Image Classification on CIFAR-100 (Standard Accuracy)
94.2AccuracyUpper Bound
Evaluation results
| Method | Links | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Upper Bound# Params=86M2026.02 | 94.2 | — | — | — | — | — | — | — | — | — | |
| Share# Params=0.10M2026.02 | 94.2 | — | — | — | — | 0.4 | — | — | — | — | |
| DAP# Params=0.19M2026.02 | 94.05 | — | — | — | — | 0.41 | — | — | — | — | |
| AdamBackbone=ViT-L, Fine-tuning protocol=Full Fine-tuning2026.02 | 91.7 | — | — | — | — | — | — | — | — | — | |
| CrossViT-18Pre-training Dataset=ImageNet-1K, Evaluation Protocol=Fine-tuning2021.03 | 91.36 | — | — | — | — | — | — | — | — | — | |
| AdamBackbone=ViT-B, Fine-tuning protocol=Full Fine-tuning2026.02 | 91.2 | — | — | — | — | — | — | — | — | — | |
| DeiT-SPre-training Dataset=ImageNet-1K, Evaluation Protocol=Fine-tuning2021.03 | 90.89 | — | — | — | — | — | — | — | — | — | |
| DeiT-BPre-training Dataset=ImageNet-1K, Evaluation Protocol=Fine-tuning2021.03 | 90.8 | — | — | — | — | — | — | — | — | — | |
| CrossViT-15Pre-training Dataset=ImageNet-1K, Evaluation Protocol=Fine-tuning2021.03 | 90.77 | — | — | — | — | — | — | — | — | — | |
| LoRABackbone=ViT-L, Fine-tuning protocol=LoRA2026.02 | 90.6 | — | — | — | — | — | — | — | — | — | |
| LaDP-FLPrivacy Budget (epsilon)=0.5, Backbone=ResNet-1522026.01 | 89.4 | — | — | 256,943 | — | — | — | — | — | — | |
| EASE# Params=0.28M2026.02 | 87.81 | — | — | — | — | 0.52 | — | — | — | — | |
| LoRABackbone=ViT-B, Fine-tuning protocol=LoRA2026.02 | 86.4 | — | — | — | — | — | — | — | — | — | |
| CODA-P# Params=0.42M2026.02 | 86.25 | — | — | — | — | 1.16 | — | — | — | — | |
| AdapLDPPrivacy Budget (epsilon)=0.5, Backbone=ResNet-1522026.01 | 86.12 | — | — | 265,376 | — | — | — | — | — | — | |
| DualPrompt# Params=0.35M2026.02 | 84.77 | — | — | — | — | 1.4 | — | — | — | — | |
| DPA LDPPrivacy Budget (epsilon)=0.5, Backbone=ResNet-1522026.01 | 83.57 | — | — | 254,003 | — | — | — | — | — | — | |
| L2P# Params=86M2026.02 | 83.05 | — | — | — | — | 2.12 | — | — | — | — | |
| LaDP-FLPrivacy Budget (epsilon)=0.2, Backbone=ResNet-1522026.01 | 82.44 | — | — | 257,104 | — | — | — | — | — | — | |
| CTSNArchitecture=ResNet19, T=42026.01 | 81.19 | — | — | — | — | — | — | — | — | — | |
| LlipPre-training Data=2.5B, Model Architecture=ViT-B/32, Zero-shot=true2024.12 | 80.8 | — | — | — | — | — | — | — | — | — | |
| OpenCLIPPre-training Data=DataComp-1B, Model Architecture=ViT-B/32, Zero-shot=true2024.12 | 80.4 | — | — | — | — | — | — | — | — | — | |
| SoftMatchBag Size=162024.11 | 80.14 | — | — | — | — | — | — | — | — | — | |
| Fully SupervisedFully Supervised=true2024.11 | 79.89 | — | — | — | — | — | — | — | — | — | |
| CLIFArchitecture=ResNet18, T=42026.01 | 79.69 | — | — | — | — | — | — | — | — | — | |
| Ternary SpikeArchitecture=ResNet19, T=42026.01 | 79.68 | — | — | — | — | — | — | — | — | — | |
