Writing study: Thinking with Drafts: Speculative Temporal Reasoning for Efficient Long Video Understanding

CVPR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← CVPR 2026 oral writing corpus

Pengfei Hu, Meng Cao, Yingyao Wang, Yi Wang, Jiahua Dong, Jun Song, Cheng Yu, Bo Zheng, Xiaodan Liang · Official CVPR 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: Efficiency, inference & systems (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

12 physical pages

Before references / appendix

≈ 8.08 page-equivalents

Main figures / tables

5 / 5

Attached appendix figures / tables

0 / 0

Attached appendix span

Separate file recorded below; not in this PDF total

Before references / appendix: ≈ 67.33%References / intervening material: ≈ 32.67%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	2	≈ 0.95 pp.	7.92%
2. Related Work	3	≈ 0.59 pp.	4.92%
3. Methodology	3	≈ 2.65 pp.	22.08%
4. Experiments	6	≈ 2.53 pp.	21.08%
5. Conclusions	8	≈ 0.23 pp.	1.92%
6. Acknowledgment	9	≈ 0.03 pp.	0.25%
References	9	≈ 3.92 pp.	32.67%

All detected figure and table locations

Main text

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

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 7 physical pages; 8 detected figures and 3 tables. These may include overlapping main-text material.

Located appendix only: starts on physical p.1; ≈ 6.93 page-equivalents, 8 figure labels and 3 table labels. This interval is distinct from the whole file.

Detected heading	PDF page	Estimated span
Supplementary Material	1	≈ 0.28 pp.
1. Experiment Setup	1	≈ 0.48 pp.
2. More Experimental Findings	1	≈ 1.08 pp.
3. Visualizations	2	≈ 0.55 pp.
References	3	≈ 0.09 pp.
4. Limitations	3	≈ 1.67 pp.
A. Red B. Bright yellow-green C. Blue D. Orange	5	≈ 2.78 pp.

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

Argument and organization

1. 3. Methodology

   Defines cooperation, iterative algorithm, policy optimization, and supporting data.

2. 4. Experiments

   Separates short/long-video accuracy from frame-selection, collaboration, and training ablations.

Problem, claim and contribution

The paper identifies growing visual context as an efficiency bottleneck and assigns frame exploration and reasoning to different models.

Figure 1; §1 · Main paper · Additional source 2

Argument and section order

A token-and-model interaction diagram is immediately paired with pseudocode, exposing both the loop and stopping condition.

Figure 2 and Algorithm 1 p. 4 · Main paper · Additional source 2

Figures, captions and visual explanation

A stacked latency chart decomposes vision encoding, projection, prefill, and decoding instead of showing only an aggregate speedup.

Figure 4 p. 7 · Main paper · Additional source 2

Tables, comparisons and evidence

Large-only, small-only, and collaborative variants test whether two models buy a useful accuracy/latency balance; frame selection and training strategy receive separate ablations.

Tables 3–5 p. 8 · Main paper · Additional source 2

Appendix and reproducibility

The supplement includes iteration counts, fixed frame-budget comparisons, model prompts, and explicit higher training overhead.

Supplement §§1–4 · Main paper · Additional source 2

Clarity and writing conventions

The inference case study follows a question through evidence-region selection and verification; it should be read as an illustration, not proof of human-like reasoning.

Figure 5; §4.4 p. 8 · Main paper · Additional source 2

What to reuse—and what not to infer

Transferable lesson: For iterative multimodel inference, show state passed between models, stop rules, and a latency breakdown alongside budget-matched controls.

Boundary: Reported latency depends on model sizes, frame budget, and hardware; an inference saving can coexist with more expensive training.

Apply this lesson using the argument and evidence worksheet.

Inspection record and sources

Checked

2026-09-26

Reading status

Writing-focused manuscript scan + selected visuals inspected

Main text

Manuscript-wide writing scan of main PDF pp. 1–8: framing, section structure, method exposition, figure/table captions, evaluation organization and conclusions; not an exhaustive technical proof audit or reproduction.

Appendix / supplement

Scanned seven supplementary pages on training setup, iteration/reward/frame-budget ablations, long-video retrieval, prompts, multi-iteration cases, and training-cost limitations.

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.

• Official CVPR 2026 CVF proceedings PDF; retrieved 2026-09-26 · 12 pages
• Official CVPR 2026 separate supplementary PDF; retrieved 2026-09-26 · 7 pages