Writing study: Latent Speech-Text Transformer

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Yen-Ju Lu, Yashesh Gaur, Wei Zhou, Benjamin Muller, Jesus Villalba, Najim Dehak, Luke Zettlemoyer, Gargi Ghosh, Mike Lewis, Srini Iyer, Duc Le · 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

18 physical pages

Before references / appendix

≈ 10.24 page-equivalents

Main figures / tables

5 / 6

Attached appendix figures / tables

1 / 7

Attached appendix span

≈ 5.25 page-equivalents

Before references / appendix: ≈ 56.89%References / intervening material: ≈ 13.94%Attached appendix: ≈ 29.17%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.	9.56%
2 BACKGROUND	3	≈ 0.94 pp.	5.22%
3 LATENT SPEECH-TEXT TRANSFORMERS	4	≈ 1.4 pp.	7.78%
4 EXPERIMENTAL SETUP	5	≈ 1.83 pp.	10.17%
5 RESULTS	7	≈ 1.88 pp.	10.44%
6 RELATED WORK	9	≈ 1.14 pp.	6.33%
7 LIMITATIONS	10	≈ 0.2 pp.	1.11%
8 CONCLUSION	10	≈ 0.34 pp.	1.89%
ETHICS STATEMENT	11	≈ 0.14 pp.	0.78%
REFERENCES	11	≈ 2.5 pp.	13.89%
A APPENDIX	13	≈ 5.25 pp.	29.17%

All detected figure and table locations

Main text

• Figure 1 · p. 2
• Figure 2 · p. 3
• Figure 3 · p. 4
• Figure 4 · p. 5
• Table 1 · p. 6
• Table 2 · p. 6
• Table 3 · p. 7
• Table 4 · p. 7
• Table 5 · p. 8
• Figure 5 · p. 9
• Table 6 · p. 10

Attached appendix

• Table 7 · p. 15
• Table 8 · p. 15
• Table 9 · p. 16
• Figure 6 · p. 16
• Table 10 · p. 17
• Table 11 · p. 17
• Table 12 · p. 17
• Table 13 · p. 18

Argument and organization

1. Motivation and background

   Establish the speech/text sequence-length mismatch before defining tokenization and interleaving.

2. Method and controlled evaluation

   Introduce hierarchical patching, then distinguish equal-compute and equal-data comparisons.

3. Limitations and appendix

   Separate remaining scope limits from reproducibility and robustness detail.

Problem, claim and contribution

The opening makes sequence length a concrete scaling obstacle rather than using a generic multimodal-capability claim.

Section 1, pp. 1–2 · Main paper

Argument and section order

Static, aligned, mixed and curriculum patching are defined before their comparative ablation, so variant names have an operational meaning.

Section 3.1 and Table 6 · Main paper

Figures, captions and visual explanation

Figure 2 gives speech and text distinct token colors, keeps the global transformer central, and uses dashed alignments to connect compression with decoding; its caption explains what each module contributes.

Figure 2, p. 3 · Main paper

Tables, comparisons and evidence

Adjacent Tables 3 and 4 reuse task columns while explicitly changing the budget constraint. This makes an efficiency claim auditable without conflating extra processed data with a better architecture.

Tables 3–4, p. 7 · Main paper

Appendix and reproducibility

The appendix supplies data licenses, local/global dimensions, optimizer details, three-run variability, alignment perturbations and token/FLOP accounting rather than only additional favorable examples.

Appendix A.1–A.12, pp. 13–18 · Main paper

What to reuse—and what not to infer

Transferable lesson: Separate efficiency evidence into equal-compute and equal-data comparisons, and keep the same task columns across both. Define the computational unit before presenting a speedup.

Boundary: Avoid blanket claims that conceal exceptions: Table 5 contains a worse test-other TTS CER for LST. Also reconcile scope wording: the limitation about downstream adaptation sits awkwardly beside the ASR/TTS fine-tuning experiments.

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, Sections 1–8 and Ethics Statement; references pp. 10–13 also inspected

Appendix / supplement

Appendix A.1–A.12, PDF pp. 13–18, including data, architecture, robustness, accounting and LLM-use disclosure

Visually inspected PDF pages

3, 7

• Official ICLR 2026 proceedings conference PDF; PDF creation metadata: Mon Mar 2 20:47:13 2026 EST · 18 pages