Image Classification on 8-task Vision Benchmark
95.8Average AccuracyFinetuned
Evaluation results
| Method | Links | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| FinetunedBackbone=ViT-L/142026.03 | 95.8 | — | — | — | — | — | — | — | — | — | — | — | |
| theta_B fine-tuneN=–, Evaluation Protocol=fine-tune, Target Model=ViT-L/14, Pre-training=Datacomp-XL2026.05 | 94.74 | 95.22 | 86.11 | 99.33 | 99.29 | 99.74 | 97.58 | 82.4 | 98.21 | — | — | 24.45 | |
| Iso-CTSBackbone=ViT-L/142026.03 | 94.7 | — | — | — | — | — | — | — | — | — | — | — | |
| Iso-CTS + RIBackbone=ViT-L/142026.03 | 94.7 | — | — | — | — | — | — | — | — | — | — | — | |
| FinetunedBackbone=ViT-B/162026.03 | 94.6 | — | — | — | — | — | — | — | — | — | — | — | |
| WUDI + RIBackbone=ViT-L/142026.03 | 94.3 | — | — | — | — | — | — | — | — | — | — | — | |
| Iso-C + RIBackbone=ViT-L/142026.03 | 94.3 | — | — | — | — | — | — | — | — | — | — | — | |
| Iso-CBackbone=ViT-L/142026.03 | 94.2 | — | — | — | — | — | — | — | — | — | — | — | |
| WUDIBackbone=ViT-L/142026.03 | 94 | — | — | — | — | — | — | — | — | — | — | — | |
| theta_A fine-tuneN=–, Evaluation Protocol=fine-tune, Model=ViT-B/16, Pre-training=Datacomp-XL2026.05 | 93.57 | 92.5 | 84.25 | 98.96 | 99.06 | 99.77 | 96.71 | 79.25 | 98.02 | — | — | 23.28 | |
| TSV-M + RIBackbone=ViT-L/142026.03 | 93.5 | — | — | — | — | — | — | — | — | — | — | — | |
| DOGE AMBackbone=ViT-L/14, Category=Test-Time Adaptation2025.08 | 93 | — | — | — | — | — | — | — | — | — | — | — | |
| theta_A fine-tuneEvaluation Protocol=fine-tune, Source Model=ViT-B/16-plus (LAION-400M), Scaling Direction=wide to narrow2026.05 | 92.9 | 90.11 | 82.55 | 99 | 99.19 | 99.81 | 96.46 | 78.29 | 97.8 | — | — | 30.71 | |
| θB fine-tuneN=–2026.05 | 92.9 | 90.11 | 82.55 | 99 | 99.19 | 99.81 | 96.46 | 78.29 | 97.8 | — | — | 35.24 | |
| theta_B fine-tuneEvaluation Protocol=fine-tune, Target Model=ViT-B/16 (LAION-2B), Scaling Direction=wide to narrow2026.05 | 92.88 | 92.1 | 83.46 | 94.67 | 99.03 | 99.72 | 96.81 | 79.35 | 97.89 | — | — | 30.69 | |
| θA fine-tuneN=–2026.05 | 92.88 | 92.1 | 83.46 | 94.67 | 99.03 | 99.72 | 96.81 | 79.35 | 97.89 | — | — | 35.22 | |
| FinetunedBackbone=ViT-B/322026.03 | 92.8 | — | — | — | — | — | — | — | — | — | — | — | |
| TSV-MBackbone=ViT-L/142026.03 | 92.8 | — | — | — | — | — | — | — | — | — | — | — | |
| Iso-CTSBackbone=ViT-L/14, Category=Data-Free Methods2025.08 | 92.8 | — | — | — | — | — | — | — | — | — | — | — | |
| Iso-CBackbone=ViT-L/14, Category=Data-Free Methods2025.08 | 92.4 | — | — | — | — | — | — | — | — | — | — | — | |
| Fine-tunedBackbone=ViT-B/162025.08 | 92.3 | 85.9 | 82.3 | 99 | 99 | 99.7 | 96.6 | 78.9 | 97.6 | — | — | — | |
| RegMean++Backbone=ViT-L/14, Category=Training-Free Methods2025.08 | 91.9 | — | — | — | — | — | — | — | — | — | — | — | |
| Layer-wise AdaMergingBackbone=ViT-L/14, Category=Test-Time Adaptation2025.08 | 91.8 | — | — | — | — | — | — | — | — | — | — | — | |
