Image Classification on CelebA
90.3WG ScoreSPARE**
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| SPARE**Annot. Tr/Val=✗/✓2024.07 | 90.3 | 91.1 | — | — | — | |
| Contrastive AdapterBackbone=CLIP RN-50, Evaluation Protocol=Contrastive Adaptation2022.07 | 90 | 90.7 | 0.7 | — | — | |
| Contrastive AdapterBackbone=CLIP ResNet-502022.07 | 90 | 90.7 | 0.7 | — | — | |
| PruSCAnnot. Tr/Val=✗/✗2024.07 | 89.7 | 91 | — | — | — | |
| DFR (Up)Backbone=CLIP RN-50, Evaluation Protocol=Upsample DFR2022.07 | 89.6 | 91.8 | 2.2 | — | — | |
| DFR (Upsample)Backbone=CLIP ResNet-502022.07 | 89.6 | 91.8 | 2.2 | — | — | |
| DeDiER**Annot. Tr/Val=✗/✓2024.07 | 89.6 | 93.2 | — | — | — | |
| LISAGroup Info (train/val)=✓/✓2026.03 | 89.3 | 92.4 | — | — | — | |
| HSFMGroup Info (train/val)=✗/✓2026.03 | 89.2 | 90.6 | — | — | — | |
| Group DROGroup Info (train/val)=✓/✓2026.03 | 88.9 | 92.9 | — | — | — | |
| CnC**Annot. Tr/Val=✗/✓2024.07 | 88.9 | 93.9 | — | — | — | |
| GroupDROAnnot. Tr/Val=✓/✓2024.07 | 88.7 | 93 | — | — | — | |
| DFRGroup Info (train/val)=✗/✓✓2026.03 | 88.3 | 91.3 | — | — | — | |
| DFRAnnot. Tr/Val=✓/✓2024.07 | 88.3 | 91.3 | — | — | — | |
| SliceLensRepair Protocol=Discriminative Fine-tuning for Repair (DFR) [18]2025.12 | 88.2 | — | — | — | — | |
| LadderRepair Protocol=Native ensembling procedure2025.12 | 87.8 | — | — | — | — | |
| DDBGroup Info (train/val)=✗/✓2026.03 | 85.8 | 87.3 | — | — | — | |
| WiSE-FTBackbone=CLIP RN-50, Evaluation Protocol=Fine-tuning2022.07 | 85.6 | 88.6 | 3 | — | — | |
| WISE-FTBackbone=CLIP ResNet-502022.07 | 85.6 | 88.6 | 3 | — | — | |
| Contrastive AdapterBackbone=CLIP ViT-L/142022.07 | 84.6 | 90.4 | 5.8 | — | — | |
| SELF**6Annot. Tr/Val=✗/✓2024.07 | 83.9 | — | — | — | — | |
| DATBackbone=AltCLIP2026.06 | 83.89 | 86.64 | 2.75 | — | — | |
| DAT*Backbone=AltCLIP2026.06 | 83.89 | 86.68 | 2.79 | — | — | |
| DFR (Upsample)Backbone=CLIP ViT-L/142022.07 | 83.7 | 91.2 | 7.5 | — | — | |
| ROBOSHOTBackbone=ALIGN2026.06 | 83.4 | 86.3 | 2.9 | — | — | |
| DATBackbone=ALIGN2026.06 | 82.76 | 84.82 | 2.06 | — | — | |
| DominoRepair Protocol=Discriminative Fine-tuning for Repair (DFR) [18]2025.12 | 82.7 | — | — | — | — | |
| DAT*Backbone=ALIGN2026.06 | 82.69 | 84.8 | 2.11 | — | — | |
| DaCGroup Info (train/val)=✗/✓2026.03 | 81.9 | 91.4 | — | — | — | |
| JTTGroup Info (train/val)=✗/✓2026.03 | 81.1 | 88 | — | — | — | |
| WISE-FTBackbone=CLIP ViT-L/142022.07 | 80 | 87.4 | 7.4 | — | — | |
| ZSBackbone=AltCLIP2026.06 | 79.7 | 82.3 | 2.6 | — | — | |
| Group PromptBackbone=AltCLIP2026.06 | 79 | 82.3 | 3.3 | — | — | |
| ZSBackbone=ALIGN2026.06 | 77.2 | 81.8 | 4.6 | — | — | |
| ROBOSHOTBackbone=AltCLIP2026.06 | 77.2 | 86 | 8.8 | — | — | |
