Writing study: Towards Hierarchy–Uniformity Equilibrium: Recovering Semantic Depth in Hypergraph Contrastive Learning

ICML overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICML 2026 oral writing corpus

Ruiting Zhao, Ming Li, Lixin Cui, Lu Bai, Feilong Cao, Ke Lv, Pietro Lio · 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: 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

27 physical pages

Before references / appendix

≈ 9.19 page-equivalents

Main figures / tables

6 / 2

Attached appendix figures / tables

5 / 4

Attached appendix span

≈ 14.91 page-equivalents

Before references / appendix: ≈ 34.04%References / intervening material: ≈ 10.74%Attached appendix: ≈ 55.22%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	≈ 1.1 pp.	4.07%
2. Related Work	2	≈ 0.92 pp.	3.41%
3. Preliminaries	3	≈ 0.35 pp.	1.3%
4. Theoretical Formulation and Geometric	4	≈ 0.72 pp.	2.67%
5. Proposed Method	4	≈ 2.33 pp.	8.63%
6. Theoretical Analysis	7	≈ 0.59 pp.	2.19%
7. Experiments	7	≈ 2.13 pp.	7.89%
8. Conclusion	9	≈ 0.2 pp.	0.74%
Impact Statement	10	≈ 0.08 pp.	0.3%
Acknowledgments	10	≈ 0.06 pp.	0.22%
References	10	≈ 2.9 pp.	10.74%
A. Notation Summary	13	≈ 0.59 pp.	2.19%
B. Additional Details for Theoretical Analysis	13	≈ 4.41 pp.	16.33%
C. Details for Algorithm Implementation	18	≈ 1.13 pp.	4.19%
D. Further Details for Experimental Studies	19	≈ 8.78 pp.	32.52%

All detected figure and table locations

Main text

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

Attached appendix

• Table 3 · p. 13
• Table 4 · p. 19
• Table 5 · p. 20
• Table 6 · p. 22
• Figure 7 · p. 23
• Figure 8 · p. 24
• Figure 9 · p. 25
• Figure 10 · p. 26
• Figure 11 · p. 27

Argument and organization

1. 1 Introduction; 2 Related Work; 3 Preliminaries; 4 Theoretical Formulation

   Names the hierarchy–uniformity conflict and defines the competing geometric preferences.

2. 5 Proposed Method; 6 Theoretical Analysis

   Separates spectral encoding, hierarchical alignment and local contrastive learning, then states idealized subspace assumptions.

3. 7 Experiments; 8 Conclusion

   Maps four questions to benchmark comparisons, component removals, prototype visualizations and parameter sensitivity.

Problem, claim and contribution

A named failure mode connects the motivation, geometric formulation and two complementary objectives, but the formal equilibrium claim is conditional on spectral separation.

Sections 1,4,6, physical pp.1–4,7 · Main paper

Argument and section order

The method identifies where local and global losses act before the theoretical section states exact separation assumptions.

Sections 5.4–6.1, pp.6–7 · Main paper

Figures, captions and visual explanation

The architecture diagram links two spectral branches with two losses; the later three-panel prototype display compares global-only, local-only and joint objectives beside quantitative ablation bars.

Figure 2, p.5; Figures 4–5, p.9 · Main paper

Tables, comparisons and evidence

Self-supervised pretraining uses the full graph structure/features, followed by a frozen-encoder linear probe on10/10/80 label splits across five seeds. Supervised baselines must retain their distinct label-access condition.

Section 7.2, pp.7–8; Appendix D.2, p.19 · Main paper

Appendix and reproducibility

The implementation appendix separates periodic detached prototype re-anchoring from gradient updates and supplies a validation search space plus dataset-specific settings.

Appendix D.2–D.3 and Algorithm 3, pp.19–20 · Main paper

Clarity and writing conventions

Ablation, hierarchy visualization and sensitivity each answer a different stated question; a qualitative prototype layout is used as explanatory evidence rather than a new accuracy metric.

Sections 7.4–7.6, pp.8–9 · Main paper

What to reuse—and what not to infer

Transferable lesson: When two losses are claimed to cooperate, show their computational paths, state the assumptions for theoretical compatibility, and compare single-loss versus joint behavior both quantitatively and visually.

Boundary: Existence of an equilibrium under ideal subspace assumptions does not prove that finite, gated spectral filters achieve exact separation or that training reaches that equilibrium. Learned prototype hierarchy is not independently verified semantic ground truth, and full-graph self-supervised pretraining is a transductive protocol, not unseen-graph deployment evidence.

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 scan across physical PDF pp.1–9, covering framing, section progression, definitions, method/evidence explanations, captions and scope. Physical pages extracted individually to avoid form-feed pagination errors. Selected full-page rasters inspected; not exhaustive line-by-line technical verification or experimental reproduction.

Appendix / supplement

Appendix A–D heading inventory checked; selected detailed reading of spectral projector definitions B.1 (p.13), linear-evaluation protocol and prototype rectification D.2 (p.19), and hyperparameter search/algorithm D.3 (p.20). Long proofs and all dataset-specific sensitivity results were not independently verified.

Visually inspected PDF pages

5, 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 proceedings PDF, PMLR volume 306, retrieved 2026-10-07 · 27 pages