Writing study: Compositional Diffusion with Guided search for Long-Horizon Planning

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Utkarsh Mishra, David He, Yongxin Chen, Danfei Xu · 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: Inference, algorithms & systems efficiency (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

15 physical pages

Before references / appendix

≈ 10.46 page-equivalents

Main figures / tables

7 / 5

Attached appendix figures / tables

0 / 0

Attached appendix span

Separate file recorded below; not in this PDF total

Before references / appendix: ≈ 69.73%References / intervening material: ≈ 30.27%Estimated layout envelopes, not text/artwork area. Unresolved appendix boundaries remain within the undivided post-body material.

Extraction notes: No attached appendix boundary detected. Zero means no recognized caption, not independently verified absence.

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	≈ 1.24 pp.	8.27%
2 BACKGROUND	2	≈ 1.56 pp.	10.4%
3 METHOD	4	≈ 2.15 pp.	14.33%
4 EXPERIMENTAL RESULTS: ROBOTIC PLANNING	6	≈ 2.13 pp.	14.2%
5 CDGS FOR LONG CONTENT GENERATION	8	≈ 1.68 pp.	11.2%
6 RELATED WORK	10	≈ 0.3 pp.	2%
7 CONCLUSION	10	≈ 0.13 pp.	0.87%
8 LIMITATIONS	10	≈ 0.35 pp.	2.33%
9 REPRODUCIBILITY STATEMENT	11	≈ 0.19 pp.	1.27%
10 LLM USAGE	11	≈ 0.09 pp.	0.6%
ACKNOWLEDGMENTS	11	≈ 0.08 pp.	0.53%
REFERENCES	11	≈ 4.54 pp.	30.27%

All detected figure and table locations

Main text

• Figure 1 · p. 1
• Figure 2 · p. 2
• Figure 3 · p. 3
• Figure 4 · p. 4
• Table 1 · p. 7
• Table 2 · p. 7
• Table 3 · p. 7
• Figure 5 · p. 8
• Figure 6 · p. 9
• Table 4 · p. 9
• Table 5 · p. 9
• Figure 7 · p. 10

Attached appendix

No recognized caption labels.

Separate source files—excluded from canonical-PDF totals

A separate author version may repeat main text and references. Whole-file lengths and counts are not extra appendix material and are never added to the canonical PDF totals. Read the study's source scope for which sections received substantive reading versus layout inspection.

Additional source 1

Separate PDF (or ZIP containing it): whole-file layout of 27 physical pages; 11 detected figures and 6 tables. These may include overlapping main-text material.

Detected heading	PDF page	Estimated span

No top-level headings detected; no section lengths inferred.

• Figure 8 · p. 7
• Figure 9 · p. 8
• Figure 10 · p. 9
• Figure 11 · p. 9
• Figure 12 · p. 10
• Figure 13 · p. 11
• Table 6 · p. 11
• Table 7 · p. 12
• Table 8 · p. 13
• Table 9 · p. 14
• Figure 14 · p. 14
• Figure 15 · p. 15
• Table 10 · p. 23
• Table 11 · p. 23
• Figure 16 · p. 25
• Figure 17 · p. 25
• Figure 18 · p. 26

Argument and organization

1. Mode-averaging problem

   Use a common local-to-global failure across planning, panoramas, and video to motivate search.

2. Compositional diffusion with guided search

   Separate candidate generation by resampling from pruning by local feasibility.

3. Robotics and long-content experiments

   Test the same inference principle across domains, then disclose assumptions and implementation details.

Problem, claim and contribution

The running two-route example shows why individually plausible local modes can produce an invalid global sequence, giving the method a concrete target.

Figures 1 and 3, pp. 1, 3 · Main paper · Additional source 2

Argument and section order

Paired pseudocode blocks distinguish the outer population search from compositional-score resampling, avoiding an opaque monolithic algorithm.

Algorithms 1–2, p. 5 · Main paper · Additional source 2

Figures, captions and visual explanation

The pipeline uses blue and yellow magnified panels to separate resampling and pruning, with a shared candidate-population view above.

Figure 4, p. 4 · Main paper · Additional source 2

Tables, comparisons and evidence

Result tables explicitly identify privileged planning information and report visual-generation trade-offs rather than only success examples.

Table 3, p. 7; Tables 4–5, p. 9 · Main paper · Additional source 2

Appendix and reproducibility

A navigable supplement includes prompts, model training, hardware, domain-specific sampling code, wall-clock scaling, and the additional dynamics correction used for TAMP.

Supplement Appendices F–M, printed pp. 26–42 · Main paper · Additional source 2

Clarity and writing conventions

The main-text oracle-matching language should be narrowed to particular settings; task success and aesthetic quality do not improve uniformly.

Abstract; Tables 1, 3, 5; Appendix L · Main paper · Additional source 2

What to reuse—and what not to infer

Transferable lesson: A transferable algorithm is easier to explain when a small running failure case, a two-component diagram, and domain-specific evaluations all use the same local/global vocabulary.

Boundary: The curvature statistic is a heuristic feasibility proxy, not an exact likelihood or feasibility guarantee. Equation 5's product-of-exponentials direction conflicts with minimization prose and the appendix's mean-curvature code. TAMP also uses learned forward dynamics; longer-video tests use few prompts and increased compute.

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

Sections 1–10, pp. 1–11; references pp. 11–15

Appendix / supplement

Separate 27-page supplement, printed pp. 16–42: contents and Appendices A–M

Visually inspected PDF pages

4, 9

• ICLR 2026 official proceedings PDF, retrieved 2026-09-19 · 15 pages
• ICLR 2026 official proceedings PDF, retrieved 2026-09-19 · 27 pages