Writing study: Navigating the Latent Space Dynamics of Neural Models

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Marco Fumero, Luca Moschella, Emanuele Rodolà, Francesco Locatello · 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

29 physical pages

Before references / appendix

≈ 10.67 page-equivalents

Main figures / tables

5 / 0

Attached appendix figures / tables

13 / 4

Attached appendix span

≈ 12.89 page-equivalents

Before references / appendix: ≈ 36.79%References / intervening material: ≈ 18.76%Attached appendix: ≈ 44.45%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	≈ 0.91 pp.	3.14%
2 METHOD	2	≈ 1.66 pp.	5.72%
3 THEORETICAL REMARKS	4	≈ 3.49 pp.	12.03%
4 EXPERIMENTS ON VISION FOUNDATION MODELS	7	≈ 1.67 pp.	5.76%
5 RELATED WORK	9	≈ 0.54 pp.	1.86%
6 CONCLUSIONS AND DISCUSSION	9	≈ 1.3 pp.	4.48%
ACKNOWLEDGMENTS	11	≈ 0.18 pp.	0.62%
ETHICS STATEMENT	11	≈ 0.26 pp.	0.9%
REPRODUCIBILITY STATEMENT	11	≈ 0.12 pp.	0.41%
REFERENCES	11	≈ 5.44 pp.	18.76%
A PROOFS AND DERIVATION	17	≈ 4.71 pp.	16.24%
B ADDITIONAL EXPERIMENTS	21	≈ 6.66 pp.	22.97%
C REGULARIZED AUTOENCODER ENFORCE CONTRACTIVENESS	28	≈ 0.97 pp.	3.34%
D ADDITIONAL IMPLEMENTATION DETAILS	29	≈ 0.55 pp.	1.9%

All detected figure and table locations

Main text

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

Attached appendix

• Figure 6 · p. 22
• Figure 7 · p. 22
• Figure 8 · p. 22
• Figure 9 · p. 22
• Table 1 · p. 23
• Figure 10 · p. 23
• Figure 11 · p. 24
• Figure 12 · p. 24
• Figure 13 · p. 25
• Figure 14 · p. 25
• Figure 15 · p. 26
• Figure 16 · p. 26
• Figure 17 · p. 27
• Figure 18 · p. 27
• Table 2 · p. 28
• Table 3 · p. 29
• Table 4 · p. 29

Argument and organization

1. Latent dynamics and theoretical remarks

   Introduces an encoder–decoder iteration, then interprets trajectories and attractors.

2. Controlled models and foundation-model experiments

   Moves from observable two-dimensional dynamics to probing and OOD applications.

3. Limitations and appendix

   Separates exploratory extensions beyond autoencoders from the core setup.

Problem, claim and contribution

Figure 1 makes the otherwise abstract vector field visible before equations; several random initializations communicate that one attractive image is not the sole example.

Figure 1, p. 1 · Main paper

Argument and section order

Experiments repeatedly use Setting, Analysis of results and a boxed Takeaway, giving readers a stable question–evidence–interpretation rhythm.

Sections 3.2–4.2, pp. 5–9 · Main paper

Figures, captions and visual explanation

Figure 5 combines a quantitative table with matched score histograms, linking OOD metrics to the distributional separation that produces them.

Figure 5, p. 8 · Main paper

Tables, comparisons and evidence

The main OOD comparison uses only KNN; the appendix adds Mahalanobis and reconstruction-error controls, which should be signposted more prominently.

Section 4.2 and Appendix B.3, Table 1 · Main paper

Appendix and reproducibility

The proof, diagnostic, regularizer and implementation sections form a useful navigation hierarchy; later SSL and language-model experiments are explicitly preliminary.

Appendices A–D, especially B.8 · Main paper

Clarity and writing conventions

The formal presentation warrants careful qualification: an open convex domain is not necessarily complete, and the squared-norm decomposition in the prototype bound omits a cross term or multiplicative factor.

Theorem A.6, p. 18; Proposition A.8 proof, p. 20 · Main paper

What to reuse—and what not to infer

Transferable lesson: Pair a low-dimensional visual explanation with application-level evidence, and repeat a compact setup–analysis–takeaway structure.

Boundary: Do not use polished theorem formatting as proof of validity. The appendix contains substantive argument gaps; contraction, score alignment and generalization claims need independent mathematical review. Reduced output variance alone does not establish a contraction mapping.

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 1–6.1

Appendix / supplement

Pages 17–29, Appendices A–D: all derivations, supplementary experiments, regularizer taxonomy and implementation

Visually inspected PDF pages

1, 8

• ICLR 2026 final proceedings · 29 pages