Image Classification on CIFAR-100 (test) (Standard Metrics)
89.5Top-1 Accuracy+DnA (A_{h}^{Q^{\pm}V^{\pm}})
Evaluation Results
| Method | Links | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| +DnA (A_{h}^{Q^{\pm}V^{\pm}})setup=transfer learning, epochs=1002026.06 | 89.5 | — | — | — | — | — | 98.1 | — | — | — | — | — | |
| ViT-Bsetup=transfer learning, epochs=1002026.06 | 88.8 | — | — | — | — | — | 98 | — | — | — | — | — | |
| VS2++Backbone=ViT-B/16, Steering Setting=RAG-enhanced (oracle unlabeled external data)2025.06 | 85.4 | — | — | — | — | — | — | — | — | — | — | — | |
| PEFT Ensembleε=8, Label strategy=Base Dataset (Baseline)2024.11 | 85 | — | — | — | — | — | — | — | — | — | — | — | |
| PEFT Ensembleε=8, Label strategy=Public Labels (Ours)2024.11 | 84.9 | — | — | — | — | — | — | — | — | — | — | — | |
| PEFT Ensembleε=8, Label strategy=Label Oracle (Baseline)2024.11 | 84.4 | — | — | — | — | — | — | — | — | — | — | — | |
| PEFT Ensembleε=8, Label strategy=Release Labels (Ours)2024.11 | 84.4 | — | — | — | — | — | — | — | — | — | — | — | |
| CLIP Steering VectorBackbone=ViT-B/16, Steering Setting=RAG-enhanced (oracle unlabeled external data)2025.06 | 84.12 | — | — | — | — | — | — | — | — | — | — | — | |
| PATTeacher model=ConvNeXt-T, Student model=ResMLP-S12, Student architecture category=MLP-based, Average of 3 independent trials=true2026.06 | 83.5 | — | — | — | — | — | — | — | — | — | — | — | |
| SPOFATeacher model=Swin-T, Student model=ResNet18, Student architecture category=CNN-based, Average of 3 independent trials=true2026.06 | 82.39 | — | — | — | — | — | — | — | — | — | — | — | |
| SPOFATeacher model=ConvNeXt-T, Student model=ResMLP-S12, Student architecture category=MLP-based, Average of 3 independent trials=true2026.06 | 82.32 | — | — | — | — | — | — | — | — | — | — | — | |
| asymmetric knowledge distillation frameworkDistortion type=Gaussian blur with kernel size 37 and σ = 9 (GB), Evaluation protocol=Transfer learning2026.04 | 82.19 | — | — | — | — | — | — | — | — | — | — | — | |
| VS2++Backbone=ViT-B/32, Steering Setting=RAG-enhanced (oracle unlabeled external data)2025.06 | 81.95 | — | — | — | — | — | — | — | — | — | — | — | |
| SPOFATeacher model=Mixer-B/16, Student model=ResNet18, Student architecture category=CNN-based, Average of 3 independent trials=true2026.06 | 81.89 | — | — | — | — | — | — | — | — | — | — | — | |
| CLIP Steering VectorBackbone=ViT-B/32, Steering Setting=RAG-enhanced (oracle unlabeled external data)2025.06 | 81.85 | — | — | — | — | — | — | — | — | — | — | — | |
| HRPNoise mode=Sym., Noise ratio=20%2026.05 | 81.6 | — | — | — | — | — | — | — | — | — | — | — | |
| SPOFATeacher model=ViT-S, Student model=ResNet18, Student architecture category=CNN-based, Average of 3 independent trials=true2026.06 | 81.26 | — | — | — | — | — | — | — | — | — | — | — | |
| SPOFATeacher model=Swin-T, Student model=MobileNetV2, Student architecture category=CNN-based, Average of 3 independent trials=true2026.06 | 81.25 | — | — | — | — | — | — | — | — | — | — | — | |
| PATTeacher model=Swin-T, Student model=ResNet18, Student architecture category=CNN-based, Average of 3 independent trials=true2026.06 | 81.22 | — | — | — | — | — | — | — | — | — | — | — | |
| SPOFATeacher model=ConvNeXt-T, Student model=Swin-P, Student architecture category=ViT-based, Average of 3 independent trials=true2026.06 | 81.17 | — | — | — | — | — | — | — | — | — | — | — | |
| UncertParam. (M)=6.74, T=42026.05 | 81.06 | — | — | — | — | — | — | — | — | — | — | — | |
| QKFormerParam. (M)=6.74, T=42026.05 | 80.98 | — | — | — | — | — | — | — | — | — | — | — | |
| OFATeacher model=Swin-T, Student model=MobileNetV2, Student architecture category=CNN-based, Average of 3 independent trials=true2026.06 | 80.98 | — | — | — | — | — | — | — | — | — | — | — | |
