Writing study: Actions Speak Louder than Prompts: A Large-Scale Study of LLMs for Graph Inference

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Ben Finkelshtein, Silviu Cucerzan, Sujay Kumar Jauhar, Ryen White · 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: Empirical analysis, diagnosis & auditing (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

23 physical pages

Before references / appendix

≈ 9.42 page-equivalents

Main figures / tables

4 / 3

Attached appendix figures / tables

0 / 9

Attached appendix span

≈ 9.89 page-equivalents

Before references / appendix: ≈ 40.96%References / intervening material: ≈ 16.04%Attached appendix: ≈ 43%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	≈ 1.05 pp.	4.57%
2 RELATED WORK	2	≈ 0.92 pp.	4%
3 AXES OF VARIABILITY	3	≈ 1.83 pp.	7.96%
4 EXPERIMENTS ACROSS AXES OF VARIABILITY	5	≈ 2.25 pp.	9.78%
5 EXPERIMENTS ON FEATURES, STRUCTURE, AND LABELS DEPENDENCIES	7	≈ 1.97 pp.	8.57%
6 CONCLUSIONS	9	≈ 0.47 pp.	2.04%
ETHICS STATEMENT	10	≈ 0.19 pp.	0.83%
REPRODUCIBILITY STATEMENT	10	≈ 0.12 pp.	0.52%
REFERENCES	10	≈ 3.69 pp.	16.04%
A ADDITIONAL RELATED WORKS	14	≈ 0.48 pp.	2.09%
B ADDITIONAL AXES OF VARIABILITY	14	≈ 1.72 pp.	7.48%
C TOKEN AND LATENCY EFFICIENCY	16	≈ 0.99 pp.	4.3%
D ADDITIONAL EXPERIMENTS	17	≈ 3.47 pp.	15.09%
E AN EXAMPLE OF GRAPH-AS-CODE REASONING	20	≈ 1 pp.	4.35%
F DETAILS ON LLM-GRAPH INTERACTION MODES	21	≈ 1.89 pp.	8.22%
G DATASET STATISTICS	23	≈ 0.33 pp.	1.43%

All detected figure and table locations

Main text

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

Attached appendix

• Table 4 · p. 15
• Table 5 · p. 15
• Table 6 · p. 16
• Table 7 · p. 16
• Table 8 · p. 18
• Table 9 · p. 18
• Table 10 · p. 19
• Table 11 · p. 20
• Table 12 · p. 23

Argument and organization

1. Axes of variability

   Define interaction modes and graph regimes before reporting named findings.

2. Experiments and dependencies

   Separate ordinary accuracy comparisons from controlled information deletion.

3. Additional experiments and templates

   Test summarization and shuffled adjacency, contextualize supervised baselines, and disclose executable interfaces.

Problem, claim and contribution

The central question concerns orchestration of frozen LLMs rather than a new graph encoder. The six variation axes help readers distinguish a systems comparison from model-training advances.

Introduction and Section 3. · Main paper

Argument and section order

Numbered findings make a long empirical paper navigable, but universal wording exceeds individual rows: Texas label propagation exceeds the LLM variants, and Graph-as-Code is not best on every dataset.

Findings 2–4; Tables 1–3. · Main paper

Figures, captions and visual explanation

Figure 1 uses matched three-column layouts to contrast one-shot prompting, fixed actions and executable code. Identical graph and agent icons make the changed interface immediately visible.

Figure 1, page 3. · Main paper

Tables, comparisons and evidence

Paired heatmaps share color scales and deletion axes, exposing information dependencies that one aggregate accuracy number would hide. Truncating text is distinct from uniformly deleting edges or labels.

Figures 2–4; Section 5. · Main paper

Appendix and reproducibility

Templates and an execution transcript show precisely which labels the environment exposes; supervised hybrid results are explicitly taken from another study rather than presented as controlled reruns.

Appendices D.4–G, pages 20–23. · Main paper

Clarity and writing conventions

Several numerical checks need clarification: Table 10's overall prompting MSE exceeds every disjoint path-length bin; the 1,000-test-node sampling description also needs a small-dataset exception or replacement rule.

Table 10, page 19; Section 5 and Table 12. · Main paper

What to reuse—and what not to infer

Transferable lesson: Use matched interface diagrams and controlled perturbation grids to turn a broad benchmark into actionable conditional guidance.

Boundary: This is not proof of graph reasoning independent of tool access. Code can inspect non-neighbor features when prompts cannot; closed-model size comparisons and unqualified no-risk language need qualification.

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 1–6, ethics and reproducibility; references 10–13.

Appendix / supplement

Pages 14–23: Appendices A–G, including model comparisons, latency, controls, execution example, templates and dataset splits.

Visually inspected PDF pages

3, 7

• ICLR 2026 final proceedings · 23 pages