Writing study: EmotionThinker: Prosody-Aware Reinforcement Learning for Explainable Speech Emotion Reasoning

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Dingdong WANG, Shujie LIU, Tianhua Zhang, Youjun Chen, Jinyu Li, Helen Meng · 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

26 physical pages

Before references / appendix

≈ 9.2 page-equivalents

Main figures / tables

4 / 7

Attached appendix figures / tables

5 / 1

Attached appendix span

≈ 12.82 page-equivalents

Before references / appendix: ≈ 35.38%References / intervening material: ≈ 15.31%Attached appendix: ≈ 49.31%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	≈ 1.16 pp.	4.46%
2 RELATED WORK	2	≈ 0.97 pp.	3.73%
3 METHOD	3	≈ 2.96 pp.	11.38%
4 EXPERIMENT	6	≈ 3.03 pp.	11.65%
5 CONCLUSION	9	≈ 0.31 pp.	1.19%
6 ACKNOWLEDGEMENT	10	≈ 0.09 pp.	0.35%
REFERENCES	10	≈ 3.98 pp.	15.31%
Supplementary Material	14	≈ 1 pp.	3.85%
A THE USE OF LARGE LANGUAGE MODELS	15	≈ 0.14 pp.	0.54%
B EMOTIONCOT-35K	15	≈ 1.14 pp.	4.38%
C EMOTIONTHINKER-BASE	16	≈ 1.66 pp.	6.38%
D REWARD MODEL	18	≈ 0.34 pp.	1.31%
E EMOTION REASONING CRITERIA	18	≈ 0.11 pp.	0.42%
F CASE STUDY	18	≈ 8.44 pp.	32.46%

All detected figure and table locations

Main text

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

Attached appendix

• Figure 5 · p. 17
• Figure 7 · p. 18
• Figure 6 · p. 19
• Table 8 · p. 20
• Figure 8 · p. 25
• Figure 9 · p. 26

Argument and organization

1. Three obstacles to emotional explanation

   Maps dataset quality, perception and reward weaknesses onto three contributions.

2. Prosody supervision and progressive rewards

   Separates preparatory SFT from policy optimization.

3. Recognition and explanation evaluation

   Uses distinct accuracy and rubric dimensions rather than one success measure.

Problem, claim and contribution

The three-challenge introduction provides a direct outline for dataset construction, prosody-centric SFT and reward design. However, plausible explanations of an emotion label are not necessarily causal explanations of the model's actual decision.

Section 1, pp.1–2 · Main paper

Argument and section order

The reward equation makes correctness, formatting and explanation separable. Its exponential trust factor only attenuates a conflicting reward; contrary to the gate metaphor, it never switches that reward off. The case of an all-correct or all-wrong group is unspecified.

Section 3.3.2, p.6 · Main paper

Figures, captions and visual explanation

Figure 3 combines a training diagram with an enlarged reward inset. Figure 4 color-codes acoustic evidence and contradictions in matched outputs, making a correct final label with an inconsistent explanation visibly different from successful reasoning.

Figures 3–4, pp.5, 8 · Main paper

Tables, comparisons and evidence

Automated ratings receive a human check, but that check covers 100 outputs total, only 20 per model. Table 6's K=8 result differs from the stated main setting, while Table 4 labels V5 with-progressive although the prose describes without-progressive.

Tables 3–7 and Sections 4.2–4.3, pp.7–9 · Main paper

Appendix and reproducibility

Full synthesis and scoring prompts expose an important dependency: generated rationales are given the target emotion, and GPT-4o participates in both dataset generation and evaluation. Appendix C also clarifies the second, LoRA-only cold-start SFT stage.

Appendices B–E, pp.15–19, 24–26 · Main paper

Clarity and writing conventions

The extended example is not wholly consistent with the main illustration: Table 8's proposed-model rationale ends in frustration or irritation rather than the displayed sad answer, and one baseline quotes another transcript. Such mismatches weaken exemplary-evidence claims.

Figure 4 and Table 8, pp.8, 20–23 · Main paper

What to reuse—and what not to infer

Transferable lesson: Tie each proposed component to a named obstacle, expose annotation and judge prompts, and check qualitative examples as carefully as benchmark tables.

Boundary: Treat emotion and speaker-attribute labels as dataset judgments, not definitive private mental states. Synthetic demographic cues, judge dependence and sparse human assessment limit interpretability and bias claims.

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 1–6; references 10–13

Appendix / supplement

Pages 14–26, supplementary contents and Appendices A–F, including all prompts and case-study outputs

Visually inspected PDF pages

5, 8

• official ICLR 2026 proceedings PDF · 26 pages