| SPOFATeacher model=Mixer-B/16, Student model=MobileNetV2, Student architecture category=CNN-based, Average of 3 independent trials=true2026.06 | 80.96 | — | — | — | — | — | — | — | — | — | — | — | |
| PATTeacher model=Swin-T, Student model=ResMLP-S12, Student architecture category=MLP-based, Average of 3 independent trials=true2026.06 | 80.94 | — | — | — | — | — | — | — | — | — | — | — | |
| PATTeacher model=ConvNeXt-T, Student model=Swin-P, Student architecture category=ViT-based, Average of 3 independent trials=true2026.06 | 80.74 | — | — | — | — | — | — | — | — | — | — | — | |
| BaselineDistortion type=Gaussian blur with kernel size 37 and σ = 9 (GB), Evaluation protocol=Transfer learning2026.04 | 80.56 | — | — | — | — | — | — | — | — | — | — | — | |
| OFATeacher model=Swin-T, Student model=ResNet18, Student architecture category=CNN-based, Average of 3 independent trials=true2026.06 | 80.54 | — | — | — | — | — | — | — | — | — | — | — | |
| SPOFATeacher model=Mixer-B/16, Student model=Swin-P, Student architecture category=ViT-based, Average of 3 independent trials=true2026.06 | 80.41 | — | — | — | — | — | — | — | — | — | — | — | |
| DKDTeacher model=Swin-T, Student model=ResNet18, Student architecture category=CNN-based, Average of 3 independent trials=true2026.06 | 80.26 | — | — | — | — | — | — | — | — | — | — | — | |
| STAttenT=42026.05 | 80.2 | — | — | — | — | — | — | — | — | — | — | — | |
| OFATeacher model=ViT-S, Student model=ResNet18, Student architecture category=CNN-based, Average of 3 independent trials=true2026.06 | 80.15 | — | — | — | — | — | — | — | — | — | — | — | |
| PATTeacher model=ViT-S, Student model=ResNet18, Student architecture category=CNN-based, Average of 3 independent trials=true2026.06 | 80.11 | — | — | — | — | — | — | — | — | — | — | — | |
| PATTeacher model=Mixer-B/16, Student model=ResNet18, Student architecture category=CNN-based, Average of 3 independent trials=true2026.06 | 80.07 | — | — | — | — | — | — | — | — | — | — | — | |
| SPOFATeacher model=ViT-S, Student model=MobileNetV2, Student architecture category=CNN-based, Average of 3 independent trials=true2026.06 | 80.01 | — | — | — | — | — | — | — | — | — | — | — | |
| PEFT Ensembleε=1, Label strategy=Base Dataset (Baseline)2024.11 | 80 | — | — | — | — | — | — | — | — | — | — | — | |
| PEFT Ensembleε=1, Label strategy=Public Labels (Ours)2024.11 | 80 | — | — | — | — | — | — | — | — | — | — | — | |
| SPOFATeacher model=Swin-T, Student model=ResMLP-S12, Student architecture category=MLP-based, Average of 3 independent trials=true2026.06 | 79.81 | — | — | — | — | — | — | — | — | — | — | — | |
| PATTeacher model=ConvNeXt-T, Student model=DeiT-T, Student architecture category=ViT-based, Average of 3 independent trials=true2026.06 | 79.59 | — | — | — | — | — | — | — | — | — | — | — | |
| SWformerParam. (M)=7.51, T=42026.05 | 79.3 | — | — | — | — | — | — | — | — | — | — | — | |
| VS2++Backbone=ViT-B/16, Steering Setting=RAG-enhanced (non-oracle unlabeled external data)2025.06 | 79.14 | — | — | — | — | — | — | — | — | — | — | — | |
| RRLNoise mode=Sym., Noise ratio=20%2026.05 | 79.1 | — | — | — | — | — | — | — | — | — | — | — | |
| Cosine Similarity Classifierε=8, Label strategy=Label Oracle (Baseline)2024.11 | 79 | — | — | — | — | — | — | — | — | — | — | — | |
| Cosine Similarity Classifierε=8, Label strategy=Release Labels (Ours)2024.11 | 79 | — | — | — | — | — | — | — | — | — | — | — | |
| CLIP Steering VectorBackbone=ViT-B/16, Steering Setting=RAG-enhanced (non-oracle unlabeled external data)2025.06 | 78.97 | — | — | — | — | — | — | — | — | — | — | — | |
