Writing study: Through the Lens of Contrast: Self-Improving Visual Reasoning in VLMs

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Zhiyu Pan, Yizheng Wu, Jiashen Hua, Junyi Feng, Shaotian Yan, Bing Deng, Zhiguo Cao, Jieping Ye · 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: Learning & optimization algorithms (primary) · Datasets, benchmarks & measurement resources.

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

19 physical pages

Before references / appendix

≈ 9.33 page-equivalents

Main figures / tables

6 / 4

Attached appendix figures / tables

3 / 0

Attached appendix span

≈ 3.42 page-equivalents

Before references / appendix: ≈ 49.11%References / intervening material: ≈ 32.89%Attached appendix: ≈ 18%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	≈ 1.3 pp.	6.84%
2 RELATED WORKS	2	≈ 1.36 pp.	7.16%
3 VISUAL CONTRASTIVE SELF-TAUGHT REASONER (VC-STAR)	4	≈ 2.25 pp.	11.84%
4 EXPERIMENTS	6	≈ 3.69 pp.	19.42%
5 CONCLUSION	10	≈ 0.16 pp.	0.84%
REFERENCES	10	≈ 6.25 pp.	32.89%
A APPENDIX	16	≈ 3.42 pp.	18%

All detected figure and table locations

Main text

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

Attached appendix

• Figure 7 · p. 17
• Figure 8 · p. 19
• Figure 9 · p. 19

Argument and organization

1. Contrastive observation and data construction

   Turn an observed failure-repair pattern into a training-data hypothesis

2. Think, contrast and rethink

   Separate VLM visual comparison from LLM rationale rewriting

3. Baseline comparisons and design questions

   Test data selection, counterparts and model transfer

Problem, claim and contribution

A skateboard example motivates the narrower question of repairing visually ungrounded rationales. The opening observation is concrete, but correct final answers and selected examples do not alone verify every intermediate visual claim.

Section 1; Figure 1, page 2 · Main paper

Argument and section order

Figures 3 and 4 distinguish counterpart curation from rationale generation. This is especially useful because training uses paired images, answer hints and a 72B text rewriter, while deployment returns to ordinary single-image inference.

Sections 3.1–3.2 and 4.1; page 5 · Main paper

Figures, captions and visual explanation

The mechanism diagram carries matching colors across stages and shows discarded easy/hard examples. Visual inspection also exposes a serious Figure 5 assembly error: the fork-count example repeats the traffic-sign answer from its neighboring column.

Page 5, Figures 3–4; page 8, Figure 5 bottom-right cell · Main paper

Tables, comparisons and evidence

Table 1 gives an average of 58.1 versus 55.5, a 2.6-point difference, whereas the prose says 2.4%. Improvements are not category-wise best: LPT scores higher on MMStar and R1-Onevision slightly higher on MathVision.

Page 7, Table 1 and first results paragraph · Main paper

Appendix and reproducibility

The 240K-to-86K-to-55K funnel and complete three-stage prompts clarify supervision absent from the headline. Ground-truth answers appear in thinking, contrasting and rethinking; this is not label-free SSL despite the contrast terminology.

Appendix A.2–A.3, pages 16–18 · Main paper

Clarity and writing conventions

Question-style ablation headings make the design discussion easy to scan. However, the easy-data conclusion should not be called monotonic: Hallusion improves from 52.2 at +20K to 55.7 at +40K, although both remain below 56.3 without additions.

Section 4.3; Table 3, page 9 · Main paper

What to reuse—and what not to infer

Transferable lesson: Separate data construction from deployment in both diagrams and prose, and audit every qualitative comparison cell as carefully as numeric tables.

Boundary: Comparative prompting is not an InfoNCE objective or proof of faithful reasoning. Supervision, external rewriting, source-dataset mixture and filtering complicate causal attribution; no pathology validation is shown.

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

Pages 1–10, Sections 1–5; references 10–16

Appendix / supplement

Pages 16–19, A.1–A.4 cognitive interpretation, dataset construction, all prompts and qualitative examples

Visually inspected PDF pages

5, 8

• official ICLR 2026 proceedings PDF · 19 pages