Image Classification on ImageNet-100 (test)
93.82Clean AccuracySS-CA
Evaluation results
| Method | Links | |||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| SS-CAModel=ViT2025.11 | 93.82 | — | — | — | — | — | — | — | — | — | — | — | — | |
| Xiao et al.Model=ViT2025.11 | 93.55 | — | — | — | — | — | — | — | — | — | — | — | — | |
| Chen et al.Model=ViT2025.11 | 93.48 | — | — | — | — | — | — | — | — | — | — | — | — | |
| Conventional TrainingModel=ViT2025.11 | 93.45 | — | — | — | — | — | — | — | — | — | — | — | — | |
| SS-CAModel=ResNet-1012025.11 | 91.95 | — | — | — | — | — | — | — | — | — | — | — | — | |
| FoCA CBMBackbone=ResNet2026.06 | 91.88 | — | — | — | — | — | — | — | — | — | — | 57.3 | 1.862 | |
| Xiao et al.Model=ResNet-1012025.11 | 91.45 | — | — | — | — | — | — | — | — | — | — | — | — | |
| Chen et al.Model=ResNet-1012025.11 | 91.42 | — | — | — | — | — | — | — | — | — | — | — | — | |
| Conventional TrainingModel=ResNet-1012025.11 | 91.33 | — | — | — | — | — | — | — | — | — | — | — | — | |
| SS-CAModel=CLIP (ViT /32b)2025.11 | 91.14 | — | — | — | — | — | — | — | — | — | — | — | — | |
| RENGBackbone=ViT-B16, Evaluation Protocol=Fine-tuning, LoRA rank (r)=4, LoRA alpha (α)=42026.01 | 90.1 | — | — | — | — | — | — | — | — | — | — | — | — | |
| Chen et al.Model=CLIP (ViT /32b)2025.11 | 89.83 | — | — | — | — | — | — | — | — | — | — | — | — | |
| R-KalmanBackbone=ViT-B16, Evaluation Protocol=Fine-tuning, LoRA rank (r)=4, LoRA alpha (α)=42026.01 | 89.8 | — | — | — | — | — | — | — | — | — | — | — | — | |
| Xiao et al.Model=CLIP (ViT /32b)2025.11 | 89.77 | — | — | — | — | — | — | — | — | — | — | — | — | |
| NGDBackbone=ViT-B16, Evaluation Protocol=Fine-tuning, LoRA rank (r)=4, LoRA alpha (α)=42026.01 | 89.6 | — | — | — | — | — | — | — | — | — | — | — | — | |
| Conventional TrainingModel=CLIP (ViT /32b)2025.11 | 89.5 | — | — | — | — | — | — | — | — | — | — | — | — | |
| AdamWBackbone=ViT-B16, Evaluation Protocol=Fine-tuning, LoRA rank (r)=4, LoRA alpha (α)=42026.01 | 89 | — | — | — | — | — | — | — | — | — | — | — | — | |
| RINGBackbone=ViT-B16, Evaluation Protocol=Fine-tuning, LoRA rank (r)=4, LoRA alpha (α)=42026.01 | 89 | — | — | — | — | — | — | — | — | — | — | — | — | |
| AGF-guided Adaptive RouterRoute Ratio (Full:Small)=48.5 : 51.5, Est. Cost=0.92x2026.03 | 88.78 | — | — | — | — | 4.46 | — | — | — | — | — | — | — | |
| Static FullRoute Ratio (Full:Small)=100 : 0, Est. Cost=1.00x2026.03 | 88.74 | — | — | — | — | 4.42 | — | — | — | — | — | — | — | |
| AugMixBackbone=ResNet-18, JSD Consistency Loss=true2021.12 | 88.7 | 60.7 | 73.1 | — | — | — | — | — | — | — | — | — | — | |
| DREAM-ID2023.09 | 88.46 | — | — | — | — | — | — | — | — | — | — | — | — | |
| FoCA CBM-NBackbone=ResNet2026.06 | 88.36 | — | — | — | — | — | — | — | — | — | — | 66.5 | 2.15 | |
