Writing study: Uncover Underlying Correspondence for Robust Multi-view Clustering

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Haochen Zhou, Guofeng Ding, Mouxing Yang, Peng Hu, Yijie Lin, Xi Peng · 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).

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

20 physical pages

Before references / appendix

≈ 9.86 page-equivalents

Main figures / tables

3 / 2

Attached appendix figures / tables

5 / 2

Attached appendix span

≈ 6.89 page-equivalents

Before references / appendix: ≈ 49.3%References / intervening material: ≈ 16.25%Attached appendix: ≈ 34.45%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	≈ 1.61 pp.	8.05%
2 RELATED WORK	3	≈ 0.53 pp.	2.65%
3 METHOD	3	≈ 4.67 pp.	23.35%
4 EXPERIMENTS	8	≈ 1.4 pp.	7%
5 CONCLUSION	9	≈ 0.76 pp.	3.8%
AUTHOR CONTRIBUTIONS	10	≈ 0.13 pp.	0.65%
ACKNOWLEDGMENTS	10	≈ 0.14 pp.	0.7%
REFERENCES	10	≈ 3.25 pp.	16.25%
APPENDIX	14	≈ 0.03 pp.	0.15%
A EFFICIENT SOLVER FOR JOINT DISTRIBUTION ESTIMATION (PROOF OF	14	≈ 0.97 pp.	4.85%
B CONTRASTIVE LEARNING AS A SPECIAL CASE OF CORRGEN (PROOF OF	15	≈ 0.43 pp.	2.15%
C IMPLEMENTATION DETAILS	15	≈ 0.82 pp.	4.1%
D PERFORMANCE VISUALIZATION WITH VARYING MR AND CR VALUE (Q3)	16	≈ 0.3 pp.	1.5%
E PARAMETERS ANALYSIS (Q4)	16	≈ 1.08 pp.	5.4%
F ABLATION STUDIES (Q5)	17	≈ 0.78 pp.	3.9%
G CONVERGENCE ANALYSIS	18	≈ 0.98 pp.	4.9%
H ANALYSIS OF CATEGORY-LEVEL MISMATCH RATIO	19	≈ 1.24 pp.	6.2%
I IMAGE-TEXT PAIR EXAMPLE OF UMPC-FOOD101	20	≈ 0.11 pp.	0.55%
J THE USE OF LARGE LANGUAGE MODELS	20	≈ 0.16 pp.	0.8%

All detected figure and table locations

Main text

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

Attached appendix

• Figure 4 · p. 17
• Figure 5 · p. 17
• Figure 6 · p. 18
• Table 3 · p. 18
• Figure 7 · p. 19
• Table 4 · p. 20
• Figure 8 · p. 20

Argument and organization

1. Introduction and related work

   Distinguish class-level false negatives from alignable and unalignable sample mismatches.

2. Generative formulation and EM procedure

   Translate the failure taxonomy into soft correspondences, GMM-guided transport and a weighted representation objective.

3. Question-led experiments and appendices

   Map five evaluation questions to comparisons, correspondence visualizations, robustness, sensitivity and ablation.

Problem, claim and contribution

Figure 1 gives one concrete image–text example before contrasting reweighting, realignment, and many-to-many correspondence generation. The two mismatch definitions then formalize exactly what those illustrations are meant to solve.

Figure 1, p. 2; Definitions 1–2, p. 4. · Main paper

Argument and section order

The method first states a latent-correspondence likelihood, derives a lower-bound perspective, and separates estimation from representation updates. The experiment introduction explicitly maps five questions to main-text or appendix destinations.

Sections 3.1–3.2, pp. 3–8; Section 4, p. 8. · Main paper

Figures, captions and visual explanation

Figure 2 carries colored clusters into marginal-density sketches and the correspondence matrix, with the virtual sample visibly attached to the E-step. Its two-panel hierarchy makes the estimator and learning update easy to distinguish.

Figure 2, p. 6. · Main paper

Tables, comparisons and evidence

Table 1 groups mismatch ratios vertically and datasets horizontally, labels three metrics per dataset, and defines bold/underline plus five-seed means. Its breadth is informative, but the very dense type and absent variability estimates limit rapid uncertainty-aware reading.

Table 1, p. 9. · Main paper

Appendix and reproducibility

Appendix C reveals the retained DIVIDE architecture, identity warm start, within-view identity mixing, fixed virtual mass, pretrained feature sources, and batch realignment. These details prevent interpreting the result as training an entirely new image–text foundation model.

Appendix C, pp. 15–16; Appendix F, pp. 17–18. · Main paper

What to reuse—and what not to infer

Transferable lesson: Build a failure taxonomy first, give each proposed component a visible place in the mechanism diagram, and assign each experimental question an explicit evidence destination.

Boundary: The exact-EM convergence argument assumes a tight posterior and improving M-step; the practical GMM/transport estimate needs separate justification. Appendix A calls an entropy-regularized maximization objective convex rather than concave, and some Scene15 ablation values differ from Table 2. Preserve feature-based clustering protocols and do not generalize these results to raw-image foundation-model training.

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

Sections 1–5, pp. 1–10; references pp. 10–13.

Appendix / supplement

Appendices A–J, pp. 14–20: transport derivation, InfoNCE special case, implementation, robustness/sensitivity/ablation, convergence, mismatch metric, noisy examples and LLM-use statement.

Visually inspected PDF pages

6, 9

• ICLR 2026 official proceedings PDF, retrieved 2026-09-19 · 20 pages