Writing study: Bad Seeing or Bad Thinking? Rewarding Perception for Multimodal Reasoning

ICML overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICML 2026 oral writing corpus

Haozhe Wang, Qixin Xu, Changpeng Wang, Taofeng Xue, Chong Peng, Wenhu Chen, Fangzhen Lin · Official ICML 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

20 physical pages

Before references / appendix

≈ 8.1 page-equivalents

Main figures / tables

4 / 4

Attached appendix figures / tables

0 / 2

Attached appendix span

≈ 7.91 page-equivalents

Before references / appendix: ≈ 40.5%References / intervening material: ≈ 19.95%Attached appendix: ≈ 39.55%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	≈ 1.7 pp.	8.5%
2. Methodology	3	≈ 2.55 pp.	12.75%
3. Experiments	5	≈ 2.72 pp.	13.6%
4. Related Work	8	≈ 0.41 pp.	2.05%
Impact Statement	9	≈ 0.05 pp.	0.25%
References	9	≈ 3.99 pp.	19.95%
A. Training Setup and Implementation Details	13	≈ 0.19 pp.	0.95%
B. Benchmark Categorization and Details	13	≈ 0.4 pp.	2%
C. Modality-Rich Training Dataset	13	≈ 1.41 pp.	7.05%
D. Universal Verbal Program (VP) Prompt	15	≈ 1.73 pp.	8.65%
E. Experiment Details	16	≈ 0.97 pp.	4.85%
F. Examples	17	≈ 3.22 pp.	16.1%

All detected figure and table locations

Main text

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

Attached appendix

• Table 5 · p. 14
• Table 6 · p. 14

Argument and organization

1. 1 Introduction

   Frames failed visual reasoning as ambiguous perception-versus-reasoning credit assignment.

2. 2 Methodology

   Defines tagged actions, a text-only perception proxy, structured outcome verification and reward routing.

3. 3 Experiments; 4 Related Work

   Organizes task performance, component ablations and verifier reliability before contextualizing the method.

Problem, claim and contribution

The title question is converted into two distinct verification problems, giving each proposed reward component a specific diagnostic role.

Introduction and Sections 2.3-2.5, pp.1-5 · Main paper

Argument and section order

The paper defines perception/reasoning actions before the oracle proxy and combines the two reward signals only after explaining each separately.

Sections 2.1-2.5, pp.3-5 · Main paper

Figures, captions and visual explanation

The worked document example uses tagged text blocks and repeated role labels to expose the intended action decomposition; the reward-evaluation page uses a confusion matrix rather than only an overall agreement score.

Figure 2, p.3; Table 3, p.8 · Main paper

Tables, comparisons and evidence

The ablation caption fixes training data and base model, while separate human-majority and repeated-verdict tests address proxy correctness and consistency as different quantities.

Tables 2-4, p.8 · Main paper

Appendix and reproducibility

Training settings, benchmark categories and human annotation instructions are retained separately, including the 979-item sufficiency test and the challenge set’s composition.

Appendices A-C and E, pp.13,17 · Main paper

Clarity and writing conventions

The oracle-fallibility discussion acknowledges false negatives, but later deterministic-reward wording is stronger than the reported 92.3 percent repeated-verdict consistency.

Section 2.3, p.4; Table 4 and Related Work, p.8 · Main paper

What to reuse—and what not to infer

Transferable lesson: When a learning method relies on a judge, evaluate the judge’s error modes separately from downstream task gains and keep terminology consistent with measured reliability.

Boundary: Textual sufficiency is a proxy for perception, not direct access to internal seeing; the prose also mislabels the 58.3 SlideVQA table score as DUDE on p.7, so numerical claims need cross-checking.

Apply this lesson using the argument and evidence worksheet.

Inspection record and sources

Checked

2026-10-07

Reading status

Writing-focused manuscript scan + selected visuals inspected

Main text

Writing-oriented scan of the main argument through physical p.8, with focused physical-page text inspection on 1, 3, 4, 5, 6, 7, 8. Framing, section progression, mechanism/proof explanation, displays, comparison conditions and conclusions were inspected. This is not exhaustive technical reading, proof verification or reproduction.

Appendix / supplement

Appendix organization and main-text pointers were checked; focused text inspection of physical pages 13, 17. Remaining appendix proofs and details were not exhaustively read. No separately linked supplement inspected.

Visually inspected PDF pages

3, 8

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.

• Published proceedings PDF, PMLR volume 306, retrieved 2026-10-07 · 20 pages