Writing study: Instilling an Active Mind in Avatars via Cognitive Simulation

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Jianwen Jiang, Weihong Zeng, Zerong Zheng, Jiaqi Yang, Chao Liang, Wang Liao, Han Liang, Weifeng Chen, XING WANG, Yuan Zhang, Mingyuan Gao · 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

33 physical pages

Before references / appendix

≈ 10.11 page-equivalents

Main figures / tables

5 / 4

Attached appendix figures / tables

8 / 2

Attached appendix span

≈ 15.89 page-equivalents

Before references / appendix: ≈ 30.64%References / intervening material: ≈ 21.21%Attached appendix: ≈ 48.15%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	≈ 1.8 pp.	5.45%
2 RELATED WORK	3	≈ 1.17 pp.	3.55%
3 APPROACH	4	≈ 2.21 pp.	6.7%
4 EXPERIMENTS	6	≈ 3.59 pp.	10.88%
5 CONCLUSION	10	≈ 0.77 pp.	2.33%
REFERENCES	11	≈ 7 pp.	21.21%
A APPENDIX	18	≈ 0.36 pp.	1.09%
B IMPLEMENTATION AND TRAINING DETAILS	18	≈ 1.2 pp.	3.64%
C EVALUATION DETAILS	19	≈ 1.44 pp.	4.36%
D FURTHER ANALYSIS OF THE AGENTIC SYSTEM	21	≈ 7.17 pp.	21.73%
E ADDITIONAL QUALITATIVE RESULTS AND VISUAL ABLATIONS	28	≈ 3.26 pp.	9.88%
F LATENCY OF THE AGENTIC SYSTEM	31	≈ 0.22 pp.	0.67%
G STATEMENTS	31	≈ 1.74 pp.	5.27%
H LIMITATIONS	33	≈ 0.5 pp.	1.52%

All detected figure and table locations

Main text

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

Attached appendix

• Figure 6 · p. 20
• Table 5 · p. 26
• Table 6 · p. 26
• Figure 7 · p. 29
• Figure 8 · p. 29
• Figure 9 · p. 30
• Figure 10 · p. 31
• Figure 11 · p. 31
• Figure 12 · p. 32
• Figure 13 · p. 33

Argument and organization

1. Introduction / Related Work

   Connects semantic-motion limitations to a qualified dual-system analogy.

2. Approach

   Separates semantic planning, reactive video synthesis and modality-conflict remedies.

3. Experiments

   Begins with component ablations before generalization and external comparisons.

4. Appendices A–H

   Provides an explicit reading map to implementation, prompts, visual evidence and failure cases.

Problem, claim and contribution

The cognitive analogy supplies a memorable motivation, but a footnote explicitly disclaims neuroscientific replication. The technical problem becomes conflicting reference-image, audio and semantic guidance, not a claim that the avatar has an actual mind.

§1, pp.1–3, especially p.2 footnote · Main paper

Argument and section order

The overview separates standard VAE/flow-matching components from the new planning and conditioning designs. The method then explains the reference-frame artifact before introducing the pseudo-last-frame remedy.

§§3.1–3.3, pp.4–6 · Main paper

Figures, captions and visual explanation

Figure 2 combines a compact full pipeline with three magnified module panels. Its training/inference rows make the pseudo-frame positional shift visible; Figure 9 compares reference attention against three RoPE offsets using aligned temporal frames.

Figure 2, p.4; Figure 9, p.30 · Main paper

Tables, comparisons and evidence

Tables 1–2 deliberately pair proxy metrics with human preferences because visual fidelity alone misses contextual motion. The strongest evidence is component-specific; the full model does not win every objective metric, despite some broad prose.

§4.2 and Tables 1–2, pp.7–8 · Main paper

Appendix and reproducibility

The appendix specifies three training phases, conditioning details, data filtering and evaluation sample counts. It also declares that reflection and direct reasoning-latent injection were excluded from quantitative comparisons, preventing illustrated extensions from being mistaken for the tested baseline.

Appendices B–C, pp.18–20; Appendix E, p.30 · Main paper

Clarity and writing conventions

The latency and limitations sections expose fixed planning delay, identity drift, hand artifacts and over-articulation. These concrete costs are more reusable than claims that additional thinking time is inherently necessary or desirable.

Appendices F, H, pp.31–33 · Main paper

What to reuse—and what not to infer

Transferable lesson: Explain the conditioning failure before the architecture, visualize train/test differences, and triangulate proxy scores, human judgments and explicit failure cases.

Boundary: GSB is a signed win-minus-loss fraction, not an absolute percentage improvement in quality. Static paper frames do not establish video temporal quality. Several commercial baselines are anonymized, and detailed disclosure does not imply public release of all weights/data or exact reproducibility.

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

PDF pp.1–10, §§1–5 and ethics/reproducibility statements; references pp.11–17

Appendix / supplement

PDF pp.18–33, Appendices A–H, including training, human/MLLM evaluation, three worked planner examples, qualitative ablations, latency and limitations

Visually inspected PDF pages

4, 7, 8, 30

• official ICLR 2026 proceedings PDF · 33 pages