O.O.D. Classification on MetaShift Cat vs. Dog
92.4AccuracyOpenCLIP ZS (full image)
Evaluation Results
| Method | Links | |
|---|---|---|
| OpenCLIP ZS (full image)Backbone=ViT-L-14, Requires Group Labels=false2026.06 | 92.4 | |
| A^2_ZSBackbone (attn × emb)=ViT-S × OpenCLIP ViT-L-14, Requires Group Labels=false2026.06 | 89.7 | |
| TTRBackbone=OpenCLIP ViT-L-14, Requires Group Labels=false2026.06 | 88.1 | |
| iFAM (K=8)Backbone=ViT-B (86M), Requires Group Labels=false2026.06 | 81.1 | |
| LR (full image)Backbone=ViT-G (1.1B), Requires Group Labels=false2026.06 | 78.6 | |
| iFAM (K=4)Backbone=ViT-B (86M), Requires Group Labels=false2026.06 | 78.3 | |
| A^2_LR cross-modelBackbone (attn × emb)=ViT-S × ViT-G, Requires Group Labels=false2026.06 | 77.3 | |
| DFRBackbone=ViT-G (1.1B), Requires Group Labels=true2026.06 | 76 | |
| A^2_LRBackbone (attn × emb)=ViT-S × ViT-S, Requires Group Labels=false2026.06 | 69.8 | |
| DFR† + A^2Backbone (attn × emb)=ViT-S × ViT-S, Requires Group Labels=true2026.06 | 69.4 | |
| DFRBackbone=ViT-B (86M), Requires Group Labels=true2026.06 | 68.6 | |
| LR (full image)Backbone=ViT-S (21M), Requires Group Labels=false2026.06 | 59.8 | |
| DFRBackbone=ViT-S (21M), Requires Group Labels=true2026.06 | 58.7 |