| SPOFATeacher model=ConvNeXt-T, Student model=DeiT-T, Student architecture category=ViT-based, Average of 3 independent trials=true2026.06 | 78.91 | — | — | — | — | — | — | — | — | — | — | — | |
| UNICONNoise mode=Sym., Noise ratio=20%2026.05 | 78.9 | — | — | — | — | — | — | — | — | — | — | — | |
| FitNetTeacher model=Swin-T, Student model=ResNet18, Student architecture category=CNN-based, Average of 3 independent trials=true2026.06 | 78.87 | — | — | — | — | — | — | — | — | — | — | — | |
| Cosine Similarity Classifierε=8, Label strategy=Base Dataset (Baseline)2024.11 | 78.8 | — | — | — | — | — | — | — | — | — | — | — | |
| Cosine Similarity Classifierε=8, Label strategy=Public Labels (Ours)2024.11 | 78.8 | — | — | — | — | — | — | — | — | — | — | — | |
| PATTeacher model=Swin-T, Student model=MobileNetV2, Student architecture category=CNN-based, Average of 3 independent trials=true2026.06 | 78.78 | — | — | — | — | — | — | — | — | — | — | — | |
| KDTeacher model=Swin-T, Student model=ResNet18, Student architecture category=CNN-based, Average of 3 independent trials=true2026.06 | 78.74 | — | — | — | — | — | — | — | — | — | — | — | |
| PEFT Ensembleε=1, Label strategy=Label Oracle (Baseline)2024.11 | 78.5 | — | — | — | — | — | — | — | — | — | — | — | |
| Cosine Similarity Classifierε=1, Label strategy=Release Labels (Ours)2024.11 | 78.4 | — | — | — | — | — | — | — | — | — | — | — | |
| S-TransformerParam. (M)=10.28, T=42026.05 | 78.4 | — | — | — | — | — | — | — | — | — | — | — | |
| Cosine Similarity Classifierε=1, Label strategy=Label Oracle (Baseline)2024.11 | 78.3 | — | — | — | — | — | — | — | — | — | — | — | |
| PEFT Ensembleε=1, Label strategy=Release Labels (Ours)2024.11 | 78.3 | — | — | — | — | — | — | — | — | — | — | — | |
| SpikformerParam. (M)=9.32, T=42026.05 | 78.21 | — | — | — | — | — | — | — | — | — | — | — | |
| HRPNoise mode=Sym., Noise ratio=50%2026.05 | 77.9 | — | — | — | — | — | — | — | — | — | — | — | |
| DISTTeacher model=Swin-T, Student model=ResNet18, Student architecture category=CNN-based, Average of 3 independent trials=true2026.06 | 77.75 | — | — | — | — | — | — | — | — | — | — | — | |
| CRDTeacher model=Swin-T, Student model=ResNet18, Student architecture category=CNN-based, Average of 3 independent trials=true2026.06 | 77.63 | — | — | — | — | — | — | — | — | — | — | — | |
| LongReMixNoise mode=Sym., Noise ratio=20%2026.05 | 77.6 | — | — | — | — | — | — | — | — | — | — | — | |
| ProMixNoise mode=Sym., Noise ratio=20%2026.05 | 77.5 | — | — | — | — | — | — | — | — | — | — | — | |
| UNICONNoise mode=Sym., Noise ratio=50%2026.05 | 77.3 | — | — | — | — | — | — | — | — | — | — | — | |
| CLIP Steering VectorBackbone=ViT-B/32, Steering Setting=RAG-enhanced (non-oracle unlabeled external data)2025.06 | 77.22 | — | — | — | — | — | — | — | — | — | — | — | |
| VS2++Backbone=ViT-B/32, Steering Setting=RAG-enhanced (non-oracle unlabeled external data)2025.06 | 77.11 | — | — | — | — | — | — | — | — | — | — | — | |
| SPOFATeacher model=Mixer-B/16, Student model=DeiT-T, Student architecture category=ViT-based, Average of 3 independent trials=true2026.06 | 77.02 | — | — | — | — | — | — | — | — | — | — | — | |
| V-HMNParams (M)=7.122026.03 | 76.58 | — | — | — | — | — | — | — | — | — | — | — | |
| Weighted RAGBackbone=ViT-B/32, Steering Setting=RAG-enhanced (oracle unlabeled external data)2025.06 | 76.43 | — | — | — | — | — | — | — | — | — | — | — | |
| PSSCLNoise mode=Sym., Noise ratio=50%2026.05 | 76.3 | — | — | — | — | — | — | — | — | — | — | — | |
| OFATeacher model=ConvNeXt-T, Student model=DeiT-T, Student architecture category=ViT-based, Average of 3 independent trials=true2026.06 | 75.76 | — | — | — | — | — | — | — | — | — | — | — | |
