Writing study: Learning with Dual-level Noisy Correspondence for Multi-modal Entity Alignment

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Haobin Li, Yijie Lin, Peng Hu, Mouxing Yang, 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) · Architecture & representation design.

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

32 physical pages

Before references / appendix

≈ 9.54 page-equivalents

Main figures / tables

5 / 3

Attached appendix figures / tables

10 / 15

Attached appendix span

≈ 19.9 page-equivalents

Before references / appendix: ≈ 29.81%References / intervening material: ≈ 8%Attached appendix: ≈ 62.19%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	≈ 1.53 pp.	4.78%
2 METHOD	3	≈ 4.49 pp.	14.03%
3 EXPERIMENTS	7	≈ 2.52 pp.	7.88%
4 CONCLUSION	10	≈ 0.16 pp.	0.5%
AUTHOR CONTRIBUTIONS	10	≈ 0.11 pp.	0.34%
ACKNOWLEDGMENTS	10	≈ 0.11 pp.	0.34%
REFERENCES	10	≈ 2.57 pp.	8.03%
A PPENDIX	13	≈ 1.11 pp.	3.47%
B NOISE STATISTICS IN REAL-WORLD BENCHMARKS	14	≈ 1.89 pp.	5.91%
C RELATED WORK	16	≈ 1 pp.	3.13%
D EVIDENTIAL LEARNING	17	≈ 1 pp.	3.13%
E PROOF OF THEOREM 1	18	≈ 0.35 pp.	1.09%
F MORE IMPLEMENTATION DETAILS	18	≈ 3.65 pp.	11.41%
G MORE EXPERIMENT RESULTS	22	≈ 8 pp.	25%
H CASE STUDY OF DUAL-LEVEL NOISY CORRESPONDENCE	30	≈ 0.31 pp.	0.97%
I CASE STUDY OF TEST-TIME CORRESPONDENCE REASONING	30	≈ 2.59 pp.	8.09%

All detected figure and table locations

Main text

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

Attached appendix

• Figure 6 · p. 14
• Table 4 · p. 18
• Table 5 · p. 22
• Table 6 · p. 22
• Figure 7 · p. 23
• Table 7 · p. 23
• Table 8 · p. 23
• Table 9 · p. 24
• Figure 8 · p. 24
• Table 10 · p. 24
• Figure 9 · p. 25
• Table 11 · p. 25
• Table 12 · p. 25
• Table 13 · p. 26
• Table 14 · p. 26
• Table 15 · p. 27
• Table 16 · p. 27
• Table 17 · p. 28
• Table 18 · p. 28
• Figure 10 · p. 30
• Figure 11 · p. 31
• Figure 12 · p. 31
• Figure 13 · p. 32
• Figure 14 · p. 32
• Figure 15 · p. 32

Argument and organization

1. A two-level corruption taxonomy

   Separates entity–attribute corruption from cross-graph matching errors before defining uncertainty and consensus.

2. Training robustness and inference reasoning

   Connects reliability to fusion/loss design, then adds a separately evaluated MLLM reranking stage.

Problem, claim and contribution

Concrete mismatches explain why noisy correspondence differs from merely missing modalities; a manual sample audit supports the practical motivation.

Figure 1, pp. 1–2; Appendix B, pp. 14–15. · Main paper

Argument and section order

Definitions of uncertainty, consensus and pair subsets create a shared vocabulary reused by both fusion and robust learning.

Sections 2.2–2.4, pp. 3–6. · Main paper

Figures, captions and visual explanation

The pipeline allocates separate panels to reliability estimation, training modules and test-time reasoning, revealing which components affect training versus deployment.

Figure 2, p. 4. · Main paper

Tables, comparisons and evidence

The ablation table explicitly labels train and test stages, and adjacent reliability plots connect numerical gains to the proposed noise-handling mechanism.

Table 3 and Figures 4–5, p. 9. · Main paper

Appendix and reproducibility

Exact MLLM prompts, smaller-model comparisons, runtime/memory results and unsuccessful reranking cases give deployment context beyond accuracy.

Appendices F.5, G.7–G.8 and I, pp. 19–21, 25–26, 30–32. · Main paper

Clarity and writing conventions

The repeated word 'guarantee' overstates empirical reranking improvements; the appendix itself shows missing-knowledge and shortcut failures.

Sections 2.5 and 4; Appendix I. · Main paper

What to reuse—and what not to infer

Transferable lesson: When proposing robustness at multiple stages, name the corruption types explicitly, label where each component acts, and pair ablations with failure cases and deployment costs.

Boundary: Default inference adds a 72B VLM beyond the shared CLIP backbone, so headline comparisons are not equal-compute. Synthetic corruption is only one noise model; semantic relatedness is not necessarily entity identity. Some supplementary metric rows appear inconsistent and require checking before reuse.

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 1–4 and all main figures/tables.

Appendix / supplement

Pages 14–32, Appendices A–I, including proofs, prompt templates, expanded tables and image-only failure/missing-modality cases.

Visually inspected PDF pages

4, 9, 30, 31, 32

• ICLR 2026 official proceedings PDF, retrieved 2026-09-19 · 32 pages