Writing study: LoongRL: Reinforcement Learning for Advanced Reasoning over Long Contexts

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Siyuan Wang, Gaokai Zhang, Li Lyna Zhang, Ning Shang, Fan Yang, Dongyao Chen, Mao Yang · 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) · Datasets, benchmarks & measurement resources.

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

20 physical pages

Before references / appendix

≈ 9.63 page-equivalents

Main figures / tables

4 / 6

Attached appendix figures / tables

5 / 4

Attached appendix span

≈ 6.89 page-equivalents

Before references / appendix: ≈ 48.15%References / intervening material: ≈ 17.4%Attached appendix: ≈ 34.45%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	≈ 1.29 pp.	6.45%
2 RELATED WORKS	2	≈ 0.73 pp.	3.65%
3 METHODOLOGY	3	≈ 3.21 pp.	16.05%
4 EXPERIMENTS	6	≈ 3.54 pp.	17.7%
5 CONCLUSION	10	≈ 0.14 pp.	0.7%
REPRODUCIBILITY STATEMENT	10	≈ 0.14 pp.	0.7%
REFERENCES	10	≈ 3.48 pp.	17.4%
A APPENDIX	14	≈ 6.89 pp.	34.45%

All detected figure and table locations

Main text

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

Attached appendix

• Figure 5 · p. 15
• Figure 6 · p. 15
• Figure 7 · p. 18
• Figure 8 · p. 18
• Table 7 · p. 18
• Table 8 · p. 19
• Figure 9 · p. 19
• Table 9 · p. 19
• Table 10 · p. 20

Argument and organization

1. Data difficulty as the bottleneck

   Motivates synthetic question-hiding chains, then explains filtering, contextual distractors and verifiable RL.

2. Generalization, preservation and attribution

   Tests long reasoning, retrieval and short tasks, then isolates KeyChain data and verifier choices.

Problem, claim and contribution

The introduction turns a vague need for better reasoning into three obstacles: task difficulty, expensive long rollouts and short-task retention.

Introduction, pp. 1–2. · Main paper

Argument and section order

Data construction precedes the standard GRPO objective, appropriately emphasizing the main contribution rather than presenting the optimizer as novel.

Sections 3.1–3.2, pp. 3–6. · Main paper

Figures, captions and visual explanation

The construction figure separates filtering, context filling and chain insertion; contrasting target and distractor chains makes an unfamiliar training task tangible.

Figure 2, p. 3. · Main paper

Tables, comparisons and evidence

Adjacent ablation tables distinguish the effect of KeyChain data from the choice of answer verifier, rather than crediting the whole gain to a single component.

Tables 5–6, p. 10. · Main paper

Appendix and reproducibility

Exact prompts, sample data and expanded benchmark protocols make qualitative examples interpretable, including YaRN extension and different reasoning/completion prompting.

Appendices A.2–A.10, pp. 15–19. · Main paper

Clarity and writing conventions

Named reasoning stages make trajectories readable, but a few annotated traces support illustrative interpretation rather than a causal proof of internal cognitive processes.

Figure 1, p. 2; Appendix A.11, p. 20. · Main paper

What to reuse—and what not to infer

Transferable lesson: For a data-centric method, show a concrete transformation of one example, then separately test the data recipe, training objective and transfer conditions.

Boundary: Training contexts are approximately 16K, with some listed examples above 20K. Evaluation reaches 128K using already long-context-capable bases and YaRN. Perfect needle retrieval is not perfect long-context reasoning, and substring matching is not a guarantee against semantic errors or reward exploitation.

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 1–5, all tables and reproducibility statement.

Appendix / supplement

Pages 14–20, Appendix A.1–A.11, prompts, example trajectories, training dynamics and expanded evaluations.

Visually inspected PDF pages

3, 10

• ICLR 2026 official proceedings PDF, retrieved 2026-09-19 · 20 pages