Image Classification on ImageNet-R (OOD Evaluation)
79.5AccuracyDBD
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| DBDAttack=PGD-100, Epsilon=4/255, Backbone=CLIP-ViT-B/162026.06 | 79.5 | 99.4 | |
| CLIPAttack=PGD-100, Epsilon=4/255, Backbone=CLIP-ViT-B/162026.06 | 77.1 | 0 | |
| R-TPTAttack=PGD-100, Epsilon=4/255, Backbone=CLIP-ViT-B/162026.06 | 77 | 57.6 | |
| DBDAttack=PGD-100, Epsilon=4/255, Backbone=CLIP-ViT-B/322026.06 | 71.7 | 98.2 | |
| R-TPTAttack=PGD-100, Epsilon=4/255, Backbone=CLIP-ViT-B/322026.06 | 70 | 47.9 | |
| CLIPAttack=PGD-100, Epsilon=4/255, Backbone=CLIP-ViT-B/322026.06 | 68.3 | 0 | |
| TTCAttack=PGD-100, Epsilon=4/255, Backbone=CLIP-ViT-B/162026.06 | 63.3 | 12.4 | |
| DBDAttack=PGD-10, Epsilon=1/255, Backbone=CLIP-ResNet502026.06 | 63 | 94 | |
| SFT-EMModel=Qwen2.5-VL-3B2026.02 | 61.74 | 59.44 | |
| SFT-MModel=Qwen2.5-VL-3B2026.02 | 60.14 | 59.02 | |
| CLIPAttack=PGD-10, Epsilon=1/255, Backbone=CLIP-ResNet502026.06 | 60 | 0.4 | |
| R-TPTAttack=PGD-10, Epsilon=1/255, Backbone=CLIP-ResNet502026.06 | 57.7 | 46.6 | |
| GRPOModel=Qwen2.5-VL-3B2026.02 | 57.66 | 58.31 | |
| SFT-EMModel=Qwen2.5-VL-7B2026.02 | 55.9 | 62.1 | |
| SFT-MModel=Qwen2.5-VL-7B2026.02 | 55.79 | 60.41 | |
| SFTModel=Qwen2.5-VL-3B2026.02 | 54.22 | 56.5 | |
| TTCAttack=PGD-100, Epsilon=4/255, Backbone=CLIP-ViT-B/322026.06 | 53.4 | 28.1 | |
| GRPOModel=Qwen2.5-VL-7B2026.02 | 51.38 | 59.48 | |
| SFTModel=Qwen2.5-VL-7B2026.02 | 49.58 | 57.62 | |
| TeCoAAttack=PGD-10, Epsilon=1/255, Backbone=CLIP-ResNet502026.06 | 40.6 | 25.7 | |
| TeCoAAttack=PGD-100, Epsilon=4/255, Backbone=CLIP-ViT-B/322026.06 | 38 | 12.7 |