Continual Learning on ImageNet-R 10 tasks
87.18Average ACC@10LoDA + CA
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| LoDA + CAFeature Replay=true2026.02 | 87.18 | — | 82.55 | — | — | — | — | |
| LoDAFeature Replay=false2026.02 | 86.9 | — | 81.93 | — | — | — | — | |
| MACILFeature Replay=true2026.02 | 85.95 | — | 81.88 | — | — | — | — | |
| DIAFeature Replay=true2026.02 | 85.61 | — | 79.03 | — | — | — | — | |
| CoSOFeature Replay=false2026.02 | 85.56 | — | 81.1 | — | — | — | — | |
| LoRA-P&MFeature Replay=false2026.02 | 85.29 | — | 79.95 | — | — | — | — | |
| SSIATFeature Replay=true2026.02 | 83.63 | — | 79.38 | — | — | — | — | |
| SLCAFeature Replay=true2026.02 | 83.29 | — | 79.35 | — | — | — | — | |
| SD-LoRAFeature Replay=false2026.02 | 82.04 | — | 77.34 | — | — | — | — | |
| Bi-LoRAFeature Replay=false2026.02 | 81.52 | — | 77.95 | — | — | — | — | |
| jointBackbone=ViT-B/16, Pre-training=ImageNet-21K2024.03 | 81.14 | 81.14 | — | — | — | — | — | |
| InfLoRABackbone=ViT-B/16, Pre-training=ImageNet-21K, Inserted blocks=All Transformer blocks2024.03 | 80.82 | 75.65 | — | — | — | — | — | |
| InfLoRAFeature Replay=false2026.02 | 80.67 | — | 74.75 | — | — | — | — | |
| PLANFeature Replay=false2026.02 | 80.41 | — | 75.25 | — | — | — | — | |
| VQ-PromptFeature Replay=true2026.02 | 80.02 | — | 75.68 | — | — | — | — | |
| InfLoRA-b5Backbone=ViT-B/16, Pre-training=ImageNet-21K, Inserted blocks=Bottom 5 Transformer blocks2024.03 | 78.54 | 74.13 | — | — | — | — | — | |
| CODAPromptFeature Replay=false2026.02 | 78.47 | — | 73.31 | — | — | — | — | |
| Non CL Baseline (FiLM)Differential Privacy=non-DP, Backbone=ViT-B, Pre-training=ImageNet-21k2024.11 | 78.24 | — | — | — | — | — | — | |
| SEMABackbone=CLIP ViT-B/162024.03 | 78.18 | 80.94 | — | — | — | — | — | |
| Janus-LoRABackbone=SAM2026.05 | 77.9 | 70.22 | — | — | 6.47 | — | — | |
| L2PFeature Replay=false2026.02 | 77.36 | — | 72.34 | — | — | — | — | |
| InfLoRABackbone=SAM2026.05 | 77.21 | 69.64 | — | — | 5.9 | — | — | |
| Zero-shotBackbone=CLIP ViT-B/162024.03 | 77.08 | 79.17 | — | — | — | — | — | |
| LAEBackbone=ViT-B/16, Pre-training=ImageNet-21K2024.03 | 76.71 | 71.7 | — | — | — | — | — | |
| Janus-LoRABackbone=DINO2026.05 | 76.15 | 68.2 | — | — | 5.73 | — | — | |
| InfLoRABackbone=DINO2026.05 | 75.98 | 67.83 | — | — | 7.79 | — | — | |
| CODA-PBackbone=ViT-B/16, Pre-training=ImageNet-21K2024.03 | 75.82 | 71.47 | — | — | — | — | — | |
| LoRA¯DRSBackbone=SAM2026.05 | 75.73 | 68.12 | — | — | 4.78 | — | — | |
| C-LoRABackbone=ViT-B/16, Pre-training=ImageNet-21K2024.03 | 75.33 | 71.89 | — | — | — | — | — | |
| PEFT Ensemble (FiLM)Differential Privacy=non-DP, Backbone=ViT-B, Pre-training=ImageNet-21k2024.11 | 74.32 | — | — | 6.07 | — | — | — | |
| DualPromptFeature Replay=false2026.02 | 74.28 | — | 69.1 | — | — | — | — | |
| Non CL Baseline (FiLM)Differential Privacy=epsilon = 8, delta = 1e-5, Backbone=ViT-B, Pre-training=ImageNet-21k2024.11 | 73.77 | — | — | — | — | — | — | |
| LoRA-GPMBackbone=DINO2026.05 | 73.34 | 62.15 | — | — | -16.56 | — | — | |
| L2PBackbone=SAM2026.05 | 73.21 | 66.01 | — | — | 7.75 | — | — | |
| Coda-PromptBackbone=DINO2026.05 | 72.43 | 64.22 | — | — | 6.81 | — | — | |
| BiLoRABackbone=DINO2026.05 | 72.33 | 65.84 | — | — | 3.91 | — | — | |
| LoRA¯DRSBackbone=DINO2026.05 | 72.07 | 64.39 | — | — | 4.36 | — | — | |
| BiLoRABackbone=SAM2026.05 | 70.95 | 65.66 | — | — | 2.65 | — | — | |
| ADAMBackbone=CLIP ViT-B/162024.03 | 70.9 | 72.06 | — | — | — | — | — | |
| Coda-PromptBackbone=SAM2026.05 | 70.88 | 61.92 | — | — | 19.01 | — | — | |
| Fine-TuningBackbone=DINO2026.05 | 70.5 | 56.79 | — | — | 24.36 | — | — | |
| DualPromptBackbone=ViT-B/16, Pre-training=ImageNet-21K2024.03 | 69.39 | 65.41 | — | — | — | — | — | |
| L2PBackbone=DINO2026.05 | 69.26 | 62.48 | — | — | 5.37 | — | — | |
