# Vision-centric Visual Question Answering on MRAG-Bench (test)

Source: [Vision-centric Visual Question Answering on MRAG-Bench (test)](<https://www.sota2.com/research/sota/vision-centric-visual-question-answering-on-mrag-bench-test>)

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

Data updated: 2026-02-26T05:34:54.909Z

Evaluation of Vision-centric Visual Question Answering performance on the test set of the MRAG-Bench dataset, reporting metrics related to robustness factors such as Angle, Occlusion, Temporal consistency, and Deformation.

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 | Vision-centric Visual Question Answering |
| Dataset (including any stated split) | MRAG-Bench (test) |
| Metric | Angle Robustness Score |
| Unit | PERCENT |
| Direction | Higher is better |
| Results loaded | 12 of 12 |
| 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

- [Angle Robustness Score](<https://www.sota2.com/research/sota/vision-centric-visual-question-answering-on-mrag-bench-test.md?metric=Angle%20Robustness%20Score>): Higher is better; unit: PERCENT.

- [Partial Occlusion Robustness Score](<https://www.sota2.com/research/sota/vision-centric-visual-question-answering-on-mrag-bench-test.md?metric=Partial%20Occlusion%20Robustness%20Score>): Higher is better; unit: PERCENT.

- [Scope Robustness Score](<https://www.sota2.com/research/sota/vision-centric-visual-question-answering-on-mrag-bench-test.md?metric=Scope%20Robustness%20Score>): Higher is better; unit: PERCENT.

- [Occlusion Robustness Score](<https://www.sota2.com/research/sota/vision-centric-visual-question-answering-on-mrag-bench-test.md?metric=Occlusion%20Robustness%20Score>): Higher is better; unit: PERCENT.

- [Temporal Consistency Score](<https://www.sota2.com/research/sota/vision-centric-visual-question-answering-on-mrag-bench-test.md?metric=Temporal%20Consistency%20Score>): Higher is better; unit: PERCENT.

- [Deformation Robustness Score](<https://www.sota2.com/research/sota/vision-centric-visual-question-answering-on-mrag-bench-test.md?metric=Deformation%20Robustness%20Score>): Higher is better; unit: PERCENT.

- [Incomplete Information Robustness Score](<https://www.sota2.com/research/sota/vision-centric-visual-question-answering-on-mrag-bench-test.md?metric=Incomplete%20Information%20Robustness%20Score>): Higher is better; unit: PERCENT.

- [Biological Context Robustness Score](<https://www.sota2.com/research/sota/vision-centric-visual-question-answering-on-mrag-bench-test.md?metric=Biological%20Context%20Robustness%20Score>): Higher is better; unit: PERCENT.

- [Other Robustness Factors Score](<https://www.sota2.com/research/sota/vision-centric-visual-question-answering-on-mrag-bench-test.md?metric=Other%20Robustness%20Factors%20Score>): Higher is better; unit: PERCENT.

- [Overall VQA Score](<https://www.sota2.com/research/sota/vision-centric-visual-question-answering-on-mrag-bench-test.md?metric=Overall%20VQA%20Score>): Higher is better; unit: PERCENT.

## Recorded results

| Method | Angle Robustness Score | Result date | Report paper |
| --- | --- | --- | --- |
| Qwen2.5-VL-7B-Instruct + R3G (Ours) | 67.7 | Not recorded | 69818125a8b01e76cc8e22e2 |
| Qwen2.5-VL-7B-Instruct + MRAG | 61.18 | Not recorded | 69818125a8b01e76cc8e22e2 |
| LLaVA-OneVision + R3G (Ours) | 54.66 | Not recorded | 69818125a8b01e76cc8e22e2 |
| LLaVA-OneVision + MRAG | 50.93 | Not recorded | 69818125a8b01e76cc8e22e2 |
| Mantis-8B-siglip-llama3 + R3G (Ours) | 48.41 | Not recorded | 69818125a8b01e76cc8e22e2 |
| LLaVA-NeXT-Interleave-7B + R3G (Ours) | 46.58 | Not recorded | 69818125a8b01e76cc8e22e2 |
| Mantis-8B-siglip-llama3 + MRAG | 42.55 | Not recorded | 69818125a8b01e76cc8e22e2 |
| Mantis-8B-clip-llama3 + R3G (Ours) | 40.33 | Not recorded | 69818125a8b01e76cc8e22e2 |
| LLaVA-NeXT-Interleave-7B + MRAG | 40.06 | Not recorded | 69818125a8b01e76cc8e22e2 |
| Deepseek-VL-7B-chat + R3G (Ours) | 39.44 | Not recorded | 69818125a8b01e76cc8e22e2 |
| Mantis-8B-clip-llama3 + MRAG | 36.65 | Not recorded | 69818125a8b01e76cc8e22e2 |
| Deepseek-VL-7B-chat + MRAG | 33.54 | Not recorded | 69818125a8b01e76cc8e22e2 |

### Qwen2.5-VL-7B-Instruct + R3G (Ours)

Angle Robustness Score: 67.7. Result ID: 6997909b5698665ec4e5f732.

[Reporting paper](<https://www.sota2.com/research/paper/69818125a8b01e76cc8e22e2.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69818125a8b01e76cc8e22e2.md>)

Recorded evaluation configuration: {"Backbone":"Qwen2.5-VL-7B-Instruct","Framework":"R3G","Retriever":"EVA-CLIP"}. This record does not assert independent reproduction.

