Writing study: Unsupervised Partner Design Enables Robust Ad-hoc Teamwork

ICML overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICML 2026 oral writing corpus

Constantin Ruhdorfer, Matteo Bortoletto, Victor Oei, Anna Penzkofer, Andreas Bulling · Official ICML 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: Training, objectives & optimization (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

27 physical pages

Before references / appendix

≈ 8.61 page-equivalents

Main figures / tables

8 / 2

Attached appendix figures / tables

14 / 5

Attached appendix span

≈ 13.91 page-equivalents

Before references / appendix: ≈ 31.89%References / intervening material: ≈ 16.59%Attached appendix: ≈ 51.52%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	≈ 0.95 pp.	3.52%
2. Related Work	2	≈ 0.55 pp.	2.04%
3. Preliminaries	2	≈ 1.08 pp.	4%
4. Unsupervised Partner Design	3	≈ 1.8 pp.	6.67%
5. Experiments in Fixed Environments	5	≈ 2.74 pp.	10.15%
6. Experiments with Procedural Generation	8	≈ 0.5 pp.	1.85%
7. Discussion and Limitations	8	≈ 0.35 pp.	1.3%
8. Conclusion	9	≈ 0.09 pp.	0.33%
Impact Statement	9	≈ 0.2 pp.	0.74%
Acknowledgements	9	≈ 0.07 pp.	0.26%
References	9	≈ 4.48 pp.	16.59%
A. Infrastructure & Tools	14	≈ 0.27 pp.	1%
B. Reproducibility Statement	14	≈ 0.14 pp.	0.52%
C. UPD and Convention Selection Extended	14	≈ 1.02 pp.	3.78%
D. Additional Comparison with ROTATE and fine-tuned UPD	15	≈ 0.2 pp.	0.74%
E. Additional Training Details	15	≈ 3.5 pp.	12.96%
F. Observations on Computational Requirements of UPD	19	≈ 1.55 pp.	5.74%
G. Additional Environment Details	20	≈ 0.96 pp.	3.56%
H. Training Curves	21	≈ 2.73 pp.	10.11%
I. Partner Curriculum Dynamics	24	≈ 0.32 pp.	1.19%
J. Additional Results in Overcooked	24	≈ 1.51 pp.	5.59%
K. Additional User Study Details	26	≈ 1.71 pp.	6.33%

All detected figure and table locations

Main text

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

Attached appendix

• Table 3 · p. 15
• Table 4 · p. 16
• Table 5 · p. 16
• Table 6 · p. 17
• Table 7 · p. 17
• Figure 9 · p. 21
• Figure 10 · p. 22
• Figure 11 · p. 22
• Figure 12 · p. 22
• Figure 13 · p. 22
• Figure 14 · p. 23
• Figure 15 · p. 23
• Figure 16 · p. 23
• Figure 17 · p. 23
• Figure 18 · p. 24
• Figure 19 · p. 24
• Figure 20 · p. 25
• Figure 21 · p. 26
• Figure 22 · p. 27

Argument and organization

1. Problem and contribution

   Extends unsupervised environment design to partner behavior and separates fixed-level teaming from joint generalization over levels and partners.

2. Argument sequence

   The generator and selection score precede a small convention-selection example; separate experiment sections then evaluate artificial partners, humans and procedural environments.

3. Evidence and verification

   The twelve-participant study randomizes all agent-layout combinations and reports paired tests with Holm–Bonferroni correction, expressly stopping short of declaring overall superiority.

Problem, claim and contribution

Extends unsupervised environment design to partner behavior and separates fixed-level teaming from joint generalization over levels and partners.

Contributions, p. 2; Section 5, p. 5 · Main paper

Argument and section order

The generator and selection score precede a small convention-selection example; separate experiment sections then evaluate artificial partners, humans and procedural environments.

Sections 4-6, pp. 3-8 · Main paper

Figures, captions and visual explanation

A fixed-mixture sweep exposes tuning sensitivity, while the human-study figure reports metric directions and corrected comparisons beside the procedural-generalization table.

Figure 3, p. 5; Figure 7/Table 2/Figure 8, p. 8; both PNGs inspected · Main paper

Tables, comparisons and evidence

The twelve-participant study randomizes all agent-layout combinations and reports paired tests with Holm–Bonferroni correction, expressly stopping short of declaring overall superiority.

Section 5.3, pp. 7-8 · Main paper

Appendix and reproducibility

A transparency note states that earlier baseline implementation/configuration issues were corrected and rerun; later pages retain all questionnaire-response distributions.

Appendix B, p. 14; K and Figure 22, pp. 26-27 · Main paper

Clarity and writing conventions

The human-study conclusion is calibrated to sample size, and the reproducibility statement identifies a consequential source-version correction rather than silently replacing results.

Section 5.3, p. 8; Appendix B, p. 14 · Main paper

What to reuse—and what not to infer

Transferable lesson: Separate artificial, human and environment-shift evidence, and document corrected baseline versions explicitly when the public manuscript has evolved.

Boundary: Population-free does not mean generator-free or compute-free; a conceptual two-action game and a small human sample do not establish universal partner generalization.

Apply this lesson using the argument and evidence worksheet.

Inspection record and sources

Checked

2026-10-07

Reading status

Writing-focused manuscript scan + selected visuals inspected

Main text

Writing-focused distributed manuscript scan, not exhaustive technical/proof validation. Sections 1-8, pp. 1-9: partner-space curriculum motivation, online generation/selection, illustrative convention game, artificial/human teaming and procedural partner-level generalization.

Appendix / supplement

Writing-focused appendix scan; code, data and external assets not executed. Appendices A-K, pp. 14-27: infrastructure and correction note, conceptual extension, RL setup, compute observations, task/training curves, alternative scores and human-study details.

Visually inspected PDF pages

5, 8

This is a writing-focused scan of the manuscript and the stated supplementary scope: framing, organization, visual explanation, evidence presentation and wording. It is not a line-by-line proof audit, reproduction of results or exhaustive scientific review.

• Published ICML 2026 proceedings PDF, PMLR volume 306; retrieved 2026-10-07. Public OpenReview decision notes were not independently reverified. · 27 pages