| RegMeanBackbone=ViT-L/14, Category=Training-Free Methods2025.08 | 91.5 | — | — | — | — | — | — | — | — | — | — | — | |
| TSV-MBackbone=ViT-L/14, Category=Data-Free Methods2025.08 | 91.3 | — | — | — | — | — | — | — | — | — | — | — | |
| Iso-CTS + RIBackbone=ViT-B/162026.03 | 91.1 | — | — | — | — | — | — | — | — | — | — | — | |
| Iso-CTSBackbone=ViT-B/162026.03 | 91 | — | — | — | — | — | — | — | — | — | — | — | |
| WUDI + RIBackbone=ViT-B/162026.03 | 90.8 | — | — | — | — | — | — | — | — | — | — | — | |
| Iso-C + RIBackbone=ViT-B/162026.03 | 90.8 | — | — | — | — | — | — | — | — | — | — | — | |
| Iso-CBackbone=ViT-B/162026.03 | 90.3 | — | — | — | — | — | — | — | — | — | — | — | |
| Fine-tunedBackbone=ViT-B/322025.08 | 90.3 | 78.3 | 79.7 | 99 | 98.9 | 99.6 | 95.2 | 75 | 97.3 | — | — | — | |
| TSV-M + RIBackbone=ViT-B/162026.03 | 90 | — | — | — | — | — | — | — | — | — | — | — | |
| WUDIBackbone=ViT-B/162026.03 | 89.8 | — | — | — | — | — | — | — | — | — | — | — | |
| DOGE TABackbone=ViT-L/14, Category=Data-Free Methods2025.08 | 89.8 | — | — | — | — | — | — | — | — | — | — | — | |
| DOGE AMBackbone=ViT-B/16, Category=Test-Time Adaptation2025.08 | 89.6 | — | — | — | — | — | — | — | — | — | — | — | |
| TIES + RIBackbone=ViT-L/142026.03 | 89.4 | — | — | — | — | — | — | — | — | — | — | — | |
| RegMean++Backbone=ViT-B/16, Category=Training-Free Methods2025.08 | 88.9 | — | — | — | — | — | — | — | — | — | — | — | |
| TSV-MBackbone=ViT-B/162026.03 | 88.7 | — | — | — | — | — | — | — | — | — | — | — | |
| MTLBackbone=ViT-B/322025.08 | 88.6 | 76.6 | 77.7 | 97.9 | 97.7 | 99.3 | 92.2 | 72.3 | 95.5 | — | — | — | |
| KnOTS + RIBackbone=ViT-L/142026.03 | 88.4 | — | — | — | — | — | — | — | — | — | — | — | |
| Iso-CTSBackbone=ViT-B/162025.08 | 88.3 | 83.8 | 73.7 | 96 | 94.7 | 98.6 | 92.6 | 76.2 | 90.9 | — | — | — | |
| TSV-MBackbone=ViT-B/16, Category=Data-Free Methods2025.08 | 88.2 | — | — | — | — | — | — | — | — | — | — | — | |
| DOGE AMBackbone=ViT-B/162025.08 | 88.2 | 82.4 | 75.8 | 94.6 | 96.4 | 98.6 | 90.5 | 72.6 | 94.8 | — | — | — | |
| WUDI + RIBackbone=ViT-B/322026.03 | 88.1 | — | — | — | — | — | — | — | — | — | — | — | |
| Iso-CTSBackbone=ViT-B/16, Category=Data-Free Methods2025.08 | 88.1 | — | — | — | — | — | — | — | — | — | — | — | |
| RegMeanBackbone=ViT-B/16, Category=Training-Free Methods2025.08 | 88 | — | — | — | — | — | — | — | — | — | — | — | |
| Iso-CBackbone=ViT-B/162025.08 | 87.9 | 82.9 | 71.2 | 96.3 | 94.5 | 98.7 | 92.3 | 75.9 | 91.1 | — | — | — | |
| KnOTSBackbone=ViT-L/142026.03 | 87.8 | — | — | — | — | — | — | — | — | — | — | — | |
| Iso-CBackbone=ViT-B/16, Category=Data-Free Methods2025.08 | 87.8 | — | — | — | — | — | — | — | — | — | — | — | |
| DOGE AMBackbone=ViT-B/32, Category=Test-Time Adaptation2025.08 | 87.4 | — | — | — | — | — | — | — | — | — | — | — | |
