# Audio Visual Question Answering on AVQA (Robustness Evaluation)

Source: [Audio Visual Question Answering on AVQA (Robustness Evaluation)](<https://www.sota2.com/research/sota/audio-visual-question-answering-on-avqa-robustness-evaluation>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-02-26T05:26:28.839Z

Evaluation of model performance on the AVQA dataset for Audio Visual Question Answering, focusing on robustness by reporting clean accuracy, adversarial attack accuracy, and the resulting accuracy drop.

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

| Scope | Recorded value |
| --- | --- |
| Task | Audio Visual Question Answering |
| Dataset (including any stated split) | AVQA |
| Metric | AVQA Clean Accuracy |
| Unit | PERCENT |
| Direction | Higher is better |
| Results loaded | 7 of 7 |
| Coverage | All results currently returned for this benchmark |

Evaluation settings may differ between results. The original papers and recorded configurations are needed to establish comparability. A shared metric name alone does not establish an identical protocol.

## Metrics

- [AVQA Clean Accuracy](<https://www.sota2.com/research/sota/audio-visual-question-answering-on-avqa-robustness-evaluation.md?metric=AVQA%20Clean%20Accuracy>): Higher is better; unit: PERCENT.

- [AVQA Attack Accuracy](<https://www.sota2.com/research/sota/audio-visual-question-answering-on-avqa-robustness-evaluation.md?metric=AVQA%20Attack%20Accuracy>): Higher is better; unit: PERCENT.

- [AVQA Accuracy Drop](<https://www.sota2.com/research/sota/audio-visual-question-answering-on-avqa-robustness-evaluation.md?metric=AVQA%20Accuracy%20Drop>): Lower is better; unit: PERCENT.

## Recorded results

| Method | AVQA Clean Accuracy | Result date | Report paper |
| --- | --- | --- | --- |
| Negative Language Modeling Loss | 95.6 | Not recorded | 6976f519a2b28e761d87c22c |
| Encoder-Based Cosine Similarity Loss | 95.6 | Not recorded | 6976f519a2b28e761d87c22c |
| Vision Attention Suppression Loss | 95.6 | Not recorded | 6976f519a2b28e761d87c22c |
| Audio Attention Amplification Loss | 95.6 | Not recorded | 6976f519a2b28e761d87c22c |
| Attention Randomization Loss | 95.6 | Not recorded | 6976f519a2b28e761d87c22c |
| Hidden-State Similarity Loss | 95.6 | Not recorded | 6976f519a2b28e761d87c22c |
| Combined Loss (SOUNDBREAK) | 95.6 | Not recorded | 6976f519a2b28e761d87c22c |

### Negative Language Modeling Loss

AVQA Clean Accuracy: 95.6. Result ID: 699a052b5698665ec4878b75.

[Reporting paper](<https://www.sota2.com/research/paper/6976f519a2b28e761d87c22c.md>)

[Method origin paper](<https://www.sota2.com/research/paper/6976f519a2b28e761d87c22c.md>)

Recorded evaluation configuration: {"Objective":"L\_negLM"}. This record does not assert independent reproduction.

### Encoder-Based Cosine Similarity Loss

AVQA Clean Accuracy: 95.6. Result ID: 699a052b5698665ec4878b76.

[Reporting paper](<https://www.sota2.com/research/paper/6976f519a2b28e761d87c22c.md>)

[Method origin paper](<https://www.sota2.com/research/paper/6976f519a2b28e761d87c22c.md>)

Recorded evaluation configuration: {"Objective":"L^(cos)"}. This record does not assert independent reproduction.

### Vision Attention Suppression Loss

AVQA Clean Accuracy: 95.6. Result ID: 699a052b5698665ec4878b77.

[Reporting paper](<https://www.sota2.com/research/paper/6976f519a2b28e761d87c22c.md>)

[Method origin paper](<https://www.sota2.com/research/paper/6976f519a2b28e761d87c22c.md>)

Recorded evaluation configuration: {"Objective":"L^(visionatt)"}. This record does not assert independent reproduction.

### Audio Attention Amplification Loss

AVQA Clean Accuracy: 95.6. Result ID: 699a052b5698665ec4878b78.

[Reporting paper](<https://www.sota2.com/research/paper/6976f519a2b28e761d87c22c.md>)

[Method origin paper](<https://www.sota2.com/research/paper/6976f519a2b28e761d87c22c.md>)

Recorded evaluation configuration: {"Objective":"L^(audioatt)"}. This record does not assert independent reproduction.

### Attention Randomization Loss

AVQA Clean Accuracy: 95.6. Result ID: 699a052b5698665ec4878b79.

[Reporting paper](<https://www.sota2.com/research/paper/6976f519a2b28e761d87c22c.md>)

[Method origin paper](<https://www.sota2.com/research/paper/6976f519a2b28e761d87c22c.md>)

Recorded evaluation configuration: {"Objective":"L^(randatt)"}. This record does not assert independent reproduction.

### Hidden-State Similarity Loss

AVQA Clean Accuracy: 95.6. Result ID: 699a052b5698665ec4878b7a.

[Reporting paper](<https://www.sota2.com/research/paper/6976f519a2b28e761d87c22c.md>)

[Method origin paper](<https://www.sota2.com/research/paper/6976f519a2b28e761d87c22c.md>)

Recorded evaluation configuration: {"Objective":"L^(hidden-cos)"}. This record does not assert independent reproduction.

### Combined Loss (SOUNDBREAK)

AVQA Clean Accuracy: 95.6. Result ID: 699a052b5698665ec4878b7b.

[Reporting paper](<https://www.sota2.com/research/paper/6976f519a2b28e761d87c22c.md>)

[Method origin paper](<https://www.sota2.com/research/paper/6976f519a2b28e761d87c22c.md>)

Recorded evaluation configuration: {"Objective":"L^(combined)"}. This record does not assert independent reproduction.

Page 1 of 1. This is an index page, not the full collection.