Continual Model Merging on 20-task sequence
93.5AccuracyFine-Tuned
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| Fine-TunedBackbone=ViT-L/142026.04 | 93.5 | — | — | — | — | — | — | |
| Fine-TunedBackbone=ViT-B/162026.04 | 91.6 | — | — | — | — | — | — | |
| Fine-TunedBackbone=ViT-B/322026.04 | 89.8 | — | — | — | — | — | — | |
| MADE-ITBackbone=ViT-L/142026.04 | 89.6 | -1.7 | — | — | — | — | — | |
| MINGLEBackbone=ViT-L/142026.04 | 85.5 | -2.6 | — | — | — | — | — | |
| MADE-ITBackbone=ViT-B/162026.04 | 83 | -3 | — | — | — | — | — | |
| MINGLEBackbone=ViT-B/162026.04 | 81.9 | -1.9 | — | — | — | — | — | |
| MADE-ITBackbone=ViT-B/322026.04 | 81.4 | -3.8 | — | — | — | — | — | |
| Continual Fine-TunedBackbone=ViT-L/142026.04 | 77.7 | — | — | — | — | — | — | |
| MINGLEBackbone=ViT-B/322026.04 | 77.1 | -2.2 | — | — | — | — | — | |
| OPCMBackbone=ViT-L/142026.04 | 76 | -6.5 | — | — | — | — | — | |
| Average (SWA)Backbone=ViT-L/142026.04 | 71.1 | -6.4 | — | — | — | — | — | |
| OPCMBackbone=ViT-B/162026.04 | 70.3 | -6.3 | — | — | — | — | — | |
| Continual Task ArithmeticBackbone=ViT-L/142026.04 | 70.3 | -3.3 | — | — | — | — | — | |
| Continual Ties-MergingBackbone=ViT-L/142026.04 | 68.3 | -2.9 | — | — | — | — | — | |
| Continual Fine-TunedBackbone=ViT-B/162026.04 | 68.2 | — | — | — | — | — | — | |
| Continual LW AdaMergingBackbone=ViT-L/142026.04 | 66.9 | -21.7 | — | — | — | — | — | |
| OPCMBackbone=ViT-B/322026.04 | 65.7 | -7.8 | — | — | — | — | — | |
| Pre-TrainedBackbone=ViT-L/142026.04 | 65.6 | — | — | — | — | — | — | |
| Average (SWA)Backbone=ViT-B/162026.04 | 64.8 | -7.3 | — | — | — | — | — | |
| Continual Task ArithmeticBackbone=ViT-B/162026.04 | 64.2 | -3.6 | — | — | — | — | — | |
| Continual Ties-MergingBackbone=ViT-B/162026.04 | 63 | -1.5 | — | — | — | — | — | |
| Continual Fine-TunedBackbone=ViT-B/322026.04 | 62.6 | — | — | — | — | — | — | |
| Continual LW AdaMergingBackbone=ViT-B/162026.04 | 61.5 | -21.4 | — | — | — | — | — | |
| Average (SWA)Backbone=ViT-B/322026.04 | 61.1 | -7.1 | — | — | — | — | — | |
| Continual Task ArithmeticBackbone=ViT-B/322026.04 | 60 | -3.4 | — | — | — | — | — | |
| Continual Ties-MergingBackbone=ViT-B/322026.04 | 59.9 | -1.5 | — | — | — | — | — | |
| Pre-TrainedBackbone=ViT-B/162026.04 | 59.8 | — | — | — | — | — | — | |
| Continual LW AdaMergingBackbone=ViT-B/322026.04 | 59.7 | -22.7 | — | — | — | — | — | |
| Continual LoRA-WEMoEBackbone=ViT-L/142026.04 | 56.9 | -27.9 | — | — | — | — | — | |
| Pre-TrainedBackbone=ViT-B/322026.04 | 55.6 | — | — | — | — | — | — | |
| Continual LoRA-WEMoEBackbone=ViT-B/162026.04 | 55 | -25.8 | — | — | — | — | — | |
| Continual LoRA-WEMoEBackbone=ViT-B/322026.04 | 49.6 | -24.5 | — | — | — | — | — | |
| DOGE AMBackbone=ViT-B/16, Method Category=Test-Time Adaption Methods2025.08 | — | — | 60.5 | 65.5 | 69.3 | 68.9 | 69 | |
| DOGE TABackbone=ViT-B/16, Method Category=Data-Free Methods2025.08 | — | — | 63.4 | 65.8 | 66.9 | 66.7 | 67.9 | |
| Fisher MergingBackbone=ViT-B/16, Method Category=Training-Free Methods2025.08 | — | — | 63.1 | 63.3 | 63.5 | 62.6 | 63.7 | |
| Iso-CBackbone=ViT-B/16, Method Category=Data-Free Methods2025.08 | — | — | 63.6 | 65 | 63.7 | 56.6 | 45.1 | |
| Iso-CTSBackbone=ViT-B/16, Method Category=Data-Free Methods2025.08 | — | — | 63.4 | 63.5 | 54.2 | 41.9 | 26.7 | |
| Layer-wise AdaMergingBackbone=ViT-B/16, Method Category=Test-Time Adaption Methods2025.08 | — | — | 61.2 | 64.4 | 67.2 | 66.8 | 66.9 | |
| Model SoupsBackbone=ViT-B/16, Method Category=Data-Free Methods2025.08 | — | — | 62.8 | 62.9 | 64.1 | 62.8 | 63.6 | |
| RegMeanBackbone=ViT-B/16, Method Category=Training-Free Methods2025.08 | — | — | 62 | 67.6 | 69 | 69.7 | 71.7 | |
| RegMean++Backbone=ViT-B/16, Method Category=Training-Free Methods2025.08 | — | — | 62.2 | 67.8 | 69.1 | 70.2 | 72.5 | |
| Task ArithmeticBackbone=ViT-B/16, Method Category=Data-Free Methods2025.08 | — | — | 62.2 | 61.1 | 63.3 | 62.1 | 62.2 | |
| TIES-MergingBackbone=ViT-B/16, Method Category=Data-Free Methods2025.08 | — | — | 63.7 | 63.5 | 64.2 | 63 | 64.4 | |
| TSV-MBackbone=ViT-B/16, Method Category=Data-Free Methods2025.08 | — | — | 60.6 | 65 | 66.4 | 66.9 | 67.5 |