Writing study: Reasoning as Representation: Rethinking Visual Reinforcement Learning in Image Quality Assessment

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Shijie Zhao, Xuanyu Zhang, Weiqi Li, Junlin Li, Li zhang, Tianfan Xue, Jian Zhang · 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: Empirical analysis, diagnosis & auditing (primary) · Learning & optimization algorithms.

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

≈ 9.11 page-equivalents

Main figures / tables

6 / 5

Attached appendix figures / tables

3 / 6

Attached appendix span

≈ 4.89 page-equivalents

Before references / appendix: ≈ 47.95%References / intervening material: ≈ 26.32%Attached appendix: ≈ 25.74%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	≈ 1.55 pp.	8.16%
2 RELATED WORKS	3	≈ 0.34 pp.	1.79%
3 REVISITING REASONING-BASED MLLMS IN IQA	3	≈ 2.24 pp.	11.79%
4 REASONING-ALIGNED LIGHTWEIGHT IQA FRAMEWORK	5	≈ 1.53 pp.	8.05%
5 EXPERIMENTS	7	≈ 2.54 pp.	13.37%
6 CONCLUSION	9	≈ 0.36 pp.	1.89%
REFERENCES	10	≈ 5 pp.	26.32%
APPENDIX	15	≈ 0.04 pp.	0.21%
A THE RELATIONSHIP BETWEEN REASONING AND GENERALIZATION IN	15	≈ 0.54 pp.	2.84%
B LIMITATIONS AND BROADER IMPACTS	15	≈ 0.71 pp.	3.74%
C MORE DETAILS ABOUT REASONING-ALIGNED CROSS-DOMAIN TRAINING	16	≈ 1.14 pp.	6%
D MORE ABLATION STUDIES	17	≈ 0.22 pp.	1.16%
E RELATION TO MODEL LIGHTWEIGHTING	17	≈ 0.75 pp.	3.95%
F VISUALIZATION	18	≈ 0.11 pp.	0.58%
G LLM USAGE STATEMENT	18	≈ 1.39 pp.	7.32%

All detected figure and table locations

Main text

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

Attached appendix

• Table A.1 · p. 15
• Figure A.1 · p. 16
• Table A.2 · p. 16
• Table A.3 · p. 17
• Table A.4 · p. 17
• Table A.5 · p. 18
• Table A.6 · p. 18
• Figure A.2 · p. 19
• Figure A.3 · p. 19

Argument and organization

1. Investigating reasoning-based IQA

   Use attention, representation plots and encoder/LLM recombination to motivate a text bottleneck interpretation.

2. Two designs from the interpretation

   RACT aligns cross-domain training labels; RALI transfers descriptions to a compact encoder and score basis.

Problem, claim and contribution

Two explicit questions—how reasoning relates to generalization and whether reasoning is needed at inference—organize the transition from analysis to a practical lightweight system.

Introduction, page 2. · Main paper

Argument and section order

Component ablations follow the same stage order as the method diagram, and the appendix adds explicit reasoning-disabled training plus a smaller 3B teacher baseline.

Table 4, page 9; Tables A.1 and A.6, pages 15 and 18. · Main paper

Figures, captions and visual explanation

Figure 5 assigns separate columns to the teacher, contrastive alignment, compression and scoring. Frozen/trainable icons and a caption identifying the inference-only column help distinguish offline supervision from deployment.

Figure 5, page 6. · Main paper

Tables, comparisons and evidence

Separate tables for single-domain RALI and multi-domain RACT prevent their training regimes from being conflated. Both report PLCC and SRCC per dataset instead of only an overall score.

Tables 2–3, page 8. · Main paper

Appendix and reproducibility

RACT's label-generation diagram, single-score-source explanation and exact prompts make the cross-domain training recipe easier to reconstruct; the appendix also exposes the basis-count and PCA sweeps.

Appendices C–D, pages 16–18. · Main paper

Clarity and writing conventions

The narrative sometimes upgrades diagnostic evidence into proof: attention concentration and t-SNE overlap support a hypothesis, not causal identification. The introductory >95% memory-reduction claim also conflicts with the reported 14.7% memory ratio at batch 16.

Sections 3.2–3.3, page 4; introduction, page 2; Section 5.4, page 9. · Main paper

What to reuse—and what not to infer

Transferable lesson: A convincing efficiency paper can turn a mechanistic hypothesis into a design whose training and inference paths are visibly different, then evaluate quality and deployment cost separately.

Boundary: Do not generalize an IQA-specific text bottleneck to all visual reasoning. Distinguish comparable average performance from matching every dataset, and use the measured memory ratio rather than the broader introductory claim.

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 1–9: Sections 1–6 and all result/ablation tables.

Appendix / supplement

Pages 15–19: Appendices A–G, reasoning controls, limitations, prompts, hyperparameters and examples.

Visually inspected PDF pages

6, 8

• ICLR 2026 final proceedings · 19 pages