Writing study: ScaleCUA: Scaling Open-Source Computer Use Agents with Cross-Platform Data

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Zhaoyang Liu, JingJing Xie, Zichen Ding, Zehao Li, Bowen Yang, Zhenyu Wu, Xuehui Wang, Qiushi Sun, Shi Liu, Weiyun Wang, Shenglong Ye, Qingyun Li, Zeyue Tian, Gen Luo, Xiangyu Yue, Biqing Qi, Kai Chen, Bowen Zhou, Yu Qiao, Qifeng Chen, Wenhai Wang · 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: Datasets, benchmarks & measurement resources (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

60 physical pages

Before references / appendix

≈ 10.55 page-equivalents

Main figures / tables

6 / 4

Attached appendix figures / tables

26 / 16

Attached appendix span

≈ 42.89 page-equivalents

Before references / appendix: ≈ 17.58%References / intervening material: ≈ 10.93%Attached appendix: ≈ 71.48%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	≈ 1.6 pp.	2.67%
2 RELATED WORK	3	≈ 0.58 pp.	0.97%
3 CROSS-PLATFORM INTERACTIVE DATA PIPELINE	3	≈ 2.06 pp.	3.43%
4 THE DESIGN OF COMPUTER USE AGENTS	5	≈ 1.37 pp.	2.28%
5 EXPERIMENTS	7	≈ 3.5 pp.	5.83%
6 CONCLUSION	10	≈ 0.42 pp.	0.7%
ACKNOWLEDGMENTS	11	≈ 0.08 pp.	0.13%
ETHICS STATEMENT	11	≈ 0.22 pp.	0.37%
REPRODUCIBILITY STATEMENT	11	≈ 0.15 pp.	0.25%
REFERENCES	11	≈ 6.56 pp.	10.93%
A APPENDIX	18	≈ 42.89 pp.	71.48%

All detected figure and table locations

Main text

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

Attached appendix

• Table 5 · p. 19
• Table 6 · p. 20
• Table 7 · p. 21
• Table 8 · p. 21
• Table 9 · p. 22
• Table 10 · p. 22
• Table 11 · p. 23
• Table 12 · p. 23
• Figure 7 · p. 23
• Table 13 · p. 24
• Table 14 · p. 24
• Figure 8 · p. 25
• Figure 9 · p. 25
• Figure 10 · p. 26
• Table 15 · p. 27
• Table 16 · p. 33
• Figure 11 · p. 33
• Table 17 · p. 34
• Table 18 · p. 35
• Figure 12 · p. 35
• Figure 13 · p. 35
• Table 19 · p. 36
• Figure 14 · p. 36
• Figure 15 · p. 37
• Figure 16 · p. 37
• Figure 17 · p. 37
• Figure 18 · p. 38
• Figure 19 · p. 38
• Table 20 · p. 39
• Figure 20 · p. 40
• Figure 21 · p. 41
• Figure 22 · p. 42
• Figure 23 · p. 42
• Figure 24 · p. 43
• Figure 25 · p. 44
• Figure 26 · p. 44
• Figure 27 · p. 45
• Figure 28 · p. 45
• Figure 29 · p. 46
• Figure 30 · p. 47
• Figure 31b · p. 47
• Figure 31 · p. 60

Argument and organization

1. Data infrastructure before agent design

   A collection bottleneck motivates a hybrid pipeline, then three inference modes and platform-specific evaluations.

2. Engineering evidence and replication details

   Diagnostics and ablations precede an extensive appendix of collection procedures, failure examples and exact prompts.

Problem, claim and contribution

The paper makes a defensible engineering contribution explicit: collecting consistent GUI trajectories across heterogeneous platforms is the bottleneck, rather than a claimed new model architecture.

Sections 1 and 3; page 5 discussion. · Main paper

Argument and section order

Understanding, grounding and task completion recur as dataset families and evaluation dimensions. The separate grounding, direct-action and reasoned-action modes prevent conflating a localizer with an autonomous planner.

Sections 3.2, 4.2 and 5.1; Figure 3. · Main paper

Figures, captions and visual explanation

Figure 2 groups environment interaction, raw trajectories and annotated training tasks into legible containers, with separate arrows for collection and training. It makes the human-versus-agent data origins visible.

Page 4, Figure 2. · Main paper

Tables, comparisons and evidence

Table 3 separates native agents from planner–grounder workflows and labels step budgets, including unknown or larger budgets. Its platform columns expose weak macOS results instead of hiding them in one average.

Page 8, Table 3. · Main paper

Appendix and reproducibility

Platform-specific parsing failures, browser-native dialogs and benchmark revisions are documented alongside prompts. These details explain why evaluation environments and screenshots are part of the experimental specification.

Appendices A.8–A.10, pages 38–60. · Main paper

Clarity and writing conventions

Example tables require the same checking as metrics: the referring-OCR row repeats a search-bar appearance answer despite an Apple Watch screenshot, and a functionality row uses a position prompt. These visible mismatches weaken otherwise useful task definitions.

Page 34, Table 17. · Main paper

What to reuse—and what not to infer

Transferable lesson: For a data-engineering contribution, connect acquisition decisions to task-aligned ablations, and publish the inconvenient platform failures and benchmark modifications needed to reproduce the claim.

Boundary: Strong grounding does not establish universal planning or safety. WebArena-Lite-v2 modifies the original benchmark; native and GPT-4o-planner results are different systems. Cross-platform data is not uniformly better than domain-only data. Synthetic rationales can use hindsight during annotation, and public release claims were not independently audited here.

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 1–6, limitations, ethics and reproducibility; references through page 17.

Appendix / supplement

Pages 18–60, A.1–A.10: additional results, public data, actions, failures, curation, data examples, platform-specific collection lessons, benchmark revisions and all prompt templates.

Visually inspected PDF pages

4, 8, 34

• ICLR 2026 final proceedings · 60 pages