Writing study: dnaHNet: A Scalable and Hierarchical Foundation Model for Genomic Sequence Learning

ICML overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICML 2026 oral writing corpus

Arnav Shah, Junzhe Li, Parsa Idehpour, Adibvafa Fallahpour, Brandon Wang, Sukjun Hwang, Bo Wang, Patrick D Hsu, Hani Goodarzi, Albert Gu · Official ICML 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 (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

19 physical pages

Before references / appendix

≈ 9.11 page-equivalents

Main figures / tables

6 / 1

Attached appendix figures / tables

2 / 8

Attached appendix span

≈ 5.91 page-equivalents

Before references / appendix: ≈ 47.95%References / intervening material: ≈ 20.95%Attached appendix: ≈ 31.11%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	≈ 1.19 pp.	6.26%
2. Related Work	2	≈ 0.79 pp.	4.16%
3. dnaHNet	3	≈ 2.38 pp.	12.53%
4. Experiments	5	≈ 3.12 pp.	16.42%
5. Discussion	8	≈ 0.92 pp.	4.84%
6. Conclusion	9	≈ 0.13 pp.	0.68%
Impact Statement	9	≈ 0.16 pp.	0.84%
Acknowledgements	10	≈ 0.07 pp.	0.37%
References	10	≈ 3.98 pp.	20.95%
A. Appendix	14	≈ 5.91 pp.	31.11%

All detected figure and table locations

Main text

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

Attached appendix

• Table 2 · p. 14
• Table 3 · p. 14
• Table 4 · p. 15
• Figure 7 · p. 15
• Figure 8 · p. 16
• Table 5 · p. 16
• Table 6 · p. 17
• Table 7 · p. 17
• Table 8 · p. 18
• Table 9 · p. 19

Argument and organization

1. Problem and contribution

   Connects variable biological granularity to the limitations of fixed tokenization and asks whether learned segmentation is both efficient and biologically meaningful.

2. Argument sequence

   Architecture is followed by stability modifications and an explicit training/inference distinction; evaluation separately tests scaling, two prediction tasks, hierarchy and wall-clock cost.

3. Evidence and verification

   Compute-scaled downstream tests are supplemented by synonymous-perturbation and species-related confound controls rather than interpreting raw likelihood drops alone as gene function.

Problem, claim and contribution

Connects variable biological granularity to the limitations of fixed tokenization and asks whether learned segmentation is both efficient and biologically meaningful.

Introduction, pp. 1-2 · Main paper

Argument and section order

Architecture is followed by stability modifications and an explicit training/inference distinction; evaluation separately tests scaling, two prediction tasks, hierarchy and wall-clock cost.

Sections 3-4, pp. 3-8 · Main paper

Figures, captions and visual explanation

The encoder/chunking/decoder diagram explains the one-token output shift in its caption. The efficiency triptych separates throughput, peak memory and latency, with out-of-memory endpoints disclosed.

Figure 1, p. 3; Figure 6, p. 9; both PNGs inspected · Main paper

Tables, comparisons and evidence

Compute-scaled downstream tests are supplemented by synonymous-perturbation and species-related confound controls rather than interpreting raw likelihood drops alone as gene function.

Sections 4.4-4.5, pp. 6-8; Appendix A.7, pp. 16-19 · Main paper

Appendix and reproducibility

Exact numerical scaling results and per-genome results provide denominators, including total and essential-gene counts, beyond the main aggregate plots.

Tables 5-9, pp. 16-19 · Main paper

Clarity and writing conventions

Limitations bound pretraining to prokaryotic genomes, evaluation to zero-shot tasks and tested scale to one billion parameters; eukaryotic extensions remain future work.

Sections 5.1-5.2, p. 9 · Main paper

What to reuse—and what not to infer

Transferable lesson: For a domain architecture, pair representational diagrams with a biological confound check and measured hardware behavior.

Boundary: Boundary alignment and computational prediction are not experimental validation of biological function; the prokaryotic training scope must remain visible.

Apply this lesson using the argument and evidence worksheet.

Inspection record and sources

Checked

2026-10-07

Reading status

Writing-focused manuscript scan + selected visuals inspected

Main text

Writing-focused distributed manuscript scan, not exhaustive technical/proof validation. Sections 1-6, pp. 1-9: genomic tokenization motivation, recursive architecture, stabilization choices, downstream/scaling analyses and domain limitations.

Appendix / supplement

Writing-focused appendix scan; code, data and external assets not executed. Appendix A.1-A.7, pp. 14-19: architecture, training/resources, compression/baseline checks, exact downstream results and gene-essentiality confound analyses.

Visually inspected PDF pages

3, 9

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.

• Published ICML 2026 proceedings PDF, PMLR volume 306; retrieved 2026-10-07. Public OpenReview decision notes were not independently reverified. · 19 pages