Multi-task Image Classification on 7 Vision Benchmarks (test)
99.53Accuracy (MNIST)Finetuning
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| FinetuningBackbone=ViT-L/14, Fine-tuning Protocol=LoRA-fine-tuned2026.07 | 99.53 | 97.72 | 99.77 | 70.05 | 97.2 | 98.59 | 95.7 | 94.08 | |
| PoE-EBMAccuracy Type=Normalized, Backbone=ViT-L/14, Fine-tuning Protocol=LoRA-fine-tuned2026.07 | 89.99 | 77.32 | 83.56 | 81.32 | 68.31 | 61 | 83.98 | 77.93 | |
| PoE-EBMAccuracy Type=Absolute, Backbone=ViT-L/14, Fine-tuning Protocol=LoRA-fine-tuned2026.07 | 89.56 | 75.56 | 83.37 | 56.97 | 66.4 | 60.15 | 80.37 | 73.2 | |
| KnOTS-TIESAccuracy Type=Normalized, Backbone=ViT-L/14, Fine-tuning Protocol=LoRA-fine-tuned2026.07 | 82.1 | 77.14 | 84.03 | 83.22 | 70.63 | 62.62 | 83.38 | 77.59 | |
| KnOTS-TIESAccuracy Type=Absolute, Backbone=ViT-L/14, Fine-tuning Protocol=LoRA-fine-tuned2026.07 | 81.71 | 75.38 | 83.83 | 58.3 | 68.66 | 61.74 | 79.79 | 72.77 | |
| Weight averagingAccuracy Type=Normalized, Backbone=ViT-L/14, Fine-tuning Protocol=LoRA-fine-tuned2026.07 | 78.93 | 61.59 | 78.4 | 79.65 | 53.69 | 63.64 | 75.65 | 70.22 | |
| Weight averagingAccuracy Type=Absolute, Backbone=ViT-L/14, Fine-tuning Protocol=LoRA-fine-tuned2026.07 | 78.55 | 60.19 | 78.22 | 55.8 | 52.19 | 62.74 | 72.4 | 65.73 | |
| Task Arithmetic (TA)Accuracy Type=Normalized, Backbone=ViT-L/14, Fine-tuning Protocol=LoRA-fine-tuned2026.07 | 78.3 | 61.09 | 78.33 | 79.5 | 53.35 | 63.6 | 75.3 | 69.62 | |
| Task Arithmetic (TA)Accuracy Type=Absolute, Backbone=ViT-L/14, Fine-tuning Protocol=LoRA-fine-tuned2026.07 | 77.93 | 59.7 | 78.14 | 55.69 | 51.85 | 62.7 | 72.06 | 65.44 | |
| TA-DARE-TIESAccuracy Type=Normalized, Backbone=ViT-L/14, Fine-tuning Protocol=LoRA-fine-tuned2026.07 | 71.2 | 73.14 | 81.75 | 82.01 | 62.12 | 58.04 | 82.68 | 72.99 | |
| TA-DARE-TIESAccuracy Type=Absolute, Backbone=ViT-L/14, Fine-tuning Protocol=LoRA-fine-tuned2026.07 | 70.86 | 71.48 | 81.86 | 57.45 | 60.38 | 57.22 | 79.13 | 68.3 | |
| KnOTS-DARE-TIESAccuracy Type=Normalized, Backbone=ViT-L/14, Fine-tuning Protocol=LoRA-fine-tuned2026.07 | 67.75 | 68.62 | 82.88 | 83.07 | 65.98 | 60.52 | 83.6 | 73.2 | |
| KnOTS-DARE-TIESAccuracy Type=Absolute, Backbone=ViT-L/14, Fine-tuning Protocol=LoRA-fine-tuned2026.07 | 67.43 | 67.06 | 82.68 | 58.19 | 64.13 | 59.67 | 80 | 68.45 |