68%
E ALGORITHM WORKFLOW	17	≈ 0.47 pp.	2.14%
F MORE IMPLEMENTATION DETAILS	17	≈ 1.11 pp.	5.05%
G THE EFFECT OF PRE-TRAINING	18	≈ 0.73 pp.	3.32%
H EXTENDED ANALYSIS WITH ADDITIONAL EPIGENOMIC SIGNALS	19	≈ 1.69 pp.	7.68%
I CROSS-CELL GENERALIZATION	21	≈ 0.32 pp.	1.45%
J PERFORMANCE ACROSS DIFFERENT INPUT LENGTHS	21	≈ 0.42 pp.	1.91%
K PERFORMANCE ON H1 CELL LINE	21	≈ 0.45 pp.	2.05%
L LEARNING CONFOUNDER WEIGHTS WITHOUT SUPERVISION	22	≈ 0.71 pp.	3.23%

All detected figure and table locations

Main text

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

Attached appendix

• Table 4 · p. 14
• Table 5 · p. 15
• Table 6 · p. 15
• Table 7 · p. 15
• Table 8 · p. 16
• Figure 6 · p. 16
• Table 9 · p. 17
• Table 10 · p. 19
• Table 11 · p. 20
• Table 12 · p. 20
• Table 13 · p. 21
• Table 14 · p. 21
• Table 15 · p. 22
• Table 16 · p. 22

Argument and organization

1. Current methods do not benefit from long sequence input

   Test an architectural motivation empirically before proposing a replacement emphasis.

2. Method

   Separate a hypothesized causal graph from its learned weighting implementation and regularized objective.

3. Experiments and extended analyses

   Measure predictive quality and study signal, length and parameter choices.

Problem, claim and contribution

The introduction challenges context length with two complementary diagnostics: retraining at different lengths and truncating a long-trained model at test time. That distinction strengthens the motivation beyond a single favorable short-input comparison.

§§1–2, Figs.1–2, pp.1–4; Appendix B, p.14 · Main paper

Argument and section order

The causal graph, adjustment formula and implemented learned feature weights have separate subsections. This is a useful explanatory hierarchy, but the graph remains a modeling hypothesis rather than a causal relationship established by predictive accuracy.

§§3.2–3.4, pp.4–6 · Main paper

Figures, captions and visual explanation

The architecture diagram splits epigenomic input into feature and weight encoders while routing DNA separately into the predictor. Explicit dimensions and multiple weighted feature groups clarify what changes relative to signal concatenation.

Fig.4, p.5 · Main paper

Tables, comparisons and evidence

The main results report five-seed variation and distinct error/correlation metrics. Chromosome-based splits are specified, while the later mixed-cell experiment trains on both evaluated cell types; it does not establish held-out-cell generalization.

§4.1–4.2, pp.7–8; Appendix I, p.21 · Main paper

Appendix and reproducibility

The implementation appendix reveals that intermediate representations are averaged before a shared prediction head, whereas the main equation averages predictions. Stating their equivalence condition is important for reproducing the actual computation.

Eq.2–3, p.6; Appendix F.3, p.18 · Main paper

Clarity and writing conventions

Some biological and statistical language is stronger than the controls justify. Structured learned weights are not independently annotated chromatin states, and differences exceeding one standard deviation are not formal significance tests.

§4.2–4.4, pp.7–9; Appendix L, p.22 · Main paper

What to reuse—and what not to infer

Transferable lesson: Distinguish the motivating empirical failure, the hypothesized mechanism and the actual implemented operator; validate each with the corresponding control.

Boundary: The task predicts CAGE with DNA plus measured epigenomic and auxiliary features, not expression from DNA alone or WSI. Uniform learned contexts do not establish causal identification. Soft weight averages are not equivalent to binary dropout retention rates. Long-context findings are architecture/task-specific, and appendix percentage changes conflict with their raw signal-removal table.

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, §§1–6 and statements; references pp.10–13

Appendix / supplement

Appendices A–L, PDF pp.13–22, including all length, signal, pretraining, cell-line and dropout analyses

Visually inspected PDF pages

5, 8

• official ICLR 2026 proceedings PDF · 22 pages