Writing study: Learning Unmasking Policies for Diffusion Language Models

ICML overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICML 2026 oral writing corpus

Metod Jazbec, Theo X. Olausson, Louis Béthune, Pierre Ablin, Michael Kirchhof, Joao Monteiro, Victor Guilherme Turrisi Da Costa, Jason Ramapuram, Marco Cuturi · Official ICML 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: Training, objectives & optimization (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

35 physical pages

Before references / appendix

≈ 9.1 page-equivalents

Main figures / tables

10 / 0

Attached appendix figures / tables

24 / 5

Attached appendix span

≈ 21.91 page-equivalents

Before references / appendix: ≈ 26%References / intervening material: ≈ 11.4%Attached appendix: ≈ 62.6%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	≈ 0.95 pp.	2.71%
2. Background	2	≈ 1.13 pp.	3.23%
3. Learning Unmasking Policies	3	≈ 2.33 pp.	6.66%
4. Experiments	5	≈ 3.99 pp.	11.4%
5. Conclusion	9	≈ 0.3 pp.	0.86%
Impact Statement	10	≈ 0.05 pp.	0.14%
References	10	≈ 3.99 pp.	11.4%
Appendix	14	≈ 1 pp.	2.86%
A. Related work	15	≈ 0.96 pp.	2.74%
B. Policy Sampling Algorithm	16	≈ 1.04 pp.	2.97%
C. Additional Results	17	≈ 5 pp.	14.29%
D. What are RL unmasking policies actually doing?	22	≈ 3 pp.	8.57%
E. Generation Trajectories Visualizations	25	≈ 4 pp.	11.43%
F. Dynamic Plackett-Luce Sampling	29	≈ 1 pp.	2.86%
G. Expert Steering	30	≈ 1 pp.	2.86%
H. Extended Background	31	≈ 0.96 pp.	2.74%
I. Training and Policy Network Configuration	32	≈ 1 pp.	2.86%
J. Policy Architecture Diagram	33	≈ 1.04 pp.	2.97%
K. Tabular Reference for Main LLaDA Experiments	34	≈ 1.91 pp.	5.46%

All detected figure and table locations

Main text

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

Attached appendix

• Figure 10 · p. 17
• Figure 11 · p. 17
• Figure 12 · p. 18
• Figure 13 · p. 18
• Figure 14 · p. 18
• Figure 15 · p. 19
• Figure 16 · p. 19
• Figure 17 · p. 20
• Figure 18 · p. 20
• Figure 19 · p. 21
• Figure 20 · p. 21
• Figure 21 · p. 22
• Figure 22 · p. 22
• Figure 23 · p. 23
• Figure 24 · p. 23
• Figure 25 · p. 23
• Figure 26 · p. 24
• Figure 27 · p. 25
• Figure 28 · p. 25
• Figure 29 · p. 26
• Figure 30 · p. 26
• Figure 31 · p. 27
• Figure 32 · p. 27
• Figure 33 · p. 28
• Table 1 · p. 32
• Table 2 · p. 34
• Table 3 · p. 34
• Table 4 · p. 35
• Table 5 · p. 35

Argument and organization

1. Problem and contribution

   Explicitly separates improving the base model's capacity from learning how to sample it; the frozen dLLM is the environment, not the RL policy.

2. Argument sequence

   Heuristic failures motivate an MDP with separately defined state, action, transition and reward, followed by policy parameterization and GRPO training.

3. Evidence and verification

   Transfer results retain the weaker math-to-code outcome, and reward ablations explain how an additive speed incentive can favor fast incorrect answers.

Problem, claim and contribution

Explicitly separates improving the base model's capacity from learning how to sample it; the frozen dLLM is the environment, not the RL policy.

Contribution statement, p. 2 · Main paper

Argument and section order

Heuristic failures motivate an MDP with separately defined state, action, transition and reward, followed by policy parameterization and GRPO training.

Sections 2.2-3.3, pp. 2-5 · Main paper

Figures, captions and visual explanation

Solid/dashed curves distinguish semi-autoregressive from full-diffusion sampling. The transfer figure separates an extreme operating point rather than drawing a misleading connecting line.

Figure 1, p. 3; Figure 8, p. 8; both PNGs inspected · Main paper

Tables, comparisons and evidence

Transfer results retain the weaker math-to-code outcome, and reward ablations explain how an additive speed incentive can favor fast incorrect answers.

Sections 4.3-4.4, pp. 7-9 · Main paper

Appendix and reproducibility

Expert steering is explicitly training-only; additional figures convert NFE comparisons to wall-clock time, and final tables retain numerical operating points.

Appendix C.3, p. 18; G, p. 30; K, pp. 34-35 · Main paper

Clarity and writing conventions

Captions disclose omitted near-zero regimes and omitted baselines with cross-references instead of silently presenting the simplified plots as complete.

Figure 4 caption, p. 5; Figure 8 caption, p. 8 · Main paper

What to reuse—and what not to infer

Transferable lesson: Let captions explain simplifications and anomalous operating points; pair a speed proxy with a separate wall-clock check.

Boundary: NFE is not elapsed time, and transfer across one model or length does not establish domain-independent sampling policies.

Apply this lesson using the argument and evidence worksheet.

Inspection record and sources

Checked

2026-10-07

Reading status

Writing-focused manuscript scan + selected visuals inspected

Main text

Writing-focused distributed manuscript scan, not exhaustive technical/proof validation. Sections 1-5, pp. 1-9: sampling motivation, MDP formulation, policy/reward design, performance, transfer and failure diagnostics.

Appendix / supplement

Writing-focused appendix scan; code, data and external assets not executed. Appendices A-K, pp. 14-35: related work, sampling algorithm, extended results, behavior/trajectory views, expert steering, network configuration and numerical tables.

Visually inspected PDF pages

3, 8

This is a writing-focused scan of the manuscript and the stated supplementary scope: framing, organization, visual explanation, evidence presentation and wording. It is not a line-by-line proof audit, reproduction of results or exhaustive scientific review.

• Published ICML 2026 proceedings PDF, PMLR volume 306; retrieved 2026-10-07. Public OpenReview decision notes were not independently reverified. · 35 pages