| L2P-AHILBag Size=162024.11 | 78.65 | — | — | — | — | — | — | — | — | — | |
| RMP-LossArchitecture=ResNet19, T=42026.01 | 78.28 | — | — | — | — | — | — | — | — | — | |
| ASGLArchitecture=ResNet18, T=42026.01 | 77.74 | — | — | — | — | — | — | — | — | — | |
| MetaCLIPPre-training Data=2.5B, Model Architecture=ViT-B/32, Zero-shot=true2024.12 | 77.7 | — | — | — | — | — | — | — | — | — | |
| Time-Varying DPPrivacy Budget (epsilon)=0.5, Backbone=ResNet-1522026.01 | 77.59 | — | — | 243,434 | — | — | — | — | — | — | |
| GLIFArchitecture=ResNet18, T=62026.01 | 77.35 | — | — | — | — | — | — | — | — | — | |
| L2P-AHILBag Size=322024.11 | 77.3 | — | — | — | — | — | — | — | — | — | |
| LinGeoPEBackbone=Swin-S, Resolution=32x32, Training=from scratch, Window size=42025.12 | 77.2 | — | — | — | 0.1 | — | — | — | — | — | |
| GeoPEBackbone=Swin-S, Resolution=32x32, Training=from scratch, Window size=42025.12 | 76.8 | — | — | — | 0.15 | — | — | — | — | — | |
| L2P-AHILBag Size=642024.11 | 76.52 | — | — | — | — | — | — | — | — | — | |
| LieREBackbone=Swin-S, Resolution=32x32, Training=from scratch, Window size=42025.12 | 76.4 | — | — | — | 0.14 | — | — | — | — | — | |
| STRINGBackbone=Swin-S, Resolution=32x32, Training=from scratch, Window size=42025.12 | 76 | — | — | — | 0.16 | — | — | — | — | — | |
| HexFormer-HybridArchitecture=ViT-Tiny, Pre-training=None2026.01 | 75.85 | — | — | — | — | — | — | — | — | — | |
| OpenCLIPPre-training Data=LAION-2B, Model Architecture=ViT-B/32, Zero-shot=true2024.12 | 75.8 | — | — | — | — | — | — | — | — | — | |
| HexFormerArchitecture=ViT-Tiny, Pre-training=None2026.01 | 75.65 | — | — | — | — | — | — | — | — | — | |
| RoPE-MixedBackbone=Swin-S, Resolution=32x32, Training=from scratch, Window size=42025.12 | 75.6 | — | — | — | 0.18 | — | — | — | — | — | |
| AdapLDPPrivacy Budget (epsilon)=0.2, Backbone=ResNet-1522026.01 | 75.31 | — | — | 263,998 | — | — | — | — | — | — | |
| CTSNArchitecture=ResNet20, T=62026.01 | 75.12 | — | — | — | — | — | — | — | — | — | |
| CPEBackbone=Swin-S, Resolution=32x32, Training=from scratch, Window size=42025.12 | 75.1 | — | — | — | 0.2 | — | — | — | — | — | |
| CTSNArchitecture=ResNet20, T=42026.01 | 74.82 | — | — | — | — | — | — | — | — | — | |
| Sensitive DPPrivacy Budget (epsilon)=0.5, Backbone=ResNet-1522026.01 | 74.81 | — | — | 247,759 | — | — | — | — | — | — | |
| LinGeoPEBackbone=ViT-Base, Resolution=32x32, Training=from scratch, Patch size=4, Epochs=4002025.12 | 74.8 | — | — | — | 0.13 | — | — | — | — | — | |
| Euclidean ViTArchitecture=ViT-Tiny, Pre-training=None2026.01 | 74.64 | — | — | — | — | — | — | — | — | — | |
| RPBBackbone=Swin-S, Resolution=32x32, Training=from scratch, Window size=42025.12 | 74.5 | — | — | — | 0.22 | — | — | — | — | — | |
