Writing study: Learning to Theorize the World from Observation

ICML overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICML 2026 oral writing corpus

Doojin Baek, Gyubin Lee, Junyeob Baek, Hosung Lee, Sungjin Ahn · 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: Architecture & representation (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

43 physical pages

Before references / appendix

≈ 9.3 page-equivalents

Main figures / tables

7 / 3

Attached appendix figures / tables

15 / 24

Attached appendix span

≈ 30.91 page-equivalents

Before references / appendix: ≈ 21.63%References / intervening material: ≈ 6.49%Attached appendix: ≈ 71.88%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	≈ 1.57 pp.	3.65%
2. Learning to Theorize	2	≈ 0.93 pp.	2.16%
3. Neural Theorizer: A World Theory Model	3	≈ 1.93 pp.	4.49%
4. Related Works	5	≈ 0.41 pp.	0.95%
5. Experiments	6	≈ 3.5 pp.	8.14%
6. Limitations & Discussion	9	≈ 0.14 pp.	0.33%
7. Conclusion	9	≈ 0.23 pp.	0.53%
Acknowledgements	10	≈ 0.11 pp.	0.26%
Impact Statement	10	≈ 0.15 pp.	0.35%
References	10	≈ 2.78 pp.	6.47%
A. Use of Large Language Models	13	≈ 0.11 pp.	0.26%
B. Pseudocode of NEO	13	≈ 1.44 pp.	3.35%
C. Additional Experimental Details	14	≈ 13.46 pp.	31.3%
D. Additional Analysis	28	≈ 2.78 pp.	6.47%
E. Visualization Results	30	≈ 1.22 pp.	2.84%
F. Model Hyperparameters	32	≈ 10 pp.	23.26%
G. Extended Related Works	42	≈ 1.91 pp.	4.44%

All detected figure and table locations

Main text

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

Attached appendix

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

Argument and organization

1. Problem and contribution

   Defines theorizing operationally as recovering reusable programs from input-output observations without program labels, then distinguishes explaining the observed pair from transferring the same program to new inputs.

2. Argument sequence

   The general probabilistic formulation is followed by an explicit practical implementation: deterministic execution, VQ discreteness and selected-length reconstruction replace parts of the ideal objective.

3. Evidence and verification

   Three controlled domains separate in-distribution, compositional and length generalization, while codebook, grounding and length-penalty controls probe the proposed mechanism; search budgets differ by task and must accompany comparisons.

Problem, claim and contribution

Defines theorizing operationally as recovering reusable programs from input-output observations without program labels, then distinguishes explaining the observed pair from transferring the same program to new inputs.

Introduction and Section 2, pp. 1-3 · Main paper

Argument and section order

The general probabilistic formulation is followed by an explicit practical implementation: deterministic execution, VQ discreteness and selected-length reconstruction replace parts of the ideal objective.

Sections 3.1-3.5, pp. 3-5 · Main paper

Figures, captions and visual explanation

Intermediate image-editing panels expose variable program lengths next to a component table whose self-explanation and transfer columns separate success modes; the no-grounding row visibly collapses.

Figure 5 and Table 3, p. 8; p. 5 implementation and p. 8 PNGs inspected · Main paper

Tables, comparisons and evidence

Three controlled domains separate in-distribution, compositional and length generalization, while codebook, grounding and length-penalty controls probe the proposed mechanism; search budgets differ by task and must accompany comparisons.

Sections 5.1-5.7, pp. 6-9; Appendix C, pp. 15-28 · Main paper

Appendix and reproducibility

Domain construction discloses known primitive sets used for evaluation, runtime tables separate single-path inference from large search budgets, and per-domain settings identify the shared pretrained encoder-decoder substrate.

Appendix C.1-C.2, pp. 15-17; D, pp. 29-30; F, pp. 32-41 · Main paper

Clarity and writing conventions

The limitations narrow the broad theorizing language to small, discrete, short-program settings and explicitly reject guarantees of human-aligned semantics or recovery of true causal factors.

Limitations, p. 9 · Main paper

What to reuse—and what not to infer

Transferable lesson: Pair an ambitious conceptual framing with precise operational tests, distinguish the general model from the implemented surrogate, and show both intermediate structure and transfer failures.

Boundary: Controlled synthetic compositionality is not unrestricted scientific theory discovery; known evaluation primitives, pretrained encoders and deterministic low-noise tasks constrain the scope, while large test-time search can carry substantial runtime.

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 1-7, pp. 1-9: observation-to-program motivation, formal latent-program model, practical training implementation, domain evaluations, compositional transfer, ablations and limitations.

Appendix / supplement

Writing-focused appendix scan; code, data and external assets not executed. Appendices A-G, pp. 13-43: notation and algorithms, synthetic-domain construction and metrics, additional results, runtime, qualitative programs, domain-specific hyperparameters and extended related work.

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. · 43 pages