| Vanilla CBMBackbone=ResNet2026.06 | 88.27 | — | — | — | — | — | — | — | — | — | — | 66.2 | 2.197 | |
| FAN-S-AdaNCAPlacement (after)=52024.06 | 88.22 | — | — | — | — | — | — | — | 47.88 | 42.24 | — | — | — | |
| SophiaBackbone=ViT-B16, Evaluation Protocol=Fine-tuning, LoRA rank (r)=4, LoRA alpha (α)=42026.01 | 88.2 | — | — | — | — | — | — | — | — | — | — | — | — | |
| RandAugment2023.09 | 88.1 | — | — | — | — | — | — | — | — | — | — | — | — | |
| StandardBackbone=ResNet-182021.12 | 88 | 49.7 | 100 | — | — | — | — | — | — | — | — | — | — | |
| AutoAugment2023.09 | 88 | — | — | — | — | — | — | — | — | — | — | — | — | |
| MEMO2023.09 | 88 | — | — | — | — | — | — | — | — | — | — | — | — | |
| CutMix2023.09 | 87.98 | — | — | — | — | — | — | — | — | — | — | — | — | |
| AugMix2023.09 | 87.74 | — | — | — | — | — | — | — | — | — | — | — | — | |
| Generic Prompts2023.09 | 87.74 | — | — | — | — | — | — | — | — | — | — | — | — | |
| RVT-S-AdaNCAPlacement (after)=92024.06 | 87.5 | — | — | — | — | — | — | — | 44.71 | 39.12 | — | — | — | |
| FAN-S2024.06 | 87.3 | — | — | — | — | — | — | — | 33.72 | 29.44 | — | — | — | |
| CF-CBMBackbone=ResNet2026.06 | 87.3 | — | — | — | — | — | — | — | — | — | — | 67.6 | 2.448 | |
| Original (no aug)data augmentation=none2023.09 | 87.28 | — | — | — | — | — | — | — | — | — | — | — | — | |
| Dynamic Interaction#Params (M)=27.94, FLOPS (G)=4.72024.06 | 87.18 | — | — | — | — | — | — | — | 22.35 | — | — | — | — | |
| Swin-T-AdaNCAPlacement (after)=42024.06 | 87.18 | — | — | — | — | — | — | — | 36.41 | 31.74 | — | — | — | |
| RVT-S2024.06 | 87.18 | — | — | — | — | — | — | — | 38.01 | 33.14 | — | — | — | |
| MLPCBMBackbone=ResNet2026.06 | 86.88 | — | — | — | — | — | — | — | — | — | — | 65.9 | 2.21 | |
| DeepAugment2023.09 | 86.86 | — | — | — | — | — | — | — | — | — | — | — | — | |
| ViTCAscale=1, #Params (M)=28.48, FLOPS (G)=4.82024.06 | 86.78 | — | — | — | — | — | — | — | 19.47 | — | — | — | — | |
| ViTCAscale=2, #Params (M)=29.08, FLOPS (G)=4.92024.06 | 86.72 | — | — | — | — | — | — | — | 22.28 | — | — | — | — | |
| Swin-T2024.06 | 86.56 | — | — | — | — | — | — | — | 18.69 | 16.18 | — | — | — | |
| CEMBackbone=ResNet2026.06 | 86.51 | — | — | — | — | — | — | — | — | — | — | 67.8 | 2.178 | |
| DA+AugMixBackbone=ResNet-18, JSD Consistency Loss=true2021.12 | 86.5 | 73.1 | 57.3 | — | — | — | — | — | — | — | — | — | — | |
| TA (Wide)Total # Augmentations=5, Hard-coded # Augmentations=4, Backbone=ResNet182023.06 | 86.39 | — | — | — | — | — | — | — | — | — | — | — | — | |
| LFCBMBackbone=ResNet2026.06 | 86.32 | — | — | — | — | — | — | — | — | — | — | 67.6 | 2.448 | |
| DeepAugment (DA)Backbone=ResNet-182021.12 | 86.3 | 67.7 | 68.1 | — | — | — | — | — | — | — | — | — | — | |
