Image Classification on Waterbirds 95% correlation (test)
92.5Worst-group AccuracyBAP (LP)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| BAP (LP)Backbone=ConvNeXT-W, Pre-training=CLIP, Evaluation Protocol=Linear Probing2026.05 | 92.5 | 93.6 | |
| DFRBackbone=ConvNeXT-W, Pre-training=CLIP, Evaluation Protocol=Trained Evaluation2026.05 | 91.3 | 94.1 | |
| DFRGroup Info=true, Train Twice=true2025.08 | 89.7 | 93.6 | |
| UWGroup Info=true, Train Twice=false2025.08 | 89.3 | 94.5 | |
| AFRBackbone=ConvNeXT-W, Pre-training=CLIP, Evaluation Protocol=Trained Evaluation2026.05 | 88.9 | 92.9 | |
| GroupDROGroup Info=true, Train Twice=false2025.08 | 87.2 | 93.2 | |
| CnCGroup Info=false, Train Twice=true2025.08 | 86.5 | 91 | |
| JTTGroup Info=false, Train Twice=true2025.08 | 86.4 | 89.5 | |
| Data-Matched ControlBackbone=ConvNeXT-W, Pre-training=CLIP, Evaluation Protocol=Trained Evaluation2026.05 | 85.8 | 92.5 | |
| SupERGroup Info=false, Train Twice=false2025.08 | 84.4 | 87.3 | |
| LfFGroup Info=false, Train Twice=false2025.08 | 79.1 | 91.9 | |
| IRMGroup Info=true, Train Twice=false2025.08 | 76.2 | 89.4 | |
| GALSGroup Info=false, Train Twice=false2025.08 | 75.4 | 89 | |
| ERM (LP-FT)Backbone=ConvNeXT-W, Pre-training=CLIP, Evaluation Protocol=Linear Probing & Fine-Tuning2026.05 | 73.9 | 87.5 | |
| CVaR DROGroup Info=false, Train Twice=false2025.08 | 73.1 | 90.7 | |
| BAP (ZS)Backbone=ConvNeXT-W, Pre-training=CLIP, Evaluation Protocol=Zero-Shot2026.05 | 71.6 | 86.4 | |
| WiSE-FTBackbone=ConvNeXT-W, Pre-training=CLIP, Evaluation Protocol=Trained Evaluation2026.05 | 65.8 | 88.4 | |
| ERMGroup Info=false, Train Twice=false2025.08 | 64.9 | 90.7 | |
| DIALBackbone=ConvNeXT-W, Pre-training=CLIP, Evaluation Protocol=Zero-Shot2026.05 | 64.9 | 79.7 | |
| RoboShotBackbone=ConvNeXT-W, Pre-training=CLIP, Evaluation Protocol=Zero-Shot2026.05 | 59.2 | 75.8 | |
| PruSCBackbone=ConvNeXT-W, Pre-training=CLIP, Evaluation Protocol=Trained Evaluation2026.05 | 57.5 | 75.7 | |
| Native Backbone (LP)Backbone=ConvNeXT-W, Pre-training=CLIP, Evaluation Protocol=Linear Probing2026.05 | 57.1 | 81.5 | |
| Native BackboneBackbone=ConvNeXT-W, Pre-training=CLIP, Evaluation Protocol=Zero-Shot2026.05 | 47.2 | 75.3 |