Writing study: Omni-Reward: Towards Generalist Omni-Modal Reward Modeling with Free-Form Preferences

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Zhuoran Jin, Hongbang Yuan, Kejian Zhu, Jiachun Li, Pengfei Cao, Yubo Chen, Kang Liu, Jun Zhao · 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 · Architecture & representation design.

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

50 physical pages

Before references / appendix

≈ 11.11 page-equivalents

Main figures / tables

3 / 3

Attached appendix figures / tables

14 / 17

Attached appendix span

≈ 28.89 page-equivalents

Before references / appendix: ≈ 22.22%References / intervening material: ≈ 20%Attached appendix: ≈ 57.78%Estimated layout envelopes, not text/artwork area. Unresolved appendix boundaries remain within the undivided post-body material.

Source-bound correction: Omni-Reward: Figure 3 is the real task-correlation heatmap caption on physical page 9. Its label shares a Poppler line with unrelated body text, while the caption body is emitted separately. Visually checked page 9 and restored the caption.

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	≈ 1.88 pp.	3.76%
2 OMNI-REWARDBENCH	3	≈ 1.94 pp.	3.88%
3 OMNI-REWARDMODEL	5	≈ 1.38 pp.	2.76%
4 EXPERIMENTS	6	≈ 2.39 pp.	4.78%
5 ANALYSIS	9	≈ 0.47 pp.	0.94%
6 RELATED WORK	9	≈ 1.09 pp.	2.18%
7 CONCLUSION	10	≈ 0.4 pp.	0.8%
ETHICS STATEMENT	11	≈ 0.2 pp.	0.4%
REPRODUCIBILITY STATEMENT	11	≈ 0.69 pp.	1.38%
REFERENCES	12	≈ 10 pp.	20%
A LLM USAGE STATEMENT	22	≈ 0.12 pp.	0.24%
B LIMITATIONS	22	≈ 0.2 pp.	0.4%
C BROADER IMPACTS	22	≈ 0.68 pp.	1.36%
D ANNOTATION DETAILS	23	≈ 2 pp.	4%
E ETHICS AND QUALITY CONTROL	25	≈ 1 pp.	2%
F DATASET STATISTICS	26	≈ 3 pp.	6%
G IMPLEMENTATION DETAILS	29	≈ 1 pp.	2%
H ADDITIONAL EXPERIMENTAL RESULTS	30	≈ 0.66 pp.	1.32%
I ADDITIONAL ANALYSIS	30	≈ 2.34 pp.	4.68%
J DATA EXAMPLES	33	≈ 9 pp.	18%
K PROMPT TEMPLATES	42	≈ 8.89 pp.	17.78%

All detected figure and table locations

Main text

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

Attached appendix

• Figure 4 · p. 23
• Figure 5 · p. 24
• Table 4 · p. 26
• Table 5 · p. 27
• Table 6 · p. 27
• Table 7 · p. 28
• Figure 6 · p. 30
• Figure 7 · p. 30
• Table 8 · p. 31
• Table 9 · p. 31
• Figure 8 · p. 32
• Table 10 · p. 32
• Table 11 · p. 32
• Figure 9 · p. 33
• Figure 10 · p. 34
• Figure 11 · p. 35
• Figure 12 · p. 36
• Figure 13 · p. 37
• Figure 14 · p. 38
• Figure 15 · p. 39
• Figure 16 · p. 40
• Figure 17 · p. 41
• Table 12 · p. 42
• Table 13 · p. 43
• Table 14 · p. 44
• Table 15 · p. 45
• Table 16 · p. 46
• Table 17 · p. 47
• Table 18 · p. 48
• Table 19 · p. 49
• Table 20 · p. 50

Argument and organization

1. Evaluation, data, model

   Two practical limitations motivate three named deliverables that are reused as major section headings.

2. Shared reward objective, distinct model families

   Discriminative Bradley–Terry training and generative GRPO are separated before evaluation and modality-mixture analysis.

Problem, claim and contribution

Modality imbalance and fixed implicit preferences are introduced as separate obstacles; free-form criteria are part of the formal sample tuple, making the second obstacle operational.

Sections 1–2.1, pages 1–3 · Main paper

Argument and section order

Data-mixture and instruction-data ablations are distinct from the cross-task correlation analysis. This separates an intervention from an associative observation.

Table 3 and Figure 3, page 9 · Main paper

Figures, captions and visual explanation

Figure 2 contrasts scalar and generative rewards using aligned horizontal decoder diagrams, frozen-encoder symbols and example inputs. The differing supervision paths are immediately visible.

Figure 2, page 6 · Main paper

Tables, comparisons and evidence

The main table retains nine task columns and unsupported-modality dashes, which reveal why one pooled ranking is inadequate.

Table 1, page 7; Appendix G, page 29 · Main paper

Appendix and reproducibility

Annotation guidelines, rejection rates, author-annotator disclosure, source counts and full task-specific prompts support reuse. The single-turn and limited-modalities boundaries appear explicitly.

Appendices B–G and K, pages 22–29 and 42–50 · Main paper

Clarity and writing conventions

The evaluation/data/model triad creates memorable navigation, but 'generalist' needs repeated qualification: the generative variant lacks audio support and aggregate means use only supported tasks.

Sections 3.2–3.3, Table 1 and Appendix G · Main paper

What to reuse—and what not to infer

Transferable lesson: Use the same small set of conceptual deliverables in the introduction, headings, diagrams and experimental questions; expose missing capabilities alongside scores.

Boundary: Tie-inclusive scores for discriminative models use a maximized tie threshold, while model-wide means cover different supported modalities. These are not interchangeable deployment-calibrated all-modality rankings.

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 1–11: sections 1–7, ethics and reproducibility statements.

Appendix / supplement

Pages 22–50, A–K: limitations, annotation workflow/ethics, dataset statistics, training settings, extra results/analyses, task examples and prompts. References pages 12–21 excluded from substantive reading claim.

Visually inspected PDF pages

6, 9

• ICLR 2026 final proceedings · 50 pages