Writing study: WAVE: Learning Unified & Versatile Audio-Visual Embeddings with Multimodal LLM

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Changli Tang, Qinfan Xiao, Ke Mei, Tianyi Wang, Fengyun Rao, Chao 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: Architecture & representation design (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

17 physical pages

Before references / appendix

≈ 10.11 page-equivalents

Main figures / tables

1 / 7

Attached appendix figures / tables

1 / 3

Attached appendix span

≈ 2.7 page-equivalents

Before references / appendix: ≈ 59.47%References / intervening material: ≈ 24.65%Attached appendix: ≈ 15.88%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	≈ 1.01 pp.	5.94%
2 BACKGROUND	2	≈ 0.96 pp.	5.65%
3 METHODS	3	≈ 2.12 pp.	12.47%
4 EXPERIMENTAL SETTINGS	5	≈ 1.73 pp.	10.18%
5 EXPERIMENTAL RESULTS	7	≈ 2.8 pp.	16.47%
6 CONCLUSION	10	≈ 0.18 pp.	1.06%
7 REPRODUCIBILITY STATEMENT	10	≈ 0.77 pp.	4.53%
REFERENCES	11	≈ 4.2 pp.	24.71%
A THE USE OF LARGE LANGUAGE MODELS	15	≈ 0.07 pp.	0.41%
B INFERENCE PROCEDURE FOR EVALUATION TASKS	15	≈ 0.34 pp.	2%
C MORE EVALUATION RESULTS OF WAVE	15	≈ 0.16 pp.	0.94%
D CASE STUDY OF PROMPT-AWARE EMBEDDINGS	15	≈ 0.93 pp.	5.47%
E ANALYSIS OF DUAL SPEECH & AUDIO ENCODERS	16	≈ 0.58 pp.	3.41%
F THE EFFECT OF IMAGE TRAINING	17	≈ 0.63 pp.	3.71%

All detected figure and table locations

Main text

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

Attached appendix

• Table 8 · p. 16
• Figure 2 · p. 16
• Table 9 · p. 17
• Table 10 · p. 17

Argument and organization

1. Capability gap and background

   Distinguishes traditional modality encoders from LLM-based embeddings, then motivates dynamic audio/video inputs and instruction-sensitive embeddings.

2. Architecture, objectives, protocols

   Separates the multimodal token path, retrieval/QA losses and training/evaluation specifications so implementation and task definitions are not conflated.

3. Capabilities followed by controlled explanations

   Reports broad retrieval results before testing prompts, joint training and feature fusion as explanations for those results.

Problem, claim and contribution

The contribution list separates representation coverage, prompt awareness and architectural choices. Each becomes a visible experimental question, making a multi-part multimodal system easier to evaluate than a single broad claim of universality.

Section 1; Sections 5.1–5.4 · Main paper

Figures, captions and visual explanation

Figure 1 uses a consistent color legend for visual, speech, audio and text tokens. Stacked layer outputs and an explicit text-only branch explain the different pooling rules. The diagram also states that non-text inputs receive prompts, a condition essential to interpreting the model.

Figure 1, page 3; Section 3.1 · Main paper

The appendix heatmap annotates every cosine similarity and prints all four prompts beside it. Holding the video fixed lets the visual isolate a semantic-focus change, although one illustrative video is not statistical evidence of general instruction following.

Appendix D, Figure 2, page 16 · Main paper

Tables, comparisons and evidence

The evaluation table specifies query/target modalities and metrics separately from the training-data table. The common-prompt versus question-specific-prompt comparison directly tests the instruction-awareness claim; the reverse-retrieval appendix exposes a Clotho result below its reference baseline.

Tables 1–2 and 5; Appendix C, Table 8 · Main paper

Appendix and reproducibility

The inference appendix turns QA into a four-step embedding-retrieval protocol, avoiding ambiguity with free-form answer generation. Separate sections then cover audio-encoder and image-data controls. This is a useful pattern for papers whose task names conceal different evaluation procedures.

Appendices B, E and F · Main paper

Clarity and writing conventions

The tables support narrower statements than some captions: joint training improves seven of eight task columns rather than every task. Also, the overall WAVE score is 59.9 in Table 3 but 59.0 in Table 6. A consistency audit should reconcile such values and distinguish percentage-point gaps from relative percentages.

Tables 3, 5 and 6; Section 5.3 · Main paper

What to reuse—and what not to infer

Transferable lesson: For a unified model, explicitly enumerate supported input combinations, training pair types and evaluation query/target directions. Pair each capability claim with a controlled ablation and a small, interpretable example.

Boundary: Embedding-based multiple-choice QA is not unrestricted text generation. Broad universality language exceeds the tested datasets, and the paper's tables contain an unresolved overall-score discrepancy. Avoid copying the background statement that SigLIP removes negatives; it removes the global softmax normalization, not the need to distinguish matched from unmatched pairs.

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

All main sections 1–7, pages 1–10, including architecture, two losses, training/evaluation specifications and ablations.

Appendix / supplement

Appendices A–F, pages 15–17: disclosure, inference protocol, reverse retrieval, prompt case study, audio-encoder and image-data ablations.

Visually inspected PDF pages

3, 16

• Official ICLR 2026 proceedings manuscript with appendices A–F · 17 pages