Visual Question Answering on GQA (Adversarial Accuracy)
54.4Clean AccuracyNone
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| NoneCompressed Method=None, Retained Tokens (Kmodel)=144, Backbone=Qwen2.5-VL2026.01 | 54.4 | 35.5 | 35.2 | |
| VisionZIPCompressed Method=VisionZIP, Retained Tokens (Kmodel)=72, Backbone=Qwen2.5-VL2026.01 | 54 | 33.1 | 33 | |
| DivPruneCompressed Method=DivPrune, Retained Tokens (Kmodel)=72, Backbone=Qwen2.5-VL2026.01 | 53.3 | 36.4 | 35.1 | |
| AverageCompressed Method=Average, Retained Tokens (Kmodel)=72, Backbone=Qwen2.5-VL2026.01 | 53.3 | 35.2 | 34.7 | |
| VisPrunerCompressed Method=VisPruner, Retained Tokens (Kmodel)=72, Backbone=Qwen2.5-VL2026.01 | 52.5 | 36.1 | 36.1 | |
| DivPruneCompressed Method=DivPrune, Retained Tokens (Kmodel)=36, Backbone=Qwen2.5-VL2026.01 | 51.1 | 36.5 | 36 | |
| AverageCompressed Method=Average, Retained Tokens (Kmodel)=36, Backbone=Qwen2.5-VL2026.01 | 48.7 | 34.4 | 33.2 | |
| VisionZIPCompressed Method=VisionZIP, Retained Tokens (Kmodel)=36, Backbone=Qwen2.5-VL2026.01 | 47.9 | 32.3 | 30.1 | |
| VisPrunerCompressed Method=VisPruner, Retained Tokens (Kmodel)=36, Backbone=Qwen2.5-VL2026.01 | 47 | 34.4 | 33.4 | |
| NoneCompressed Method=None, Retained Tokens (Kmodel)=0, Backbone=Qwen2.5-VL2026.01 | 28.4 | 28.4 | 28.4 |