Writing study: On The Surprising Effectiveness of a Single Global Merging in Decentralized Learning

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Tongtian Zhu, Tianyu Zhang, Mingze Wang, Zhanpeng Zhou, Can Wang · 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) · Theory & formal foundations · Learning & optimization algorithms.

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

36 physical pages

Before references / appendix

≈ 10.11 page-equivalents

Main figures / tables

2 / 1

Attached appendix figures / tables

6 / 0

Attached appendix span

≈ 18.29 page-equivalents

Before references / appendix: ≈ 28.08%References / intervening material: ≈ 21.11%Attached appendix: ≈ 50.81%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	≈ 1.72 pp.	4.78%
2 RELATED WORK	3	≈ 0.32 pp.	0.89%
3 NOTATIONS AND PRELIMINARIES	3	≈ 1.15 pp.	3.19%
4 EMPIRICAL OBSERVATIONS	4	≈ 2.31 pp.	6.42%
5 THEORETICAL ANALYSIS	7	≈ 3.64 pp.	10.11%
6 IMPLICATIONS AND DISCUSSIONS	10	≈ 0.36 pp.	1%
REFERENCES	11	≈ 7 pp.	19.44%
LLM USAGE STATEMENT	18	≈ 0.35 pp.	0.97%
ETHICS STATEMENT	18	≈ 0.14 pp.	0.39%
REPRODUCIBILITY STATEMENT	18	≈ 0.12 pp.	0.33%
A LIMITATIONS AND POTENTIAL QUESTIONS	18	≈ 1.39 pp.	3.86%
B ADDITIONAL BACKGROUND AND RELATED WORK	20	≈ 2.05 pp.	5.69%
C ADDITIONAL EXPERIMENTS	22	≈ 4.95 pp.	13.75%
D THEORY	27	≈ 9.89 pp.	27.47%

All detected figure and table locations

Main text

• Figure 1 · p. 2
• Figure 2 · p. 5
• Table 1 · p. 8

Attached appendix

• Figure C.1 · p. 24
• Figure C.2 · p. 25
• Figure C.3 · p. 25
• Figure C.4 · p. 26
• Figure C.5 · p. 26
• Figure C.6 · p. 26

Argument and organization

1. Empirical observations

   Build the finding from late synchronization windows to one final merge and a no-communication control.

2. Theoretical analysis

   State assumptions before connecting consensus error and progressive sharpening to a conditional convergence claim.

3. Limitations and potential questions

   Answer optimizer, initialization, tuning and visual-task-scope objections explicitly.

Problem, claim and contribution

The opening asks when to communicate, contrasting temporal allocation with familiar graph design. Separate boxed question and phenomenon statements make the empirical discovery legible before formal machinery appears.

§1, pp.1–3 · Main paper

Argument and section order

The experiments narrow synchronization from broad windows to one final round, then inspect hypothetical averages throughout training. This progression distinguishes a final intervention from latent mergeability and includes the essential zero-communication control.

§§4.1–4.3, pp.4–6 · Main paper

Figures, captions and visual explanation

Figure 1 combines learning curves, network cartoons and an interpolation landscape. Its ring is a chosen hexadecagon coordinate layout, not a measured spatial arrangement of parameters; the appendix explains the barycentric construction needed to interpret it.

Fig.1, p.2; Appendix C.1, p.23 · Main paper

Tables, comparisons and evidence

The paper defines average global test accuracy separately from personalized local accuracy. The appendix varies optimizers, initializations, topology and heterogeneity, while final-gossip panels show that approximating global synchronization is possible but not identical to no communication.

Definition 1, p.4; Appendix C.2–C.3, pp.23–26 · Main paper

Appendix and reproducibility

The question-and-answer appendix discloses global-test-based hyperparameter tuning and that experiments are visual classification only. The proof appendix also states that arbitrarily small target error may require increasing the number of agents, a consequential qualification to matching rates.

Appendix A, p.19; Remark D.2, p.35 · Main paper

Clarity and writing conventions

Empirical discovery, theoretical explanation and practical scheduling advice are separated. However, an optimization-stationarity result should not be narrated as a direct generalization theorem or a guarantee for every independently initialized implementation.

Theorem 1, p.7; §5.3, p.10; Appendix A, pp.18–19 · Main paper

What to reuse—and what not to infer

Transferable lesson: Turn a surprising curve into a sequence of controlled questions, define the evaluated population, and explain how every diagnostic visualization is constructed.

Boundary: One final merge follows persistent sparse gossip, not isolated local training. The theorem uses shared initialization although several experiments do not. Its progressive-sharpening assumption is unusually strong: the literal strictly negative claim for every positive-semidefinite matrix includes the zero-matrix edge case and needs qualification. Test-based tuning and absent uncertainty reporting limit generalization claims; reading the proof is not independent proof certification.

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, §§1–6; references pp.11–17

Appendix / supplement

Statements and Appendices A–D, PDF pp.18–36, including all convergence proofs

Visually inspected PDF pages

2, 26

• official ICLR 2026 proceedings PDF · 36 pages