Writing study: Molmo2: Open Weights and Data for Vision-Language Models with Video Understanding and Grounding

CVPR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← CVPR 2026 oral writing corpus

Christopher Clark, Jieyu Zhang, Zixian Ma, Jae Sung Park, Rohun Tripathi, Sangho Lee, Reza Salehi, Jason Ren, Chris Dongjoo Kim, Yinuo Yang, Vincent Shao, Yue Yang, Weikai Huang, Ziqi Gao, Taira Anderson, Jianrui Zhang, Jitesh Jain, George Stoica, Ali Farhadi, Ranjay Krishna · 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: Datasets, benchmarks & measurement (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

17 physical pages

Before references / appendix

≈ 8.18 page-equivalents

Main figures / tables

3 / 7

Attached appendix figures / tables

1 / 0

Attached appendix span

≈ 8.8 page-equivalents

Before references / appendix: ≈ 48.12%References / intervening material: ≈ 0.12%Attached appendix: ≈ 51.76%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	≈ 1.64 pp.	9.65%
2. Data	3	≈ 1.4 pp.	8.24%
3. Training	4	≈ 2.13 pp.	12.53%
4. Evaluation	6	≈ 2.06 pp.	12.12%
5. Conclusion	8	≈ 0.13 pp.	0.76%
Acknowledgements	9	≈ 0.13 pp.	0.76%
References	9	≈ 0.03 pp.	0.18%
A. Bewley, J. Bingham, M. Bloesch, et al. Gemini	9	≈ 3.12 pp.	18.35%
B. Fang, J. Zhang, Y. R. Wang, S. Lee, W. Han,	12	≈ 0.79 pp.	4.65%
C. Zhu, O. Tafjord, P. Clark, and A. Kalyan. Learn	13	≈ 1.47 pp.	8.65%
D. Parikh, and M. Rohrbach. Towards VQA models	14	≈ 3.41 pp.	20.06%

All detected figure and table locations

Main text

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

Attached appendix

• Figure QA · p. 11

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 31 physical pages; 27 detected figures and 14 tables. These may include overlapping main-text material.

Located appendix only: starts on physical p.2; ≈ 29.96 page-equivalents, 27 figure labels and 14 table labels. This interval is distinct from the whole file.

Detected heading	PDF page	Estimated span
Author Contributions	1	≈ 1 pp.
Appendix	2	≈ 0.1 pp.
A. Model details	2	≈ 1.46 pp.
B. Training details	3	≈ 1.93 pp.
C. Evaluation Details	5	≈ 4.59 pp.
D. Additional results	10	≈ 0.73 pp.
E. Test time scaling with 128-frame model	10	≈ 2.77 pp.
F. Dataset details	13	≈ 5.11 pp.
G. Data examples	18	≈ 0.29 pp.
H. Related works	19	≈ 0.79 pp.
I. Limitations	19	≈ 0.75 pp.
J. Qualitative results	20	≈ 11.41 pp.

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

Argument and organization

1. Data-led capability story

   Introduce grounded video capability, then describe dataset construction before architecture/training.

2. Evaluation separated by output

   Distinguish general understanding, counting/pointing, tracking and controlled data/model ablations.

Problem, claim and contribution

The opening combines a capability gap in video grounding with transparency of weights/data/recipes, rather than claiming a wholly new model skeleton.

Introduction and Figure 1 · Main paper · Additional source 2

Argument and section order

Data receives its own early section; the architecture is described explicitly as a standard connector design, keeping novelty claims aligned with the actual contribution.

Sections 2–3 · Main paper · Additional source 2

Figures, captions and visual explanation

The data-mixture table distinguishes new datasets with color while retaining names, sampling rates, dataset counts and example counts; it makes the training composition inspectable.

Table 1, PDF p. 4 · Main paper · Additional source 2

Tables, comparisons and evidence

Counting/pointing and tracking are separated, with model groups labeled by openness and incompatible outputs shown as unavailable instead of forcing a fabricated score.

Tables 3–4, PDF p. 7 · Main paper · Additional source 2

Appendix and reproducibility

The supplement publishes annotation interfaces, sampling details, human-preference intervals and failure cases; its limitations qualify the openness claim by disclosing closed-data vision pretraining and closed text-model assistance.

Supplement Sections A–J, especially I and J, pp. 19–31 · Main paper · Additional source 2

Clarity and writing conventions

The main result bullets retain the long-video weakness and distinguish fully open from open-weight alternatives, giving headline comparisons a defined reference class.

Section 4.1, PDF p. 7 · Main paper · Additional source 2

What to reuse—and what not to infer

Transferable lesson: For foundation-model/data releases, state precisely what is open, publish the sampling mixture, and separate capability comparisons from provenance claims.

Boundary: Fully open is qualified by the supplement's closed-data SigLIP 2 encoder and closed text-only generation assistance. API prompting and point-to-mask conversions also affect comparison fairness.

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

Separate supplement pp. 1–31 scanned: model/training/evaluation details, full data mix, human preference uncertainty, tracking results, scaling, annotation interfaces, examples, limitations and failure cases.

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 · 17 pages
• Official CVPR 2026 separate supplementary PDF; retrieved 2026-09-26 · 31 pages