| SLACKTotal # Augmentations=3, Hard-coded # Augmentations=0, Backbone=ResNet182023.06 | 86.06 | — | — | — | — | — | — | — | — | — | — | — | — | |
| SCBMBackbone=ResNet2026.06 | 85.97 | — | — | — | — | — | — | — | — | — | — | 66.2 | 2.049 | |
| PRIMEBackbone=ResNet-18, JSD Consistency Loss=true2021.12 | 85.9 | 71.6 | 61 | — | — | — | — | — | — | — | — | — | — | |
| TA (RA)Total # Augmentations=5, Hard-coded # Augmentations=4, Backbone=ResNet182023.06 | 85.87 | — | — | — | — | — | — | — | — | — | — | — | — | |
| Uniform policyTotal # Augmentations=3, Hard-coded # Augmentations=0, Backbone=ResNet182023.06 | 85.78 | — | — | — | — | — | — | — | — | — | — | — | — | |
| ProbCBMBackbone=ResNet2026.06 | 85.75 | — | — | — | — | — | — | — | — | — | — | 69.3 | 2.269 | |
| DA+PRIMEBackbone=ResNet-18, JSD Consistency Loss=true2021.12 | 84.9 | 74.9 | 54.6 | — | — | — | — | — | — | — | — | — | — | |
| Random PolicyRoute Ratio (Full:Small)=50 : 50, Est. Cost=0.92x2026.03 | 84.32 | — | — | — | — | — | — | — | — | — | — | — | — | |
| Full DatasetImg/Cls=10/202022.12 | 82 | — | — | — | — | — | — | — | — | — | — | — | — | |
| FullIPC (Ratio)=10 (0.8%), Test Model=ResNet-182026.06 | 81.8 | — | — | — | — | — | — | — | — | — | — | — | — | |
| FullIPC (Ratio)=20 (1.6%), Test Model=ResNet-182026.06 | 81.8 | — | — | — | — | — | — | — | — | — | — | — | — | |
| FullIPC (Ratio)=10 (0.8%), Test Model=ResNetAP-102026.06 | 80.3 | — | — | — | — | — | — | — | — | — | — | — | — | |
| FullIPC (Ratio)=20 (1.6%), Test Model=ResNetAP-102026.06 | 80.3 | — | — | — | — | — | — | — | — | — | — | — | — | |
| FullIPC (Ratio)=10 (0.8%), Test Model=ConvNet-62026.06 | 79.9 | — | — | — | — | — | — | — | — | — | — | — | — | |
| FullIPC (Ratio)=20 (1.6%), Test Model=ConvNet-62026.06 | 79.9 | — | — | — | — | — | — | — | — | — | — | — | — | |
| Static PrunedRoute Ratio (Full:Small)=0 : 100, Est. Cost=0.85x2026.03 | 79.8 | — | — | — | — | -4.52 | — | — | — | — | — | — | — | |
| HybridCBMBackbone=ResNet2026.06 | 79.51 | — | — | — | — | — | — | — | — | — | — | 67.6 | 2.448 | |
| ZENITHBackbone=DenseNet1212026.01 | 78.2 | — | — | 4.84 | — | — | — | — | — | — | — | — | — | |
| ProdigyBackbone=DenseNet1212026.01 | 78 | — | — | 5.63 | — | — | — | — | — | — | — | — | — | |
| DAdaptBackbone=DenseNet1212026.01 | 76.9 | — | — | 5.59 | — | — | — | — | — | — | — | — | — | |
| SPSBackbone=DenseNet1212026.01 | 76.6 | — | — | 1.88 | — | — | — | — | — | — | — | — | — | |
| DoGBackbone=DenseNet1212026.01 | 76.4 | — | — | 6.7 | — | — | — | — | — | — | — | — | — | |
| DoWGBackbone=DenseNet1212026.01 | 76.1 | — | — | 7.21 | — | — | — | — | — | — | — | — | — | |
| ALIGBackbone=DenseNet1212026.01 | 75.1 | — | — | 1.75 | — | — | — | — | — | — | — | — | — | |
