Writing study: From Pixels to Tokens: A Systematic Study of Latent Action Supervision for Vision-Language-Action Models

ICML overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICML 2026 oral writing corpus

Yihan Lin, Haoyang Li, Yang Li, Haitao Shen, Yihan Zhao, Chao Shao, Jing Zhang · 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: Empirical analysis, diagnosis & auditing (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

25 physical pages

Before references / appendix

≈ 8.84 page-equivalents

Main figures / tables

6 / 5

Attached appendix figures / tables

5 / 17

Attached appendix span

≈ 13.91 page-equivalents

Before references / appendix: ≈ 35.36%References / intervening material: ≈ 9%Attached appendix: ≈ 55.64%Estimated layout envelopes, not text/artwork area. Unresolved appendix boundaries remain within the undivided post-body material.

Source-bound correction: Restore Figure6, the right-hand caption alongside Figure5 on physical p8. The original Poppler line merged both captions; p8 rendered and visually checked.

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	≈ 1.03 pp.	4.12%
2. Related Work	2	≈ 0.56 pp.	2.24%
3. Preliminaries	2	≈ 1.27 pp.	5.08%
4. Methodology	4	≈ 2.11 pp.	8.44%
5. Experiments	6	≈ 3.13 pp.	12.52%
6. Limitations	9	≈ 0.2 pp.	0.8%
7. Conclusion	9	≈ 0.08 pp.	0.32%
Acknowledgements	9	≈ 0.05 pp.	0.2%
Impact Statement	9	≈ 0.08 pp.	0.32%
References	9	≈ 2.25 pp.	9%
A. Latent Action Model Details	12	≈ 2.15 pp.	8.6%
B. Architecture Details	14	≈ 1.01 pp.	4.04%
C. VLA Baselines and Evaluation Protocols	15	≈ 0.94 pp.	3.76%
D. VLA Training Implementation Details	16	≈ 0.91 pp.	3.64%
E. Additional Experimental Results	17	≈ 2.22 pp.	8.88%
F. Real-World Experiments on the JAKA Arm	19	≈ 6.7 pp.	26.8%

All detected figure and table locations

Main text

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

Attached appendix

• Figure 7 · p. 13
• Table 6 · p. 13
• Table 7 · p. 15
• Table 8 · p. 16
• Figure 8 · p. 17
• Table 9 · p. 18
• Table 10 · p. 18
• Table 11 · p. 18
• Table 12 · p. 19
• Figure 9 · p. 20
• Figure 10 · p. 20
• Figure 11 · p. 21
• Table 13 · p. 22
• Table 14 · p. 22
• Table 15 · p. 22
• Table 16 · p. 23
• Table 17 · p. 23
• Table 18 · p. 23
• Table 19 · p. 24
• Table 20 · p. 24
• Table 21 · p. 25
• Table 22 · p. 25

Argument and organization

1. 1 Introduction; 2 Related Work; 3 Preliminaries

   Defines the comparison target as supervision design, not a contest between latent-action encoders.

2. 4 Methodology

   Maps two conceptual directions to four implementations under a shared baseline.

3. 5 Experiments; 6 Limitations; 7 Conclusion

   Separates formulation, integration, representation and scene-transfer questions.

Problem, claim and contribution

The opening explicitly studies how latent supervision affects policy learning under a unified baseline, limiting what its ranking claims compare.

Introduction, p.1 · Main paper

Argument and section order

Results are organized by comparison axis, with discrete-versus-continuous representation and placeholder-length controls answering separate alternatives.

Sections 5.2-5.6, pp.7-9 · Main paper

Figures, captions and visual explanation

The architecture panel uses shared colored token roles across all four variants; real-world displays distinguish completion scores from binary task success and label rollout counts.

Figure 3, p.4; Figure 4 and Table 3, p.7 · Main paper

Tables, comparisons and evidence

The paper preserves task-dependent rankings: direct image-latent supervision leads some tasks, while action tokens or conditional designs benefit others.

Sections 5.2-5.5, pp.7-8 · Main paper

Appendix and reproducibility

A vision-only variant tests the small action-supervision contribution, while the physical protocol fixes cameras and calibration and enumerates ten rollouts per model-task.

Table 6, p.13; Appendix F, p.19 · Main paper

Clarity and writing conventions

Mechanistic interpretations use hypothesize and suggest, and limitations leave open whether stronger latent representations would change the observed ranking.

Section 5.2, p.7; Limitations, p.9 · Main paper

What to reuse—and what not to infer

Transferable lesson: For a systematic study, make the controlled comparison axes visible and preserve task-specific exceptions to an overall recommendation.

Boundary: A four-strategy study with one physical robot does not exhaust the design space; completion-percentage scores should not be relabeled as success probabilities.

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-oriented scan of the main argument through physical p.9, with focused physical-page text inspection on 1, 3, 5, 7, 8, 9. Framing, section progression, mechanism/proof explanation, displays, comparison conditions and conclusions were inspected. This is not exhaustive technical reading, proof verification or reproduction.

Appendix / supplement

Appendix organization and main-text pointers were checked; focused text inspection of physical pages 13, 19. Remaining appendix proofs and details were not exhaustively read. No separately linked supplement inspected.

Visually inspected PDF pages

4, 7

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 proceedings PDF, PMLR volume 306, retrieved 2026-10-07 · 25 pages