### Qwen2.5-VL-7B-Instruct + MRAG

Angle Robustness Score: 61.18. Result ID: 6997909b5698665ec4e5f731.

[Reporting paper](<https://www.sota2.com/research/paper/69818125a8b01e76cc8e22e2.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa77871e209131cea73dac.md>)

Recorded evaluation configuration: {"Backbone":"Qwen2.5-VL-7B-Instruct","Framework":"MRAG","Retriever":"EVA-CLIP"}. This record does not assert independent reproduction.

### LLaVA-OneVision + R3G (Ours)

Angle Robustness Score: 54.66. Result ID: 6997909b5698665ec4e5f730.

[Reporting paper](<https://www.sota2.com/research/paper/69818125a8b01e76cc8e22e2.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69818125a8b01e76cc8e22e2.md>)

Recorded evaluation configuration: {"Backbone":"LLaVA-OneVision","Framework":"R3G","Retriever":"EVA-CLIP"}. This record does not assert independent reproduction.

### LLaVA-OneVision + MRAG

Angle Robustness Score: 50.93. Result ID: 6997909b5698665ec4e5f72f.

[Reporting paper](<https://www.sota2.com/research/paper/69818125a8b01e76cc8e22e2.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa77871e209131cea73dac.md>)

Recorded evaluation configuration: {"Backbone":"LLaVA-OneVision","Framework":"MRAG","Retriever":"EVA-CLIP"}. This record does not assert independent reproduction.

### Mantis-8B-siglip-llama3 + R3G (Ours)

Angle Robustness Score: 48.41. Result ID: 6997909b5698665ec4e5f72a.

[Reporting paper](<https://www.sota2.com/research/paper/69818125a8b01e76cc8e22e2.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69818125a8b01e76cc8e22e2.md>)

Recorded evaluation configuration: {"Backbone":"Mantis-8B-siglip-llama3","Framework":"R3G","Retriever":"EVA-CLIP"}. This record does not assert independent reproduction.

### LLaVA-NeXT-Interleave-7B + R3G (Ours)

Angle Robustness Score: 46.58. Result ID: 6997909b5698665ec4e5f72e.

[Reporting paper](<https://www.sota2.com/research/paper/69818125a8b01e76cc8e22e2.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69818125a8b01e76cc8e22e2.md>)

Recorded evaluation configuration: {"Backbone":"LLaVA-NeXT-Interleave-7B","Framework":"R3G","Retriever":"EVA-CLIP"}. This record does not assert independent reproduction.

### Mantis-8B-siglip-llama3 + MRAG

Angle Robustness Score: 42.55. Result ID: 6997909b5698665ec4e5f729.

[Reporting paper](<https://www.sota2.com/research/paper/69818125a8b01e76cc8e22e2.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa77871e209131cea73dac.md>)

Recorded evaluation configuration: {"Backbone":"Mantis-8B-siglip-llama3","Framework":"MRAG","Retriever":"EVA-CLIP"}. This record does not assert independent reproduction.

### Mantis-8B-clip-llama3 + R3G (Ours)

Angle Robustness Score: 40.33. Result ID: 6997909b5698665ec4e5f728.

[Reporting paper](<https://www.sota2.com/research/paper/69818125a8b01e76cc8e22e2.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69818125a8b01e76cc8e22e2.md>)

Recorded evaluation configuration: {"Backbone":"Mantis-8B-clip-llama3","Framework":"R3G","Retriever":"EVA-CLIP"}. This record does not assert independent reproduction.

### LLaVA-NeXT-Interleave-7B + MRAG

Angle Robustness Score: 40.06. Result ID: 6997909b5698665ec4e5f72d.

[Reporting paper](<https://www.sota2.com/research/paper/69818125a8b01e76cc8e22e2.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa77871e209131cea73dac.md>)

Recorded evaluation configuration: {"Backbone":"LLaVA-NeXT-Interleave-7B","Framework":"MRAG","Retriever":"EVA-CLIP"}. This record does not assert independent reproduction.

### Deepseek-VL-7B-chat + R3G (Ours)

Angle Robustness Score: 39.44. Result ID: 6997909b5698665ec4e5f72c.

[Reporting paper](<https://www.sota2.com/research/paper/69818125a8b01e76cc8e22e2.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69818125a8b01e76cc8e22e2.md>)

Recorded evaluation configuration: {"Backbone":"Deepseek-VL-7B-chat","Framework":"R3G","Retriever":"EVA-CLIP"}. This record does not assert independent reproduction.

### Mantis-8B-clip-llama3 + MRAG

Angle Robustness Score: 36.65. Result ID: 6997909b5698665ec4e5f727.

[Reporting paper](<https://www.sota2.com/research/paper/69818125a8b01e76cc8e22e2.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa77871e209131cea73dac.md>)

Recorded evaluation configuration: {"Backbone":"Mantis-8B-clip-llama3","Framework":"MRAG","Retriever":"EVA-CLIP"}. This record does not assert independent reproduction.

### Deepseek-VL-7B-chat + MRAG

Angle Robustness Score: 33.54. Result ID: 6997909b5698665ec4e5f72b.

[Reporting paper](<https://www.sota2.com/research/paper/69818125a8b01e76cc8e22e2.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa77871e209131cea73dac.md>)

Recorded evaluation configuration: {"Backbone":"Deepseek-VL-7B-chat","Framework":"MRAG","Retriever":"EVA-CLIP"}. This record does not assert independent reproduction.

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