Writing study: CodeV: Code with Images for Faithful Visual Reasoning via Tool-Aware Policy Optimization

CVPR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← CVPR 2026 oral writing corpus

Xinhai Hou, Shaoyuan Xu, Manan Biyani, Moyan Li, Jia Liu, Todd C. Hollon, Bryan Wang · 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: Training, objectives & optimization (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

11 physical pages

Before references / appendix

≈ 8.05 page-equivalents

Main figures / tables

6 / 2

Attached appendix figures / tables

0 / 0

Attached appendix span

Separate file recorded below; not in this PDF total

Before references / appendix: ≈ 73.18%References / intervening material: ≈ 26.82%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	≈ 1.29 pp.	11.73%
2. Faithfulness evaluation for visual agents. We propose	2	≈ 0.02 pp.	0.18%
3. Empirical gains in faithfulness and accuracy. Train-	2	≈ 0.03 pp.	0.27%
4. Open recipe and resources. We release the full train-	2	≈ 3.55 pp.	32.27%
5. Experiments	6	≈ 2.04 pp.	18.55%
6. Related Work	8	≈ 0.4 pp.	3.64%
7. Conclusion	8	≈ 0.32 pp.	2.91%
References	9	≈ 2.95 pp.	26.82%

All detected figure and table locations

Main text

• Figure 1 · p. 1
• Figure 2 · p. 3
• Figure 3 · p. 5
• Figure 4 · p. 7
• Figure 5 · p. 7
• Table 1 · p. 8
• Table 2 · p. 8
• Figure 6 · 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 13 physical pages; 4 detected figures and 6 tables. These may include overlapping main-text material.

Located appendix only: starts on physical p.1; ≈ 12.93 page-equivalents, 4 figure labels and 6 table labels. This interval is distinct from the whole file.

Detected heading	PDF page	Estimated span
Supplementary Material	1	≈ 0.09 pp.
8. Faithfulness Evaluation Protocol	1	≈ 0.45 pp.
9. Data preparation	1	≈ 0.7 pp.
10. Full results	2	≈ 2.28 pp.
11. Perturbation Analysis	4	≈ 1.46 pp.
12. Tool Use analysis	6	≈ 0.19 pp.
13. Python Sandbox Design	6	≈ 0.81 pp.
14. RL Prompt Design	7	≈ 0.5 pp.
15. Training Dynamics	7	≈ 0.5 pp.
16. Qualitative Examples	8	≈ 4.05 pp.
A. In front of the recycle bin	12	≈ 0.01 pp.
B. To the right of the recycle bin	12	≈ 0.01 pp.
C. To the left of the recycle bin	12	≈ 0.01 pp.
D. Behind the recycle bin	12	≈ 1.88 pp.

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

Argument and organization

1. Preliminaries and evidence §§2–3

   Define agentic interaction and demonstrate the failure before introducing the reward redesign.

2. Method and evaluation §§4–5

   Explain the SFT/RL stages, then separately measure answer performance, tool faithfulness and reward ablations.

Problem, claim and contribution

A correct answer with an irrelevant crop motivates a distinction between outcome accuracy and action-level evidence use, which becomes the organizing question of the paper.

Figure 1 and §1, PDF pp. 1–2 · Main paper · Additional source 2

Argument and section order

The failure analysis appears before the method; later evaluation returns to the same faithfulness definition instead of quietly substituting ordinary benchmark accuracy.

§3, PDF pp. 3–4; §5.3, PDF p. 7 · Main paper · Additional source 2

Figures, captions and visual explanation

Figure 2 uses stacked correct-answer bars to expose faithful versus unfaithful tool use; the caption names its conditioning population, essential for interpreting the proportions.

Figure 2, PDF p. 3 · Main paper · Additional source 2

Tables, comparisons and evidence

Training-stage and reward-design ablations are separate tables, while the tool-invocation trajectory diagnoses why accuracy-only reward can drift toward text-only answering.

Tables 1–2 and Figure 6, PDF p. 8 · Main paper · Additional source 2

Appendix and reproducibility

The supplement exposes judge prompts, data processing, per-benchmark rather than only averaged ablations, tool-output perturbations, training dynamics and both faithful and unfaithful trajectories.

Supplement §§8–16, PDF pp. 1–13 · Main paper · Additional source 2

Clarity and writing conventions

The paper labels the reward-hacking explanation as a hypothesis and distinguishes crop observability from genuine use of the tool output; the supplementary perturbation test addresses the latter gap.

PDF p. 2; Supplement §11, PDF p. 4 · Main paper · Additional source 2

What to reuse—and what not to infer

Transferable lesson: For agentic reasoning, define what counts as evidence use and report it separately from final-answer accuracy; show the judging protocol and counterexamples.

Boundary: A crop containing the target is a limited proxy, not proof of internal chain-of-thought faithfulness. The introductory conditional metric and supplementary joint faithful-and-correct rate have different denominators.

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 PDF pp. 1–11: framing, argument sequence, section organization, figure/table captions, evaluation interpretation and limitations; reference sections identified. This is a writing-focused scan, not an exhaustive technical proof or reproduction audit. Main PDF pages 3 and 6 were rendered and visually inspected.

Appendix / supplement

Writing-focused scan of the separate supplementary PDF, pp. 1–13, including its organization, captions, protocol/evidence detail and limitation material where present; proofs and code listings were inspected for exposition, not independently verified or executed. Supplement §§8–16, PDF pp. 1–13. Linked video/audio demonstrations and non-PDF archive contents were not assessed.

Visually inspected PDF pages

3, 6

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 open-access accepted manuscript; retrieved 2026-09-26 · 11 pages
• Official CVPR 2026 CVF supplementary PDF; retrieved 2026-09-26 · 13 pages