Writing study: A Systematic Study of Behavioral Cloning for Scientific Data Annotation

ICML overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICML 2026 oral writing corpus

Ishaan Singh Chandok, Core Francisco Park · 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

46 physical pages

Before references / appendix

≈ 9.04 page-equivalents

Main figures / tables

6 / 1

Attached appendix figures / tables

33 / 19

Attached appendix span

≈ 33.92 page-equivalents

Before references / appendix: ≈ 19.65%References / intervening material: ≈ 6.61%Attached appendix: ≈ 73.74%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	≈ 1.49 pp.	3.24%
2. Related Work	3	≈ 0.58 pp.	1.26%
3. Framework	3	≈ 0.96 pp.	2.09%
4. Results	4	≈ 3.9 pp.	8.48%
5. Discussion	8	≈ 0.89 pp.	1.93%
Impact Statement	9	≈ 0.48 pp.	1.04%
Acknowledgements	9	≈ 0.13 pp.	0.28%
References	10	≈ 3.04 pp.	6.61%
Appendix	13	≈ 0.06 pp.	0.13%
A. Tasks	13	≈ 9.49 pp.	20.63%
B. Methodology	22	≈ 5.46 pp.	11.87%
C. Additional Results	28	≈ 13 pp.	28.26%
D. Real-Data Experiments	41	≈ 5 pp.	10.87%
E. Visualizations	46	≈ 0.91 pp.	1.98%

All detected figure and table locations

Main text

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

Attached appendix

• Table 2 · p. 13
• Figure 7 · p. 15
• Figure 8 · p. 16
• Figure 9 · p. 17
• Figure 10 · p. 18
• Figure 11 · p. 19
• Figure 12 · p. 20
• Figure 13 · p. 21
• Table 3 · p. 21
• Figure 14 · p. 22
• Figure 15 · p. 23
• Table 4 · p. 23
• Figure 16 · p. 24
• Table 5 · p. 24
• Figure 17 · p. 25
• Table 6 · p. 25
• Figure 18 · p. 26
• Figure 19 · p. 28
• Figure 20 · p. 29
• Figure 21 · p. 30
• Figure 22 · p. 30
• Figure 23 · p. 31
• Figure 24 · p. 31
• Figure 25 · p. 32
• Figure 26 · p. 32
• Figure 27 · p. 33
• Figure 28 · p. 33
• Figure 29 · p. 33
• Figure 30 · p. 34
• Figure 31 · p. 34
• Table 7 · p. 35
• Figure 32 · p. 35
• Table 8 · p. 36
• Figure 33 · p. 36
• Table 9 · p. 36
• Figure 34 · p. 37
• Figure 35 · p. 38
• Table 10 · p. 38
• Table 11 · p. 39
• Figure 36 · p. 39
• Table 12 · p. 39
• Figure 37 · p. 40
• Table 13 · p. 40
• Table 14 · p. 41
• Table 15 · p. 41
• Table 16 · p. 41
• Table 17 · p. 42
• Figure 38 · p. 43
• Figure 39 · p. 44
• Table 18 · p. 45
• Table 19 · p. 45
• Table 20 · p. 45

Argument and organization

1. Problem and contribution

   Explains why synthetic annotators enable controlled experiments before investment in real scientific annotation collection.

2. Argument sequence

   Single-task behavior leads to scaling and transfer, then internal probes and a separate real-data tracing test; limitations return to scale, action-space and context boundaries.

3. Evidence and verification

   The text retains equal-FLOP failure of larger models, failed in-context adaptation, unsuccessful DAgger and below-chance mistake-probe transfer to a qualitatively different task.

Problem, claim and contribution

Explains why synthetic annotators enable controlled experiments before investment in real scientific annotation collection.

Introduction and contribution list, pp. 1-2 · Main paper

Argument and section order

Single-task behavior leads to scaling and transfer, then internal probes and a separate real-data tracing test; limitations return to scale, action-space and context boundaries.

Sections 4-5, pp. 4-9 · Main paper

Figures, captions and visual explanation

A task table pairs challenges with action-count ranges. The multitask figure separates loss, decision actions, motor actions and task trajectories rather than treating aggregate loss as behavior.

Table 1, p. 5; Figures 3-4, p. 7; both PNGs inspected · Main paper

Tables, comparisons and evidence

The text retains equal-FLOP failure of larger models, failed in-context adaptation, unsuccessful DAgger and below-chance mistake-probe transfer to a qualitatively different task.

Sections 4.2-4.4, pp. 5-8 · Main paper

Appendix and reproducibility

Task-specific definitions expose matching tolerances and topology-sensitive metrics; real-data experiments and human annotation validation have separate destinations.

Appendix A.2, p. 20; D.1-D.2, pp. 41-45 · Main paper

Clarity and writing conventions

Finding-led paragraphs pair counterintuitive results with protocol distinctions, especially teacher-forced correction versus free-running behavior.

Section 4.1 continuation, p. 5 · Main paper

What to reuse—and what not to infer

Transferable lesson: Use controlled evaluation modes and task-level failures to explain why a smooth training curve can hide missing behavior.

Boundary: Synthetic virtual-annotator conclusions are bounded by the tested tasks and 25M–320M scale; shared probe structure is not a universal concept of error.

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 1-5, pp. 1-9: procedural study design, single/multitask learning, adaptation, representation probes, connectomics transfer and limitations.

Appendix / supplement

Writing-focused appendix scan; code, data and external assets not executed. Appendices A-E, pp. 13-46: task/interface and metric specifications, training/evaluation, detailed controls, real-data validation and external-video destination; videos not executed.

Visually inspected PDF pages

5, 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 ICML 2026 proceedings PDF, PMLR volume 306; retrieved 2026-10-07. Public OpenReview decision notes were not independently reverified. · 46 pages