| asymmetric knowledge distillation frameworkDistortion type=90% random masking (RM), Evaluation protocol=Transfer learning2026.04 | 75.18 | — | — | — | — | — | — | — | — | — | — | — | |
| LongReMixNoise mode=Sym., Noise ratio=50%2026.05 | 75 | — | — | — | — | — | — | — | — | — | — | — | |
| PATTeacher model=Mixer-B/16, Student model=DeiT-T, Student architecture category=ViT-based, Average of 3 independent trials=true2026.06 | 74.66 | — | — | — | — | — | — | — | — | — | — | — | |
| RRLNoise mode=Sym., Noise ratio=50%2026.05 | 74.6 | — | — | — | — | — | — | — | — | — | — | — | |
| asymmetric knowledge distillation frameworkDistortion type=Gaussian noise with σ = 0.5 (GN), Evaluation protocol=Transfer learning2026.04 | 74.58 | — | — | — | — | — | — | — | — | — | — | — | |
| PruneFusePruning ratio (p)=0.5, Selection Metric=Least Conf, Label Budget=50%, Backbone=ResNet-1642026.03 | 74.32 | — | — | — | — | — | — | — | — | — | — | — | |
| RKDTeacher model=Swin-T, Student model=ResNet18, Student architecture category=CNN-based, Average of 3 independent trials=true2026.06 | 74.11 | — | — | — | — | — | — | — | — | — | — | — | |
| DivideMixNoise mode=Sym., Noise ratio=20%2026.05 | 73.8 | — | — | — | — | — | — | — | — | — | — | — | |
| MLP-MixerParams (M)=8.712026.03 | 73.35 | — | — | — | — | — | — | — | — | — | — | — | |
| Baseline ALSelection Metric=Least Conf, Label Budget=50%, Backbone=ResNet-1642026.03 | 73.05 | — | — | — | — | — | — | — | — | — | — | — | |
| AiTParams (M)=7.152026.03 | 72.91 | — | — | — | — | — | — | — | — | — | — | — | |
| ViTParams (M)=7.162026.03 | 72.56 | — | — | — | — | — | — | — | — | — | — | — | |
| Variational Feature CompressionTarget Model=ConvNeXt-V2, Evaluation Model=ConvNeXt-V2, Masking Strategy=integrated dynamic masking2026.04 | 72.23 | — | — | — | — | — | — | — | — | — | — | — | |
| Cosine Similarity Classifierε=1, Label strategy=Public Labels (Ours)2024.11 | 72.2 | — | — | — | — | — | — | — | — | — | — | — | |
| ZSharpNetwork=VGG-16/BN2025.05 | 72.07 | — | — | — | — | — | — | — | — | — | — | — | |
| Cosine Similarity Classifierε=1, Label strategy=Base Dataset (Baseline)2024.11 | 72 | — | — | — | — | — | — | — | — | — | — | — | |
| L2BNoise mode=Sym., Noise ratio=20%2026.05 | 71.8 | — | — | — | — | — | — | — | — | — | — | — | |
| ASAMNetwork=VGG-16/BN2025.05 | 71.7 | — | — | — | — | — | — | — | — | — | — | — | |
| DivideMixNoise mode=Sym., Noise ratio=50%2026.05 | 71.5 | — | — | — | — | — | — | — | — | — | — | — | |
| ProMixNoise mode=Sym., Noise ratio=50%2026.05 | 71.5 | — | — | — | — | — | — | — | — | — | — | — | |
| PruneFusePruning ratio (p)=0.5, Selection Metric=Least Conf, Label Budget=40%, Backbone=ResNet-1642026.03 | 71.45 | — | — | — | — | — | — | — | — | — | — | — | |
| PSSCLNoise mode=Sym., Noise ratio=20%2026.05 | 71.4 | — | — | — | — | — | — | — | — | — | — | — | |
| PruneFusePruning ratio (p)=0.8, Selection Metric=Greedy k, Label Budget=50%, Backbone=ResNet-1642026.03 | 71.29 | — | — | — | — | — | — | — | — | — | — | — | |
| DISTTeacher model=Swin-T, Student model=ResMLP-S12, Student architecture category=MLP-based, Average of 3 independent trials=true2026.06 | 71.05 | — | — | — | — | — | — | — | — | — | — | — | |
| Friendly-SAMNetwork=VGG-16/BN2025.05 | 70.99 | — | — | — | — | — | — | — | — | — | — | — | |
| SAMNetwork=VGG-16/BN2025.05 | 70.92 | — | — | — | — | — | — | — | — | — | — | — | |
| uCBOpt-adaptOptimizer=uCBOpt-adapt2025.11 | 70.5 | — | — | 1.315 | 14.5 | 0.854 | 91.3 | — | — | 0.438 | — | — | |
| Variational Feature CompressionTarget Model=ResNet152, Evaluation Model=ResNet152, Masking Strategy=integrated dynamic masking2026.04 | 70.12 | — | — | — | — | — | — | — | — | — | — | — | |
| uCBOptOptimizer=uCBOpt (ϑ → 0+)2025.11 | 70.1 | — | — | 1.278 | 13.3 | 0.848 | 91.4 | — | — | 0.438 | — | — |