| LaBoBackbone=ResNet2026.06 | 74.48 | — | — | — | — | — | — | — | — | — | — | 67.6 | 2.448 | |
| PALBackbone=DenseNet1212026.01 | 72.3 | — | — | 3.12 | — | — | — | — | — | — | — | — | — | |
| Non-private (SGD)Privacy=Non-private2026.02 | 71.14 | — | — | — | — | — | — | — | — | — | — | — | — | |
| COCOBBackbone=DenseNet1212026.01 | 68.1 | — | — | 2.83 | — | — | — | — | — | — | — | — | — | |
| L4Backbone=DenseNet1212026.01 | 68.1 | — | — | 4.69 | — | — | — | — | — | — | — | — | — | |
| Federated Averagingdataset skew (s)=100, synchronization steps (u)=200, training steps=60k2022.11 | 68 | — | — | — | — | — | — | — | — | — | — | — | — | |
| Posthoc CBMBackbone=ResNet2026.06 | 67.25 | — | — | — | — | — | — | — | — | — | — | 82 | 2.53 | |
| Non-privateDirichlet alpha (α)=12026.02 | 66.3 | — | — | — | — | — | — | — | — | — | — | — | — | |
| Federated Averagingdataset skew (s)=100, synchronization steps (u)=1000, training steps=60k2022.11 | 65.7 | — | — | — | — | — | — | — | — | — | — | — | — | |
| WW-DP-SGDPrivacy=Differential Privacy2026.02 | 65.5 | — | — | — | — | — | — | — | — | — | — | — | — | |
| Non-privateDirichlet alpha (α)=0.52026.02 | 64.74 | — | — | — | — | — | — | — | — | — | — | — | — | |
| DP-Adam-ACPrivacy=Differential Privacy2026.02 | 64.24 | — | — | — | — | — | — | — | — | — | — | — | — | |
| ADP-AdamWPrivacy=Differential Privacy2026.02 | 63.95 | — | — | — | — | — | — | — | — | — | — | — | — | |
| DP-SAMPrivacy=Differential Privacy2026.02 | 63.84 | — | — | — | — | — | — | — | — | — | — | — | — | |
| DP-ISPrivacy=Differential Privacy2026.02 | 63.69 | — | — | — | — | — | — | — | — | — | — | — | — | |
| MHD+dataset skew (s)=100, training steps=180k, public dataset scope=entire ImageNet2022.11 | 63.4 | — | — | — | — | — | — | — | — | — | — | — | — | |
| Non-privateDirichlet alpha (α)=0.32026.02 | 62.52 | — | — | — | — | — | — | — | — | — | — | — | — | |
| DP-SGDPrivacy=Differential Privacy2026.02 | 62.52 | — | — | — | — | — | — | — | — | — | — | — | — | |
| DP-SATPrivacy=Differential Privacy2026.02 | 62.14 | — | — | — | — | — | — | — | — | — | — | — | — | |
| DP-AdamPrivacy=Differential Privacy2026.02 | 61.36 | — | — | — | — | — | — | — | — | — | — | — | — | |
| LSPPretraining Prior=ImageNet†2026.03 | 59.44 | — | — | — | — | — | — | — | — | — | — | — | — | |
| REMPretraining Prior=ImageNet†2026.03 | 59.12 | — | — | — | — | — | — | — | — | — | — | — | — | |
| WW-DP-SGDDirichlet alpha (α)=12026.02 | 59.04 | — | — | — | — | — | — | — | — | — | — | — | — | |
| TUEPretraining Prior=ImageNet†2026.03 | 59 | — | — | — | — | — | — | — | — | — | — | — | — | |
| Non-privateDirichlet alpha (α)=0.12026.02 | 58 | — | — | — | — | — | — | — | — | — | — | — | — | |
| DP-Adam-ACDirichlet alpha (α)=12026.02 | 57.74 | — | — | — | — | — | — | — | — | — | — | — | — |