| TETArchitecture=ResNet19, T=42026.01 | 74.47 | — | — | — | — | — | — | — | — | — | |
| GeoPEBackbone=ViT-Base, Resolution=32x32, Training=from scratch, Patch size=4, Epochs=4002025.12 | 74.4 | — | — | — | 0.15 | — | — | — | — | — | |
| Ternary SpikeArchitecture=ResNet20, T=62026.01 | 74.12 | — | — | — | — | — | — | — | — | — | |
| Nested StrategyCorruption=15% GN2026.02 | 74.1 | — | — | — | — | — | — | — | — | — | |
| LinGeoPEBackbone=ViT-Small, Resolution=32x32, Training=from scratch, Patch size=4, Epochs=4002025.12 | 73.9 | — | — | — | 0.11 | — | — | — | — | — | |
| LieREBackbone=ViT-Base, Resolution=32x32, Training=from scratch, Patch size=4, Epochs=4002025.12 | 73.9 | — | — | — | 0.17 | — | — | — | — | — | |
| LLP-VATBag Size=162024.11 | 73.85 | — | — | — | — | — | — | — | — | — | |
| Nested StrategyCorruption=Gaussian Blur2026.02 | 73.8 | — | — | — | — | — | — | — | — | — | |
| GeoPEBackbone=ViT-Small, Resolution=32x32, Training=from scratch, Patch size=4, Epochs=4002025.12 | 73.5 | — | — | — | 0.12 | — | — | — | — | — | |
| STRINGBackbone=ViT-Base, Resolution=32x32, Training=from scratch, Patch size=4, Epochs=4002025.12 | 73.5 | — | — | — | 0.19 | — | — | — | — | — | |
| LieREBackbone=ViT-Small, Resolution=32x32, Training=from scratch, Patch size=4, Epochs=4002025.12 | 73.1 | — | — | — | 0.15 | — | — | — | — | — | |
| RoPE-MixedBackbone=ViT-Base, Resolution=32x32, Training=from scratch, Patch size=4, Epochs=4002025.12 | 73 | — | — | — | 0.2 | — | — | — | — | — | |
| ROTBag Size=162024.11 | 72.74 | — | — | — | — | — | — | — | — | — | |
| STRINGBackbone=ViT-Small, Resolution=32x32, Training=from scratch, Patch size=4, Epochs=4002025.12 | 72.6 | — | — | — | 0.18 | — | — | — | — | — | |
| ZO-MuonBackbone=ViT-B, Fine-tuning protocol=Full Fine-tuning2026.02 | 72.6 | — | — | — | — | — | — | — | — | — | |
| Mixed Barlow TwinsBackbone=ResNet-50, Epochs=2000, Evaluation Protocol=linear2023.12 | 72.51 | — | — | — | — | — | — | — | — | — | |
| CPEBackbone=ViT-Base, Resolution=32x32, Training=from scratch, Patch size=4, Epochs=4002025.12 | 72.4 | — | — | — | 0.24 | — | — | — | — | — | |
| ZO-MuonBackbone=ViT-L, Fine-tuning protocol=Full Fine-tuning2026.02 | 72.4 | — | — | — | — | — | — | — | — | — | |
| Full DPPrivacy Budget (epsilon)=0.5, Backbone=ResNet-1522026.01 | 72.33 | — | — | 239,407 | — | — | — | — | — | — | |
| RoPE-MixedBackbone=ViT-Small, Resolution=32x32, Training=from scratch, Patch size=4, Epochs=4002025.12 | 72.3 | — | — | — | 0.19 | — | — | — | — | — | |
| L2P-AHILBag Size=1282024.11 | 72.21 | — | — | — | — | — | — | — | — | — | |
| CLIPEvaluation protocol=Zero-shot2024.07 | 72.19 | — | — | — | — | — | — | — | — | — | |
| DPA LDPPrivacy Budget (epsilon)=0.2, Backbone=ResNet-1522026.01 | 72.05 | — | — | 254,731 | — | — | — | — | — | — | |
