# Visual Question Answering on MMA-Bench (Adversarial Robustness)

Source: [Visual Question Answering on MMA-Bench (Adversarial Robustness)](<https://www.sota2.com/research/sota/visual-question-answering-on-mma-bench-adversarial-robustness>)

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

Data updated: 2026-04-07T08:51:18.701Z

Evaluation of model performance on the MMA-Bench dataset for Visual Question Answering, specifically measuring clean accuracy, attack accuracy, and adversarial success rates (ASR).

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 | Visual Question |
| Dataset (including any stated split) | MMA-Bench |
| Metric | Clean Accuracy |
| Unit | PERCENT |
| Direction | Higher is better |
| Results loaded | 6 of 6 |
| 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

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

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

- [ASR (Clean Input)](<https://www.sota2.com/research/sota/visual-question-answering-on-mma-bench-adversarial-robustness.md?metric=ASR%20(Clean%20Input)>): Lower is better; unit: PERCENT.

- [ASR (Adversarial Input)](<https://www.sota2.com/research/sota/visual-question-answering-on-mma-bench-adversarial-robustness.md?metric=ASR%20(Adversarial%20Input)>): Lower is better; unit: PERCENT.

## Recorded results

| Method | Clean Accuracy | Result date | Report paper |
| --- | --- | --- | --- |
| Gemini-2.5-Flash-Lite | 96.79 | Not recorded | 69d48f7b7d0172140ec742a9 |
| Gemini-3.1-Flash-Lite-preview | 96.58 | Not recorded | 69d48f7b7d0172140ec742a9 |
| Qwen3-Omni-30B | 92.88 | Not recorded | 69d48f7b7d0172140ec742a9 |
| Qwen2.5-Omni-7B | 76.68 | Not recorded | 69d48f7b7d0172140ec742a9 |
| ChatBridge | 51.64 | Not recorded | 69d48f7b7d0172140ec742a9 |
| PandaGPT | 28.75 | Not recorded | 69d48f7b7d0172140ec742a9 |

### Gemini-2.5-Flash-Lite

Clean Accuracy: 96.79. Result ID: 69d4c56847f34ce924bcab85.

[Reporting paper](<https://www.sota2.com/research/paper/69d48f7b7d0172140ec742a9.md>)

Recorded evaluation configuration: {"Model":"Gemini-2.5-Flash-Lite"}. This record does not assert independent reproduction.

### Gemini-3.1-Flash-Lite-preview

Clean Accuracy: 96.58. Result ID: 69d4c56847f34ce924bcab86.

[Reporting paper](<https://www.sota2.com/research/paper/69d48f7b7d0172140ec742a9.md>)

Recorded evaluation configuration: {"Model":"Gemini-3.1-Flash-Lite-preview"}. This record does not assert independent reproduction.

### Qwen3-Omni-30B

Clean Accuracy: 92.88. Result ID: 69d4c56847f34ce924bcab82.

[Reporting paper](<https://www.sota2.com/research/paper/69d48f7b7d0172140ec742a9.md>)

Recorded evaluation configuration: {"Model":"Qwen3-Omni-30B"}. This record does not assert independent reproduction.

### Qwen2.5-Omni-7B

Clean Accuracy: 76.68. Result ID: 69d4c56847f34ce924bcab81.

[Reporting paper](<https://www.sota2.com/research/paper/69d48f7b7d0172140ec742a9.md>)

Recorded evaluation configuration: {"Model":"Qwen2.5-Omni-7B"}. This record does not assert independent reproduction.

### ChatBridge

Clean Accuracy: 51.64. Result ID: 69d4c56847f34ce924bcab84.

[Reporting paper](<https://www.sota2.com/research/paper/69d48f7b7d0172140ec742a9.md>)

Recorded evaluation configuration: {"Model":"ChatBridge"}. This record does not assert independent reproduction.

### PandaGPT

Clean Accuracy: 28.75. Result ID: 69d4c56847f34ce924bcab83.

[Reporting paper](<https://www.sota2.com/research/paper/69d48f7b7d0172140ec742a9.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6f7f1e209131cea46009.md>)

Recorded evaluation configuration: {"Model":"PandaGPT"}. This record does not assert independent reproduction.

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