Task Addition on 8 classification datasets (Cars, DTD, EuroSAT, GTSRB, MNIST, RESISC45, SUN397, SVHN) (test)
90.41Absolute AccuracyATT-FT+OrthoReg
Evaluation results
| Method | Links | ||
|---|---|---|---|
| ATT-FT+OrthoRegBackbone=ViT-L-14, Number of tasks=8, Evaluation protocol=Task Addition2026.04 | 90.41 | 100.05 | |
| LoRA-ATT+OrthoRegBackbone=ViT-L-14, Number of tasks=8, Evaluation protocol=Task Addition2026.04 | 89.16 | 95.5 | |
| Non-lin. FT+OrthoRegBackbone=ViT-L-14, Number of tasks=8, Evaluation protocol=Task Addition2026.04 | 88.23 | 100.08 | |
| Attention-Only Fine-tuningBackbone=ViT-L-14, Number of tasks=8, Evaluation protocol=Task Addition2026.04 | 87.81 | 93.59 | |
| TTA+OrthoRegBackbone=ViT-L-14, Number of tasks=8, Evaluation protocol=Task Addition2026.04 | 87.52 | 96.44 | |
| LoRA-ATTBackbone=ViT-L-14, Number of tasks=8, Evaluation protocol=Task Addition2026.04 | 87.02 | 93.33 | |
| Tangent Task ArithmeticBackbone=ViT-L-14, Number of tasks=8, Evaluation protocol=Task Addition2026.04 | 86.19 | 93.14 | |
| Non-linear FinetuningBackbone=ViT-L-14, Number of tasks=8, Evaluation protocol=Task Addition2026.04 | 84.07 | 89.19 | |
| ATT-FT+OrthoRegBackbone=ViT-B-16, Number of tasks=8, Evaluation protocol=Task Addition2026.04 | 83.37 | 98.77 | |
| ATT-FT+OrthoRegBackbone=ViT-B-32, Number of tasks=8, Evaluation protocol=Task Addition2026.04 | 80.87 | 99.76 | |
| Attention-Only Fine-tuningBackbone=ViT-B-16, Number of tasks=8, Evaluation protocol=Task Addition2026.04 | 80.71 | 87.64 | |
| AE-TSVMBackbone=ViT-L/14, M=50%2025.02 | 80.4 | — | |
| TTA+OrthoRegBackbone=ViT-B-16, Number of tasks=8, Evaluation protocol=Task Addition2026.04 | 79.85 | 88.02 | |
| RandSelect-TSVMBackbone=ViT-L/14, M=50%2025.02 | 79.84 | — | |
| LoRA-ATT+OrthoRegBackbone=ViT-B-16, Number of tasks=8, Evaluation protocol=Task Addition2026.04 | 79.67 | 87.7 | |
| AE-TABackbone=ViT-L/14, M=50%2025.02 | 78.99 | — | |
| Tangent Task ArithmeticBackbone=ViT-B-16, Number of tasks=8, Evaluation protocol=Task Addition2026.04 | 78.97 | 87.48 | |
| Attention-Only Fine-tuningBackbone=ViT-B-32, Number of tasks=8, Evaluation protocol=Task Addition2026.04 | 78.07 | 86.99 | |
| Non-lin. FT+OrthoRegBackbone=ViT-B-16, Number of tasks=8, Evaluation protocol=Task Addition2026.04 | 77.68 | 93.62 | |
| RandSelect-TABackbone=ViT-L/14, M=50%2025.02 | 77.47 | — | |
| TTA+OrthoRegBackbone=ViT-B-32, Number of tasks=8, Evaluation protocol=Task Addition2026.04 | 76.35 | 91.81 | |
| LoRA-ATT+OrthoRegBackbone=ViT-B-32, Number of tasks=8, Evaluation protocol=Task Addition2026.04 | 76 | 86.1 | |
| LoRA-ATTBackbone=ViT-B-16, Number of tasks=8, Evaluation protocol=Task Addition2026.04 | 75.51 | 83.17 | |
| Non-linear FinetuningBackbone=ViT-B-16, Number of tasks=8, Evaluation protocol=Task Addition2026.04 | 75.39 | 75.39 | |
| Tangent Task ArithmeticBackbone=ViT-B-32, Number of tasks=8, Evaluation protocol=Task Addition2026.04 | 74.68 | 85.27 | |
| LoRA-ATTBackbone=ViT-B-32, Number of tasks=8, Evaluation protocol=Task Addition2026.04 | 73.84 | 84.29 | |
| Non-lin. FT+OrthoRegBackbone=ViT-B-32, Number of tasks=8, Evaluation protocol=Task Addition2026.04 | 73.41 | 93.93 | |
| RandSelect-TSVMBackbone=ViT-B/16, M=50%2025.02 | 73.17 | — | |
| AE-TSVMBackbone=ViT-B/16, M=50%2025.02 | 73.03 | — | |
| AE-TABackbone=ViT-B/16, M=50%2025.02 | 71.23 | — | |
| RandSelect-TABackbone=ViT-B/16, M=50%2025.02 | 70.86 | — | |
| Non-linear FinetuningBackbone=ViT-B-32, Number of tasks=8, Evaluation protocol=Task Addition2026.04 | 70.32 | 77.56 | |
| AE-TSVMBackbone=ViT-B/32, M=50%2025.02 | 69.01 | — | |
| AE-TABackbone=ViT-B/32, M=50%2025.02 | 66.6 | — | |
| RandSelect-TSVMBackbone=ViT-B/32, M=50%2025.02 | 65.61 | — | |
| zero-shotBackbone=ViT-L-14, Number of tasks=8, Evaluation protocol=Task Addition2026.04 | 64.54 | — | |
| RandSelect-TABackbone=ViT-B/32, M=50%2025.02 | 62.04 | — | |
| zero-shotBackbone=ViT-B-16, Number of tasks=8, Evaluation protocol=Task Addition2026.04 | 54.22 | — | |
| zero-shotBackbone=ViT-B-32, Number of tasks=8, Evaluation protocol=Task Addition2026.04 | 47.74 | — |