| L2PBackbone=ViT-B/16, Pre-training=ImageNet-21K2024.03 | 67.98 | 62.54 | — | — | — | — | — | |
| Dual-PromptBackbone=SAM2026.05 | 67.77 | 61.19 | — | — | 5.27 | — | — | |
| Dual-PromptBackbone=DINO2026.05 | 66.97 | 59.6 | — | — | 5.56 | — | — | |
| LoRA-GPMBackbone=SAM2026.05 | 66.4 | 58.44 | — | — | -26.85 | — | — | |
| PEFT Ensemble (FiLM)Differential Privacy=epsilon = 8, delta = 1e-5, Backbone=ViT-B, Pre-training=ImageNet-21k2024.11 | 64.91 | — | — | 8.87 | — | — | — | |
| sequentialBackbone=ViT-B/16, Pre-training=ImageNet-21K2024.03 | 62.91 | 46.07 | — | — | — | — | — | |
| Non CL Baseline (last layer)Differential Privacy=non-DP, Backbone=ViT-B, Pre-training=ImageNet-21k2024.11 | 62.17 | — | — | — | — | — | — | |
| Non CL Baseline (FiLM)Differential Privacy=epsilon = 1, delta = 1e-5, Backbone=ViT-B, Pre-training=ImageNet-21k2024.11 | 56.62 | — | — | — | — | — | — | |
| Cosine classifierDifferential Privacy=non-DP, Backbone=ViT-B, Pre-training=ImageNet-21k2024.11 | 56.3 | — | — | 7.51 | — | — | — | |
| Fine-TuningBackbone=SAM2026.05 | 55.7 | 47.66 | — | — | 40.37 | — | — | |
| Non CL Baseline (last layer)Differential Privacy=epsilon = 8, delta = 1e-5, Backbone=ViT-B, Pre-training=ImageNet-21k2024.11 | 52.16 | — | — | — | — | — | — | |
| PEFT Ensemble (last layer)Differential Privacy=non-DP, Backbone=ViT-B, Pre-training=ImageNet-21k2024.11 | 48.16 | — | — | 6.54 | — | — | — | |
| PEFT Ensemble (last layer)Differential Privacy=epsilon = 8, delta = 1e-5, Backbone=ViT-B, Pre-training=ImageNet-21k2024.11 | 47.97 | — | — | 6.33 | — | — | — | |
| Cosine classifierDifferential Privacy=epsilon = 8, delta = 1e-5, Backbone=ViT-B, Pre-training=ImageNet-21k2024.11 | 46.17 | — | — | 12.21 | — | — | — | |
| PEFT Ensemble (FiLM)Differential Privacy=epsilon = 1, delta = 1e-5, Backbone=ViT-B, Pre-training=ImageNet-21k2024.11 | 33.19 | — | — | 12.22 | — | — | — | |
| NaiveDifferential Privacy=epsilon = 8, delta = 1e-5, Backbone=ViT-B, Pre-training=ImageNet-21k2024.11 | 23.49 | — | — | 66.92 | — | — | — | |
| Non CL Baseline (last layer)Differential Privacy=epsilon = 1, delta = 1e-5, Backbone=ViT-B, Pre-training=ImageNet-21k2024.11 | 16.57 | — | — | — | — | — | — | |
| Cosine classifierDifferential Privacy=epsilon = 1, delta = 1e-5, Backbone=ViT-B, Pre-training=ImageNet-21k2024.11 | 13.04 | — | — | 13.89 | — | — | — | |
| NaiveDifferential Privacy=non-DP, Backbone=ViT-B, Pre-training=ImageNet-21k2024.11 | 9.12 | — | — | 87.51 | — | — | — | |
| PEFT Ensemble (last layer)Differential Privacy=epsilon = 1, delta = 1e-5, Backbone=ViT-B, Pre-training=ImageNet-21k2024.11 | 7.29 | — | — | 3.54 | — | — | — | |
| NaiveDifferential Privacy=epsilon = 1, delta = 1e-5, Backbone=ViT-B, Pre-training=ImageNet-21k2024.11 | 1.52 | — | — | 37.9 | — | — | — | |
| CODA-PPre-training=iBOT-1K, Backbone=ViT-B/162026.07 | — | — | — | — | — | 66.56 | 73.14 | |
| CODA-PPre-training=DINO-1K, Backbone=ViT-B/162026.07 | — | — | — | — | — | 63.15 | 69.73 | |
| DualPromptPre-training=iBOT-1K, Backbone=ViT-B/162026.07 | — | — | — | — | — | 61.51 | 67.11 | |
| DualPromptPre-training=DINO-1K, Backbone=ViT-B/162026.07 | — | — | — | — | — | 58.57 | 64.89 | |
| HiDe-PromptPre-training=iBOT-1K, Backbone=ViT-B/162026.07 | — | — | — | — | — | 71.33 | 73.62 | |
| HiDe-PromptPre-training=DINO-1K, Backbone=ViT-B/162026.07 | — | — | — | — | — | 68.11 | 71.7 | |
| ProbPromptPre-training=iBOT-1K, Backbone=ViT-B/162026.07 | — | — | — | — | — | 72.71 | 78.23 | |
| ProbPromptPre-training=DINO-1K, Backbone=ViT-B/162026.07 | — | — | — | — | — | 69.41 | 75.5 | |
| VQ-PromptPre-training=iBOT-1K, Backbone=ViT-B/162026.07 | — | — | — | — | — | 71.68 | 76.66 | |
| VQ-PromptPre-training=DINO-1K, Backbone=ViT-B/162026.07 | — | — | — | — | — | 68.42 | 74.43 |