Writing study: Modality-free Graph In-context Alignment

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Wei Zhuo, Siqiang Luo · 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

27 physical pages

Before references / appendix

≈ 10.21 page-equivalents

Main figures / tables

4 / 5

Attached appendix figures / tables

1 / 8

Attached appendix span

≈ 11.89 page-equivalents

Before references / appendix: ≈ 37.81%References / intervening material: ≈ 18.15%Attached appendix: ≈ 44.04%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	≈ 0.93 pp.	3.44%
2 PRELIMINARIES	2	≈ 2.12 pp.	7.85%
3 MF-GIA: MODALITY-FREE GRAPH IN-CONTEXT ALIGNMENT	4	≈ 4.21 pp.	15.59%
4 EXPERIMENTS	8	≈ 1.86 pp.	6.89%
5 CONCLUSION	10	≈ 0.44 pp.	1.63%
ACKNOWLEDGEMENTS	11	≈ 0.11 pp.	0.41%
REFERENCES	11	≈ 4.89 pp.	18.11%
A RELATED WORK	16	≈ 1.05 pp.	3.89%
B PROOFS	17	≈ 4.18 pp.	15.48%
C ALGORITHMS AND COMPLEXITY ANALYSIS	21	≈ 0.43 pp.	1.59%
D TASK UNIFICATION	21	≈ 1.95 pp.	7.22%
E EXPERIMENTAL DETAILS	23	≈ 1.62 pp.	6%
F MORE EXPERIMENTS	25	≈ 2.39 pp.	8.85%
G THE USE OF LARGE LANGUAGE MODELS (LLMS)	27	≈ 0.28 pp.	1.04%

All detected figure and table locations

Main text

• Table 1 · p. 1
• Figure 1 · p. 3
• Figure 2 · p. 4
• Figure 3 · p. 6
• Table 2 · p. 9
• Table 3 · p. 9
• Table 4 · p. 10
• Figure 4 · p. 10
• Table 5 · p. 10

Attached appendix

• Table 6 · p. 23
• Figure 5 · p. 24
• Table 7 · p. 25
• Table 8 · p. 25
• Table 9 · p. 26
• Table 10 · p. 26
• Table 11 · p. 26
• Table 12 · p. 27
• Table 13 · p. 27

Argument and organization

1. Three criteria for graph ICL

   Defines post-training freedom, cross-domain alignment and modality freedom.

2. Fingerprints, alignment and episodic matching

   Decomposes the method into reusable components.

3. Node transfer, edge transfer and ablations

   Separates domain generalization from task-formulation transfer.

Problem, claim and contribution

A compact criteria table establishes which combination is claimed new before introducing the method.

Table 1 and Section 1, pp. 1–2 · Main paper

Argument and section order

Definitions precede the domain embedder, and feature-side and label-side FiLM transformations are presented in parallel subsections.

Sections 2–3.2, pp. 2–7 · Main paper

Figures, captions and visual explanation

The three-stage overview separates alignment, pretraining and frozen inference; its trainable/frozen legend is an effective device for adaptation papers.

Figure 1, p. 3 · Main paper

Tables, comparisons and evidence

The component table adds feature alignment, label alignment and episodic attention step by step, making the contribution chain easier to inspect than only a final leaderboard.

Table 4, p. 10 · Main paper

Appendix and reproducibility

Separate pretraining and inference algorithms help audit the frozen-parameter claim, but the text's extra label propagation and adaptive metric are absent from the inference pseudocode.

Section 3.4, p. 8; Algorithms 1–2, pp. 22–23 · Main paper

Clarity and writing conventions

The claimed win across every node configuration has an exception in Computers 1-shot; the Cora dataset row also repeats WikiCS statistics despite a seven-way Cora experiment.

Table 2 and Section 4.2, p. 9; Table 6, p. 23 · Main paper

What to reuse—and what not to infer

Transferable lesson: Use explicit criteria and train/freeze diagrams to define an adaptation method, then make every inference-time operation appear in the executable pseudocode.

Boundary: An upper Lipschitz bound does not prove dissimilar domains remain separated. The all-ones class projection with cross-entropy would yield identical logits and needs implementation clarification before reusing the fingerprint claim. Parameter-update-free inference still computes a support gradient; no pathology validation is supplied.

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, Sections 1–5

Appendix / supplement

PDF pp. 16–27, Appendices A–G including proofs and both algorithms

Visually inspected PDF pages

3, 10

• Official ICLR 2026 proceedings conference PDF; PDF creation metadata: Mon Feb 16 13:42:42 2026 EST · 27 pages