| TA + RIBackbone=ViT-L/142026.03 | 87.3 | — | — | — | — | — | — | — | — | — | — | — | |
| Layer-wise AdaMergingBackbone=ViT-B/16, Category=Test-Time Adaptation2025.08 | 87.2 | — | — | — | — | — | — | — | — | — | — | — | |
| RegMean++Backbone=ViT-B/162025.08 | 87.2 | 78.9 | 71 | 97.3 | 93 | 99.1 | 89.3 | 72.8 | 96 | — | — | — | |
| TSV-MBackbone=ViT-B/162025.08 | 87.1 | 80.7 | 69.7 | 96.2 | 94.1 | 99.1 | 89.7 | 73.1 | 94.1 | — | — | — | |
| Iso-C + RIBackbone=ViT-B/322026.03 | 87 | — | — | — | — | — | — | — | — | — | — | — | |
| TSV-M + RIBackbone=ViT-B/322026.03 | 87 | — | — | — | — | — | — | — | — | — | — | — | |
| Iso-CTS + RIBackbone=ViT-B/322026.03 | 86.8 | — | — | — | — | — | — | — | — | — | — | — | |
| TIESBackbone=ViT-L/142026.03 | 86.7 | — | — | — | — | — | — | — | — | — | — | — | |
| WUDIBackbone=ViT-B/322026.03 | 86.7 | — | — | — | — | — | — | — | — | — | — | — | |
| Iso-CTSBackbone=ViT-B/322026.03 | 86.6 | — | — | — | — | — | — | — | — | — | — | — | |
| DOGE TABackbone=ViT-B/16, Category=Data-Free Methods2025.08 | 86.5 | — | — | — | — | — | — | — | — | — | — | — | |
| Iso-CBackbone=ViT-B/322026.03 | 86.3 | — | — | — | — | — | — | — | — | — | — | — | |
| RegMean++Backbone=ViT-B/32, Category=Training-Free Methods2025.08 | 86.2 | — | — | — | — | — | — | — | — | — | — | — | |
| RegMeanBackbone=ViT-B/162025.08 | 86 | 78.8 | 67.9 | 96.3 | 90 | 98.8 | 89.2 | 72.6 | 94.9 | — | — | — | |
| DOGE AMBackbone=ViT-B/32, Method Category=Test-Time Adaption2025.08 | 85.8 | 74.5 | 73 | 93.7 | 95.5 | 98.8 | 88.7 | 70.6 | 91.4 | — | — | — | |
| TSV-MBackbone=ViT-B/322026.03 | 85.4 | — | — | — | — | — | — | — | — | — | — | — | |
| Layer-wise AdaMergingBackbone=ViT-B/162025.08 | 85.1 | 79.8 | 62.7 | 93.4 | 95.6 | 98.1 | 86.5 | 70.7 | 93.7 | — | — | — | |
| TABackbone=ViT-L/142026.03 | 84.8 | — | — | — | — | — | — | — | — | — | — | — | |
| TSV-MBackbone=ViT-B/32, Category=Data-Free Methods2025.08 | 84.8 | — | — | — | — | — | — | — | — | — | — | — | |
| Task ArithmeticBackbone=ViT-L/14, Category=Data-Free Methods2025.08 | 84.7 | — | — | — | — | — | — | — | — | — | — | — | |
| TIES-MergingBackbone=ViT-L/14, Category=Data-Free Methods2025.08 | 84.7 | — | — | — | — | — | — | — | — | — | — | — | |
| DOGE TABackbone=ViT-B/162025.08 | 84.7 | 77.5 | 65.3 | 95.1 | 91.4 | 98.8 | 85.9 | 70.8 | 92.7 | — | — | — | |
| RegMeanBackbone=ViT-B/32, Category=Training-Free Methods2025.08 | 84.6 | — | — | — | — | — | — | — | — | — | — | — | |
| Layer-wise AdaMergingBackbone=ViT-B/32, Category=Test-Time Adaptation2025.08 | 84.5 | — | — | — | — | — | — | — | — | — | — | — | |
| RegMean++Backbone=ViT-B/32, Method Category=Training-Free Methods2025.08 | 84.4 | 70.5 | 68.7 | 96.1 | 90.4 | 99 | 86.7 | 69.3 | 94.1 | — | — | — | |
| TIES + RIBackbone=ViT-B/162026.03 | 84.1 | — | — | — | — | — | — | — | — | — | — | — | |
| Fisher MergingBackbone=ViT-L/14, Category=Training-Free Methods2025.08 | 83.7 | — | — | — | — | — | — | — | — | — | — | — | |
