Writing study: LongWriter-Zero: Mastering Ultra-Long Text Generation via Reinforcement Learning

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Yuhao Wu, Yushi Bai, Zhiqiang Hu, Roy Ka-Wei Lee, Juanzi Li · 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: Learning & optimization algorithms (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

26 physical pages

Before references / appendix

≈ 10.62 page-equivalents

Main figures / tables

6 / 1

Attached appendix figures / tables

0 / 4

Attached appendix span

≈ 12.37 page-equivalents

Before references / appendix: ≈ 40.85%References / intervening material: ≈ 11.58%Attached appendix: ≈ 47.58%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	≈ 0.96 pp.	3.69%
2 REINFORCEMENT LEARNING FOR ULTRA-LONG TEXT GENERATION	2	≈ 4.16 pp.	16%
3 EVALUATION	6	≈ 2.98 pp.	11.46%
4 RELATED WORK	9	≈ 0.68 pp.	2.62%
5 CONCLUSION	10	≈ 0.84 pp.	3.23%
ETHICS STATEMENT	11	≈ 0.23 pp.	0.88%
REPRODUCIBILITY STATEMENT	11	≈ 0.14 pp.	0.54%
REFERENCES	11	≈ 3.01 pp.	11.58%
A APPENDIX	14	≈ 12.37 pp.	47.58%

All detected figure and table locations

Main text

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

Attached appendix

• Table 2 · p. 16
• Table 3 · p. 17
• Table 4 · p. 18
• Table 5 · p. 18

Argument and organization

1. Three explicit research questions

   Reward design, thinking and continual pretraining each receive a method/result subsection.

2. Complementary evaluation and case studies

   Critic scores, pairwise preferences, SFT comparisons and human checks precede limitations and examples.

Argument and section order

Numbered research questions become identically named subsections, letting readers map each experimental contrast back to an introductory uncertainty.

Section 1, page 2; Sections 2.2–2.4 · Main paper

Figures, captions and visual explanation

Three aligned training curves reuse colors for model variants and separate writing reward, length reward and output length, showing that these are distinct outcomes rather than interchangeable quality proxies.

Figure 2, page 4 · Main paper

A dashed divider separates automated and human preference donuts; the caption correctly explains half-credit for ties. The graphic should be read as preference under these judges, not a factuality score.

Figure 5, page 9 · Main paper

Tables, comparisons and evidence

The main table separates languages, domains and instruction requirements, then visually boxes a different Elo metric. Ablation rows are cumulative removals, not a complete factorial test of all interactions.

Table 1, page 7; Section 3.2, page 8 · Main paper

Appendix and reproducibility

Selection and judging prompts, corpus percentages, extra baselines and three full case comparisons make evaluation choices inspectable. The limitations enumerate concrete reward-hacking behaviors rather than only generic future work.

Appendix A.2–A.10, pages 14–26 · Main paper

Clarity and writing conventions

The absolute no-annotation/no-synthetic-data framing overstates the protocol: the writing reward uses human preference labels, and the final continual-pretraining corpus includes 1% generated reasoning traces.

Abstract; Sections 2.2 and 2.4; Appendix A.4, Table 2 · Main paper

What to reuse—and what not to infer

Transferable lesson: Use research-question headings and matched learning curves to connect design choices with evidence, and publish evaluation prompts plus specific observed failure modes.

Boundary: RL without SFT target responses is not training without supervised or synthetic signals. Preference judges and small human panels do not establish factual reliability; the appendix ghost-story example also contains an internal biographical inconsistency.

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 1–11, Sections 1–5 and ethics/reproducibility statements

Appendix / supplement

Pages 14–26, Appendix A.1–A.10 including complete English/Chinese case studies

Visually inspected PDF pages

4, 9

• ICLR 2026 final proceedings · 26 pages