Writing study: NextStep-1: Toward Autoregressive Image Generation with Continuous Tokens at Scale

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

chunrui han, Guopeng Li, Jingwei Wu, Quan Sun, Yan Cai, Yuang Peng, Zheng Ge, Deyu Zhou, Haomiao Tang, Hongyu Zhou, Kenkun Liu, Shu-Tao Xia, Binxing Jiao, Daxin Jiang, Xiangyu Zhang, Yibo Zhu · 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).

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

23 physical pages

Before references / appendix

≈ 10.11 page-equivalents

Main figures / tables

4 / 6

Attached appendix figures / tables

5 / 3

Attached appendix span

≈ 6.89 page-equivalents

Before references / appendix: ≈ 43.96%References / intervening material: ≈ 26.09%Attached appendix: ≈ 29.96%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	≈ 1.17 pp.	5.09%
2 METHOD	2	≈ 1.13 pp.	4.91%
3 TRAINING RECIPES	3	≈ 1.68 pp.	7.3%
4 MODEL PERFORMANCE	5	≈ 0.31 pp.	1.35%
5 DISCUSSIONS	5	≈ 3.82 pp.	16.61%
6 RELATED WORK	9	≈ 0.81 pp.	3.52%
7 CONCLUSION	10	≈ 0.19 pp.	0.83%
ACKNOWLEDGMENTS	10	≈ 0.47 pp.	2.04%
REFERENCES	11	≈ 6 pp.	26.09%
APPENDIX	17	≈ 0.04 pp.	0.17%
A MORE RESULTS	17	≈ 1.96 pp.	8.52%
B DATA PIPELINES	19	≈ 1.38 pp.	6%
C IMAGE TOKENIZER	20	≈ 0.34 pp.	1.48%
D LIMITATIONS AND CHALLENGES	20	≈ 2.73 pp.	11.87%
E LLM USAGE	23	≈ 0.44 pp.	1.91%

All detected figure and table locations

Main text

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

Attached appendix

• Figure A1 · p. 17
• Table A1 · p. 18
• Table A2 · p. 18
• Figure A2 · p. 20
• Figure A3 · p. 21
• Figure A4 · p. 22
• Figure A5 · p. 22
• Table A3 · p. 23

Argument and organization

1. Continuous-token architecture and training recipe

   A causal language backbone and patch-wise flow head are connected to tokenizer design and staged multimodal training.

2. Benchmark results followed by mechanism questions

   Generation/editing scores precede head-size and latent-stability investigations, then detailed failures and data provenance.

Argument and section order

A stage-by-stage recipe table aligns learning rates, token budgets, resolutions and data ratios, helping readers distinguish tokenizer pretraining from backbone pretraining, SFT and DPO.

Table 1, page 4; Sections 3.1–3.3 · Main paper

Figures, captions and visual explanation

The overview shows the shared causal sequence and enlarges just one patch-level flow prediction in an inset; this cleanly separates global autoregression from the local iterative sampler.

Figure 1, page 2 · Main paper

The appendix groups generation, paired editing and free-form manipulation into visual bands, making task scope immediately visible without treating selected examples as a representative success rate.

Figure A1, page 17 · Main paper

Tables, comparisons and evidence

Benchmark tables distinguish autoregressive and diffusion families and mark prompt rewriting versus Self-CoT. NextStep-1's GenEval 0.63/0.73 does not exceed all listed autoregressive alternatives, so the headline state-of-the-art claim needs benchmark-specific wording.

Table 2, page 6 · Main paper

Appendix and reproducibility

Normalization code resolves a potentially ambiguous prose description: the operation normalizes across channels per token. Failure galleries, data-pipeline categories and a labeled theoretical latency breakdown add valuable implementation context.

Appendices B–D, pages 19–23 · Main paper

Clarity and writing conventions

The tokenizer discussion acknowledges that dispersion versus robustness is not disentangled, but the head-size ablation only retrains heads for 10,000 steps; it cannot establish that a flow head has no modeling role.

Sections 5.1–5.2, pages 7–9 · Main paper

What to reuse—and what not to infer

Transferable lesson: Use an inset to show local computation inside a simple global architecture, accompany it with a staged training table, and pair appealing galleries with concrete failure examples.

Boundary: Sequential autoregression still invokes an iterative flow sampler per patch. The latency table is theoretical; proprietary training data and model-based evaluation constrain reproducibility. NextStep-1.1's additional training and qualitative artifact reduction must not be attributed to the evaluated NextStep-1 checkpoint.

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 1–7

Appendix / supplement

Pages 17–23, Appendices A–E including data pipelines, tokenizer code, failure cases and latency analysis

Visually inspected PDF pages

2, 6, 17

• ICLR 2026 final proceedings · 23 pages