| BiCoN=100, Evaluation Protocol=task-vector transfer, Source Model=ViT-B/16, Target Model=ViT-L/14, Pre-training=Datacomp-XL2026.05 | 83.18 | 93.8 | 71.17 | 80.44 | 78.63 | 98.27 | 82.33 | 74.7 | 86.1 | — | — | 12.89 | |
| KnOTS + RIBackbone=ViT-B/162026.03 | 83.1 | — | — | — | — | — | — | — | — | — | — | — | |
| TSV-MBackbone=ViT-B/32, Method Category=Data-Free Methods2025.08 | 83.1 | 71.6 | 63.8 | 93.4 | 92.5 | 98.9 | 84.7 | 67.6 | 91.9 | — | — | — | |
| DOGE TABackbone=ViT-B/32, Category=Data-Free Methods2025.08 | 83.1 | — | — | — | — | — | — | — | — | — | — | — | |
| Iso-CBackbone=ViT-B/32, Category=Data-Free Methods2025.08 | 83.1 | — | — | — | — | — | — | — | — | — | — | — | |
| Iso-CTSBackbone=ViT-B/32, Category=Data-Free Methods2025.08 | 82.8 | — | — | — | — | — | — | — | — | — | — | — | |
| Iso-CTSBackbone=ViT-B/32, Method Category=Data-Free Methods2025.08 | 82.7 | 74.6 | 68.5 | 89.1 | 90.4 | 98.1 | 86.6 | 71.1 | 83.4 | — | — | — | |
| Model SoupsBackbone=ViT-L/14, Category=Data-Free Methods2025.08 | 82.7 | — | — | — | — | — | — | — | — | — | — | — | |
| Layer-wise AdaMergingBackbone=ViT-B/32, Method Category=Test-Time Adaption2025.08 | 82.6 | 71.1 | 66.8 | 92.3 | 93.3 | 98.2 | 83.9 | 67.8 | 87.8 | — | — | — | |
| Iso-CBackbone=ViT-B/32, Method Category=Data-Free Methods2025.08 | 82.5 | 73.9 | 65.8 | 89.6 | 90.8 | 98.2 | 86 | 71 | 84.8 | — | — | — | |
| RegMeanBackbone=ViT-B/32, Method Category=Training-Free Methods2025.08 | 82.4 | 70 | 66.1 | 95.4 | 83.4 | 98.4 | 84.6 | 68.6 | 92.6 | — | — | — | |
| KnOTSBackbone=ViT-B/162026.03 | 81.7 | — | — | — | — | — | — | — | — | — | — | — | |
| BiCoN=100, Source Model=ViT-B/16-plus (LAION-400M), Target Model=ViT-B/16 (LAION-2B), Scaling Direction=wide to narrow2026.05 | 81.42 | 86.79 | 65.05 | 88.33 | 78.39 | 93.19 | 79.32 | 71.21 | 89.1 | — | — | 19.23 | |
| TA + RIBackbone=ViT-B/162026.03 | 81.4 | — | — | — | — | — | — | — | — | — | — | — | |
| DOGE TABackbone=ViT-B/32, Method Category=Data-Free Methods2025.08 | 80.6 | 69.9 | 63.8 | 89.8 | 86.8 | 98.3 | 81.9 | 67.7 | 86.2 | — | — | — | |
| TARA-variant BBackbone=CLIP-ViT-B/322026.03 | 80.5 | — | — | — | — | — | — | — | — | 5.1 | 5,922 | — | |
| Task ArithmeticBackbone=ViT-B/16, Category=Data-Free Methods2025.08 | 80.2 | — | — | — | — | — | — | — | — | — | — | — | |
| CALMBackbone=CLIP-ViT-B/322026.03 | 80 | — | — | — | — | — | — | — | — | 20 | 8,066 | — | |
| AveragingBackbone=ViT-L/142026.03 | 79.5 | — | — | — | — | — | — | — | — | — | — | — | |
| TIESBackbone=ViT-B/162026.03 | 79.5 | — | — | — | — | — | — | — | — | — | — | — | |
| TIES + RIBackbone=ViT-B/322026.03 | 79.3 | — | — | — | — | — | — | — | — | — | — | — | |
| Averaging + RIBackbone=ViT-L/142026.03 | 78.9 | — | — | — | — | — | — | — | — | — | — | — | |
| AdamergingBackbone=CLIP-ViT-B/322026.03 | 78.7 | — | — | — | — | — | — | — | — | 4.1 | 4,344 | — |