Writing study: Structured Flow Autoencoders: Learning Structured Probabilistic Representations with Flow Matching

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Yidan Xu, Yixin Wang, XuanLong Nguyen · 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: Architecture & representation design (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

22 physical pages

Before references / appendix

≈ 10.11 page-equivalents

Main figures / tables

4 / 4

Attached appendix figures / tables

6 / 0

Attached appendix span

≈ 8.89 page-equivalents

Before references / appendix: ≈ 45.95%References / intervening material: ≈ 13.64%Attached appendix: ≈ 40.41%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	≈ 2.2 pp.	10%
2 PRELIMINARIES: FLOW-BASED GENERATIVE MODELING	3	≈ 0.94 pp.	4.27%
3 STRUCTURED FLOW AUTOENCODERS	4	≈ 3.16 pp.	14.36%
4 EMPIRICAL STUDIES	7	≈ 2.58 pp.	11.73%
5 DISCUSSION	10	≈ 0.37 pp.	1.68%
ACKNOWLEDGMENTS	10	≈ 0.29 pp.	1.32%
REFERENCES	11	≈ 3 pp.	13.64%
A THEORETICAL DETAILS	14	≈ 4 pp.	18.18%
B EXPERIMENT DETAILS	18	≈ 4.36 pp.	19.82%
C THE USE OF LARGE LANGUAGE MODELS (LLMS)	22	≈ 0.53 pp.	2.41%

All detected figure and table locations

Main text

• Table 1 · p. 2
• Figure 1 · p. 2
• Figure 2 · p. 3
• Figure 3 · p. 5
• Table 2 · p. 8
• Table 3 · p. 9
• Figure 4 · p. 9
• Table 4 · p. 10

Attached appendix

• Figure 5 · p. 18
• Figure 6 · p. 19
• Figure 7 · p. 20
• Figure 8 · p. 21
• Figure 9 · p. 21
• Figure 10 · p. 22

Argument and organization

1. A two-objective modeling gap

   Contrasts generative fidelity with interpretable structured representations; the pinwheel example makes the conflict visible before formal derivation.

2. One probabilistic identity, three model families

   Introduces CNFs and flow matching, derives the posterior-averaged conditional vector field, and instantiates continuous, mixture and dynamical latents.

3. Evaluate observations and latents separately

   Defines posterior-predictive sampling and representation metrics before reporting toy, image, gene-expression and sequence evidence.

Problem, claim and contribution

The introduction frames an explicit tension: a strong marginal density model need not expose useful structured latents, while a latent-variable model need not generate realistic observations. A question links that tension to the proposed family, and numbered contributions distinguish the family, training objective and empirical instantiations. This is more informative than presenting three variants as unrelated inventions.

Abstract; Section 1; Figure 1 and Table 1 · Main paper

Argument and section order

The notation paragraph separates marginal and conditional vector fields before the main theorem. Section 3 then moves from an identity to an objective and three graphical models. Figure 2 serves as an outline of those instances. This is a useful organization for a unifying framework: teach the shared contract once, then identify exactly what changes in each case.

Sections 2–3.3; Figure 2; Theorem 3.1 · Main paper

Figures, captions and visual explanation

Figure 3 pairs a conceptual co-evolution diagram with PCA-projected observed/latent trajectories. Matching variables and time endpoints connect the picture to Equation 6. The caption explains which panel concerns conditional generation versus posterior representation; however, the plotted projection should remain illustrative rather than being treated as an identifiability proof. The small legend requires careful final-size checking.

Figure 3, PDF page 5 · Main paper

Tables, comparisons and evidence

Table 2 groups likelihood, diversity, reconstruction and clustering instead of hiding trade-offs in one score. The deterministic latent has better in-distribution NMI than stochastic SFA, while the OOD columns favor the stochastic choice; LatentFM has higher SSIM. Preserve this multi-criterion reading rather than repeating an unqualified claim of universal superiority. The prose explicitly defines posterior-predictive sampling, which is not the same evaluation as unconstrained prior generation.

Section 4; Table 2, PDF page 8; Section 4.2 · Main paper

Appendix and reproducibility

Appendix A states regularity conditions, expands the continuity-equation argument and discusses posterior collapse. Appendix B separates dataset-specific architectures and the full-sequence conditioning used by the dynamical posterior; Appendix C discloses LLM assistance. The appendix is organized by claim and experiment, but descriptions such as train-until-convergence are not a complete optimization recipe.

Appendices A.1–A.2, B.1–B.4 and C · Main paper

Clarity and writing conventions

The discussion usefully narrows the work to a methodological construction and explicitly leaves complex-image scaling and identifiable latents as future work. This restraint matters because the image experiments are MNIST, and the scientific data are RNA-seq, not whole-slide pathology. The reported expensive CNF-posterior runtime qualifies the broader efficiency language.

Section 4 runtime paragraph; Section 5 limitations; Appendix B · Main paper

What to reuse—and what not to infer

Transferable lesson: For a general framework, repeat a consistent chain: graphical-model assumption → mathematical identity → trainable approximation → separate tests of sample quality and representation utility. Put trade-offs and practical approximations next to the corresponding claim.

Boundary: The theorem concerns probability-path consistency under assumptions; it does not establish identifiable learned latents or that a finite neural optimizer reaches them. One-sample inner-expectation approximation and the absence of an explicit KL penalty should not be promoted into an unconditional no-collapse theorem. RNA-seq experiments are not WSI evidence.

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

All main Sections 1–5, PDF pages 1–10: motivation, flow preliminaries, conditional/marginal identity, SCFM and its three graphical-model instances, experiments and limitations.

Appendix / supplement

All Appendices A–C, PDF pages 14–22: regularity assumptions and continuity-equation proofs, latent-dynamics extension, all four experiment configurations/qualitative comparisons and LLM-use disclosure.

Visually inspected PDF pages

5, 8

• Official ICLR 2026 proceedings manuscript including appendices · 22 pages