Writing study: Partition Generative Modeling: Masked Modeling Without Masks

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Justin Deschenaux, Lan Tran, Caglar Gulcehre · 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) · Inference, algorithms & systems efficiency.

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

26 physical pages

Before references / appendix

≈ 9.23 page-equivalents

Main figures / tables

4 / 2

Attached appendix figures / tables

2 / 8

Attached appendix span

≈ 12.65 page-equivalents

Before references / appendix: ≈ 35.5%References / intervening material: ≈ 15.85%Attached appendix: ≈ 48.65%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	≈ 0.84 pp.	3.23%
2 BACKGROUND	2	≈ 1.83 pp.	7.04%
3 PARTITION GENERATIVE MODELING	4	≈ 1.26 pp.	4.85%
4 THE PARTITION TRANSFORMER	5	≈ 1.13 pp.	4.35%
5 EXPERIMENTS	6	≈ 1.97 pp.	7.58%
6 RELATED WORK	8	≈ 0.82 pp.	3.15%
7 CONCLUSION	9	≈ 0.45 pp.	1.73%
8 ACKNOWLEDGEMENTS	10	≈ 0.12 pp.	0.46%
REFERENCES	10	≈ 4.12 pp.	15.85%
A LIMITATIONS	14	≈ 0.18 pp.	0.69%
B SAMPLING WITH HALTON SEQUENCES (MASKGIT)	14	≈ 1 pp.	3.85%
C EXPERIMENTAL DETAILS	15	≈ 2.1 pp.	8.08%
D ADDITIONAL RESULTS	17	≈ 1.15 pp.	4.42%
E COMPUTATIONAL COSTS	18	≈ 8.22 pp.	31.62%

All detected figure and table locations

Main text

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

Attached appendix

• Figure 5 · p. 15
• Table 3 · p. 16
• Figure 6 · p. 17
• Table 4 · p. 19
• Table 5 · p. 21
• Table 6 · p. 22
• Table 7 · p. 23
• Table 8 · p. 24
• Table 9 · p. 25
• Table 10 · p. 26

Argument and organization

1. Masking versus partitioning

   Frame inference waste through a familiar baseline before introducing an architectural alternative.

2. Training, sampling, and Partition Transformer

   Explain the objective and sampling contract before detailing information-flow safeguards.

3. Language, image, and complementary-mask experiments

   Separate cross-domain usefulness from architecture and supervision ablations.

4. Limitations, protocols, and computational costs

   Document precision effects, sample metrics, samplers, and the training/inference trade-off.

Problem, claim and contribution

The core argument identifies a specific wasted operation—processing mask tokens at every step—rather than presenting speed as an unexplained headline. Prior accelerators are distinguished by which flexibility or quality they sacrifice.

Sections 1–2, pages 1–4. · Main paper

Argument and section order

Separating the probabilistic objective, generation procedure, and attention architecture helps readers verify that target tokens cannot leak into their own predictions. Distillation is introduced as compatibility with an existing method, not a newly invented algorithm.

Sections 3–4, pages 4–6; Section 5.1, page 7. · Main paper

Figures, captions and visual explanation

Figure 2 contrasts baseline training/sampling with the proposed two modes using consistent token colors and aligned callouts. Figure 3 then expands each named module and explicitly captions the omitted RoPE operations and absent decoder self-attention.

Figures 2–3, pages 3 and 5. · Main paper

Tables, comparisons and evidence

Results pair quality with throughput and parameter counts. The image tables expose guidance sweeps, sampler choice, step count, and different batch sizes for latency versus throughput rather than reporting one speed number without context.

Table 1, page 6; Tables 8–9, pages 24–25. · Main paper

Appendix and reproducibility

The appendix records FP64 sampling, entropy checks against repetitive text, loss-precision differences, and slower training. These details are essential for interpreting the efficiency claim and avoiding deceptively favorable generative perplexity.

Appendices A, C.4–C.5, D.4, and E, pages 14–19. · Main paper

Clarity and writing conventions

Explicit takeaways make architecture ablations easy to scan, but the takeaway must be checked against its own rows: the claimed 1.95 LM1B gain belongs to complementary-masked MDLM, not the listed PGM.

Section 5.1, page 7; Table 5, page 21. · Main paper

What to reuse—and what not to infer

Transferable lesson: Show the baseline's wasted computation, the exact information-flow change that removes it, and the resulting quality–speed trade-off; preserve training overhead, sampler settings, and diversity checks.

Boundary: Do not reuse the stated 1.95 PGM perplexity gain: Table 1 gives 27.67−26.80=0.87; 1.95 is 27.67−25.72 for complementary-masked MDLM. Several latency columns labeled milliseconds conflict with seconds in the main table, so retain throughput rather than converting ambiguous latency. PGM is not demonstrated as a multimodal model, and long-context quality beyond 1024 is untested.

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 1–8; reference list excluded from analytical reading.

Appendix / supplement

Pages 14–26, Appendices A–E, all algorithms and extended tables.

Visually inspected PDF pages

3, 5, 24, 25

• Official ICLR 2026 proceedings conference PDF; PDF creation metadata: Tue Feb 17 09:00:49 2026 EST · 26 pages