Image Classification on TALL-20
93.5AccuracyFine-tuned
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Fine-tunedBackbone=CLIP ViT-L/142025.12 | 93.5 | 100 | |
| MTLBackbone=CLIP ViT-L/142025.12 | 91.8 | 98.2 | |
| SAMergingBackbone=CLIP ViT-L/14, k=16002025.12 | 89.9 | 96.1 | |
| Fine-tunedBackbone=CLIP ViT-B/322025.12 | 89.8 | 100 | |
| MTLBackbone=CLIP ViT-B/322025.12 | 88.9 | 99 | |
| Isotropic MergingBackbone=CLIP ViT-L/142025.12 | 84.8 | 90.7 | |
| RegMean++Backbone=CLIP ViT-L/14, k=16002025.12 | 82.5 | 88.2 | |
| SAMergingBackbone=CLIP ViT-B/32, k=16002025.12 | 81.1 | 90.3 | |
| AdaMerging LWBackbone=CLIP ViT-L/14, k=160002025.12 | 79 | 84.5 | |
| RegMeanBackbone=CLIP ViT-L/14, k=16002025.12 | 78.8 | 87.8 | |
| SWUDI → AdaMergingBackbone=B/322026.06 | 78.12 | — | |
| SWUDI-A → AdaMergingBackbone=B/322026.06 | 78.03 | — | |
| SWUDI-ABackbone=B/322026.06 | 75.62 | — | |
| SWUDIBackbone=B/322026.06 | 75.6 | — | |
| RegMean++Backbone=CLIP ViT-B/32, k=16002025.12 | 74 | 82.4 | |
| Isotropic MergingBackbone=CLIP ViT-B/322025.12 | 73.5 | 81.8 | |
| TSV MergingBackbone=B/322026.06 | 73.22 | — | |
| PCB MergingBackbone=CLIP ViT-L/142025.12 | 72.6 | 77.6 | |
| τ^cf = DC†Backbone=B/322026.06 | 72.54 | — | |
| TIES-MergingBackbone=CLIP ViT-L/142025.12 | 72.3 | 77.3 | |
| AdaMerging LWBackbone=CLIP ViT-L/14, k=16002025.12 | 71.5 | 76.5 | |
| Simple AveragingBackbone=CLIP ViT-L/142025.12 | 71.1 | 76 | |
| Task ArithmeticBackbone=CLIP ViT-L/142025.12 | 71.1 | 76 | |
| FisherBackbone=CLIP ViT-L/14, k=16002025.12 | 70.4 | 75.3 | |
| Iso-CBackbone=B/322026.06 | 70.35 | — | |
| RegMeanBackbone=CLIP ViT-B/32, k=16002025.12 | 70 | 78 | |
| AdaMerging LWBackbone=CLIP ViT-B/32, k=160002025.12 | 69.4 | 77.3 | |
| PretrainedBackbone=CLIP ViT-L/142025.12 | 65.6 | 70.2 | |
| PCB MergingBackbone=CLIP ViT-B/322025.12 | 64.1 | 71.4 | |
| TIES MergingBackbone=B/322026.06 | 62.76 | — | |
| TIES-MergingBackbone=CLIP ViT-B/322025.12 | 62.7 | 69.8 | |
| TIES w/ DAREBackbone=B/322026.06 | 62.22 | — | |
| FisherBackbone=CLIP ViT-B/32, k=16002025.12 | 62.2 | 69.3 | |
| OptMergeBackbone=B/322026.06 | 61.71 | — | |
| AdaMerging LWBackbone=CLIP ViT-B/32, k=16002025.12 | 61.5 | 68.5 | |
| Simple AveragingBackbone=CLIP ViT-B/322025.12 | 61.1 | 68 | |
| Task ArithmeticBackbone=CLIP ViT-B/322025.12 | 61.1 | 68 | |
| Weight AverageBackbone=B/322026.06 | 61.1 | — | |
| WUDI MergingBackbone=B/322026.06 | 61.06 | — | |
| Task ArithmeticBackbone=B/322026.06 | 60.62 | — | |
| TA w/ DAREBackbone=B/322026.06 | 60.55 | — | |
| PretrainedBackbone=CLIP ViT-B/322025.12 | 55.6 | 61.9 |