| Mixed Barlow TwinsBackbone=ResNet-50, Epochs=1000, Evaluation Protocol=linear2023.12 | 71.98 | — | — | — | — | — | — | — | — | — | |
| Nested StrategyCorruption=30% GN2026.02 | 71.9 | — | — | — | — | — | — | — | — | — | |
| CPEBackbone=ViT-Small, Resolution=32x32, Training=from scratch, Patch size=4, Epochs=4002025.12 | 71.8 | — | — | — | 0.21 | — | — | — | — | — | |
| LLP-VATBag Size=322024.11 | 71.62 | — | — | — | — | — | — | — | — | — | |
| IOTAShot=16-shot, Backbone=ViT-B/162026.01 | 71.37 | — | — | — | — | — | — | — | — | — | |
| DLLPBag Size=162024.11 | 71.28 | — | — | — | — | — | — | — | — | — | |
| APEBackbone=ViT-Base, Resolution=32x32, Training=from scratch, Patch size=4, Epochs=4002025.12 | 71.2 | — | — | — | 0.28 | — | — | — | — | — | |
| COSMOSPre-training Data=Merged-30M, Model Architecture=ViT-B/32, Zero-shot=true2024.12 | 70.8 | — | — | — | — | — | — | — | — | — | |
| Parallel StrategyCorruption=15% GN2026.02 | 70.7 | — | — | — | — | — | — | — | — | — | |
| NOVIC (FT9)Evaluation protocol=Zero-shot, Frequency Threshold=92024.07 | 70.65 | — | — | — | — | — | — | — | — | — | |
| APEBackbone=ViT-Small, Resolution=32x32, Training=from scratch, Patch size=4, Epochs=4002025.12 | 70.5 | — | — | — | 0.25 | — | — | — | — | — | |
| Alternating StrategyCorruption=15% GN2026.02 | 70.5 | — | — | — | — | — | — | — | — | — | |
| CentralizedLocation=Client, Split Layer=Layer-52025.12 | 70.49 | — | — | — | — | — | — | — | — | — | |
| Peer LossNoise=Random (0.2)2020.11 | 70.43 | — | — | — | — | — | — | — | — | — | |
| Total Variation divergenceNoise=Random (0.2)2020.11 | 70.4 | — | — | — | — | — | — | — | — | — | |
| Jensen-Shannon divergenceNoise=Random (0.2)2020.11 | 70.28 | — | — | — | — | — | — | — | — | — | |
| CentralizedLocation=Server, Split Layer=Layer-52025.12 | 70.03 | — | — | — | — | — | — | — | — | — | |
| NOVIC (FT6)Evaluation protocol=Zero-shot, Frequency Threshold=62024.07 | 69.93 | — | — | — | — | — | — | — | — | — | |
| DLLPBag Size=322024.11 | 69.92 | — | — | — | — | — | — | — | — | — | |
| Alternating StrategyCorruption=Gaussian Blur2026.02 | 69.7 | — | — | — | — | — | — | — | — | — | |
| BYOLBackbone=ResNet-50, Epochs=1000, Evaluation Protocol=linear2023.12 | 69.59 | — | — | — | — | — | — | — | — | — | |
| NOVIC (FT2)Evaluation protocol=Zero-shot, Frequency Threshold=22024.07 | 69.43 | — | — | — | — | — | — | — | — | — | |
| ROTBag Size=322024.11 | 69.31 | — | — | — | — | — | — | — | — | — | |
| DreamLIPPre-training Data=Merged-30M, Model Architecture=ViT-B/32, Zero-shot=true2024.12 | 69.2 | — | — | — | — | — | — | — | — | — | |
| Total Variation divergenceNoise=Uniform2020.11 | 69.15 | — | — | — | — | — | — | — | — | — | |
| Jensen-Shannon divergenceNoise=Uniform2020.11 | 69.13 | — | — | — | — | — | — | — | — | — | |
| LViTArchitecture=ViT-Base, Pre-training=None2026.01 | 69.11 | — | — | — | — | — | — | — | — | — |