| DFR (Sub)Backbone=CLIP RN-50, Evaluation Protocol=Subsample DFR2022.07 | 76.9 | 92.5 | 15.6 | — | — | |
| DFR (Subsample)Backbone=CLIP ResNet-502022.07 | 76.9 | 92.5 | 15.6 | — | — | |
| DFR (Subsample)Backbone=CLIP ViT-L/142022.07 | 76.3 | 92.1 | 15.8 | — | — | |
| Zero-ShotBackbone=CLIP RN-50, Evaluation Protocol=Zero-Shot2022.07 | 74 | 81.9 | 7.9 | — | — | |
| Zero-shot (ZS)Backbone=CLIP ResNet-502022.07 | 74 | 81.9 | 7.9 | — | — | |
| DCWPAnnot. Tr/Val=✗/✗2024.07 | 73.2 | 88.9 | — | — | — | |
| Group Prompt ZSBackbone=CLIP ViT-L/142022.07 | 72.4 | 81.8 | 9.4 | — | — | |
| Group PromptBackbone=CLIP RN-50, Evaluation Protocol=Prompting2022.07 | 70.8 | 82.6 | 11.8 | — | — | |
| Group Prompt ZSBackbone=CLIP ResNet-502022.07 | 70.8 | 82.6 | 11.8 | — | — | |
| DFRclassAnnot. Tr/Val=✗/✗2024.07 | 67.7 | 84.9 | — | — | — | |
| Group PromptBackbone=ALIGN2026.06 | 67.4 | 78.3 | 10.9 | — | — | |
| Zero-shot (ZS)Backbone=CLIP ViT-L/142022.07 | 62.1 | 71.9 | 9.8 | — | — | |
| FACTSRepair Protocol=Discriminative Fine-tuning for Repair (DFR) [18]2025.12 | 60 | — | — | — | — | |
| Reweight OnlyModel type=Post-hoc CBM, Feedback=Automated, Component=Label and Reweight step2025.09 | 53.3 | 73.9 | — | — | — | |
| CBDebugModel type=Post-hoc CBM, Feedback=Automated2025.09 | 51.3 | 68.7 | — | — | — | |
| ERM (baseline)Annot. Tr/Val=✗/✗2024.07 | 49.7 | 90.1 | — | — | — | |
| Augment OnlyModel type=Post-hoc CBM, Feedback=Automated, Component=Augment step2025.09 | 38.9 | 71.5 | — | — | — | |
| ERM AdapterBackbone=CLIP ViT-L/142022.07 | 36.7 | 94.2 | 57.5 | — | — | |
| ERM AdapterBackbone=CLIP RN-50, Evaluation Protocol=Adapter2022.07 | 36.1 | 94.2 | 58.1 | — | — | |
| ERM AdapterBackbone=CLIP ResNet-502022.07 | 36.1 | 94.2 | 58.1 | — | — | |
| Base (ERM)Group Info (train/val)=✗/✗2026.03 | 30.6 | 95.83 | — | — | — | |
| ERM Linear ProbeBackbone=CLIP ViT-L/142022.07 | 28.3 | 94.7 | 66.4 | — | — | |
| RetrainModel type=Post-hoc CBM, Feedback=Automated2025.09 | 22.2 | 79.9 | — | — | — | |
| ERM LPBackbone=CLIP RN-50, Evaluation Protocol=Linear Probing2022.07 | 11.9 | 94.7 | 82.8 | — | — | |
| ERM Linear ProbeBackbone=CLIP ResNet-502022.07 | 11.9 | 94.7 | 82.8 | — | — | |
| OriginalModel type=Post-hoc CBM, Feedback=Automated2025.09 | 8.7 | 76.2 | — | — | — | |
| RemoveModel type=Post-hoc CBM, Feedback=Automated2025.09 | 6.5 | 19.9 | — | — | — | |
| CBMDevice=CPU, Measurement=Relative to CBM2025.06 | — | — | — | 1 | 1 | |
| CGMCDDevice=CPU, Measurement=Relative to CBM2025.06 | — | — | — | 2.209 | 82.5712 | |
| CGMpriorDevice=CPU, Measurement=Relative to CBM2025.06 | — | — | — | 2.227 | 18.6351 | |
| CREAMDevice=CPU, Measurement=Relative to CBM2025.06 | — | — | — | 1.963 | 1.122 |