Task Negation on 8 Vision
77.44Control Accuracyzero-shot
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| zero-shotBackbone=ViT-L-14, Number of tasks=8 tasks2026.04 | 77.44 | — | 64.54 | |
| Attention-Only Fine-tuningBackbone=ViT-L-14, Number of tasks=8 tasks2026.04 | 76.42 | — | 24.85 | |
| LoRA-ATTBackbone=ViT-L-14, Number of tasks=8 tasks2026.04 | 75.81 | — | 22.17 | |
| Pre-trainedBackbone=ViT-L/142026.02 | 75.5 | 65 | — | |
| Pre-trainedBackbone=ViT-L/142026.05 | 75.5 | 65 | — | |
| ATT-FT+OrthoRegBackbone=ViT-L-14, Number of tasks=8 tasks2026.04 | 75.4 | — | 14.67 | |
| Non-lin. FT+OrthoRegBackbone=ViT-L-14, Number of tasks=8 tasks2026.04 | 75.39 | — | 16.33 | |
| TTA+OrthoRegBackbone=ViT-L-14, Number of tasks=8 tasks2026.04 | 74.87 | — | 8.36 | |
| DELTABackbone=ViT-L/142026.05 | 74.7 | 11.7 | — | |
| LoRA-ATT+OrthoRegBackbone=ViT-L-14, Number of tasks=8 tasks2026.04 | 74.45 | — | 13.94 | |
| Non-linear FinetuningBackbone=ViT-L-14, Number of tasks=8 tasks2026.04 | 74.39 | — | 18.09 | |
| Tangent Task ArithmeticBackbone=ViT-L-14, Number of tasks=8 tasks2026.04 | 74.39 | — | 8.36 | |
| τJpBackbone=ViT-L/142026.05 | 74.2 | 1.8 | — | |
| TaLoSBackbone=ViT-L/142026.05 | 74.1 | 23 | — | |
| TaLoSBackbone=ViT-L/14, Dataless=true2026.02 | 73.6 | 10.7 | — | |
| TAKBackbone=ViT-L/142026.05 | 73.6 | 3.2 | — | |
| Attn. Only FTBackbone=ViT-L/142026.05 | 73.3 | 17.9 | — | |
| τJpBackbone=ViT-L/14, Dataless=false2026.02 | 73 | 3.7 | — | |
| TAKBackbone=ViT-L/14, Dataless=true2026.02 | 72.6 | 3.5 | — | |
| Non-linear FTBackbone=ViT-L/142026.02 | 72.4 | 18.1 | — | |
| Non-Linear FTBackbone=ViT-L/142026.05 | 72.3 | 18.1 | — | |
| Linear FTBackbone=ViT-L/142026.02 | 72.1 | 7.5 | — | |
| Linear FTBackbone=ViT-L/142026.05 | 72.1 | 7.1 | — | |
| zero-shotBackbone=ViT-B-16, Number of tasks=8 tasks2026.04 | 68.34 | — | 54.22 | |
| Pre-trainedBackbone=ViT-B/162026.02 | 68.3 | 55.4 | — | |
| Attention-Only Fine-tuningBackbone=ViT-B-16, Number of tasks=8 tasks2026.04 | 67.75 | — | 19.2 | |
| Non-lin. FT+OrthoRegBackbone=ViT-B-16, Number of tasks=8 tasks2026.04 | 67.42 | — | 19.51 | |
| LoRA-ATTBackbone=ViT-B-16, Number of tasks=8 tasks2026.04 | 67.28 | — | 19.44 | |
| LoRA-ATT+OrthoRegBackbone=ViT-B-16, Number of tasks=8 tasks2026.04 | 67.08 | — | 17.25 | |
| Tangent Task ArithmeticBackbone=ViT-B-16, Number of tasks=8 tasks2026.04 | 66.82 | — | 9.33 | |
| ATT-FT+OrthoRegBackbone=ViT-B-16, Number of tasks=8 tasks2026.04 | 66.81 | — | 14.78 | |
| TTA+OrthoRegBackbone=ViT-B-16, Number of tasks=8 tasks2026.04 | 66.73 | — | 7.49 | |
| zero-shotBackbone=ViT-B-32, Number of tasks=8 tasks2026.04 | 66.7 | — | 47.74 | |
| TAKBackbone=ViT-B/16, Dataless=true2026.02 | 66.4 | 3.4 | — | |
| Non-linear FinetuningBackbone=ViT-B-16, Number of tasks=8 tasks2026.04 | 66.38 | — | 20.29 | |
| TaLoSBackbone=ViT-B/16, Dataless=true2026.02 | 66.1 | 10.6 | — | |
| τJpBackbone=ViT-B/16, Dataless=false2026.02 | 66 | 4.7 | — | |
| Linear FTBackbone=ViT-B/162026.02 | 65.5 | 8.3 | — | |
| Non-linear FTBackbone=ViT-B/162026.02 | 65.3 | 20.4 | — | |
| Attention-Only Fine-tuningBackbone=ViT-B-32, Number of tasks=8 tasks2026.04 | 64.9 | — | 19.39 | |
| LoRA-ATTBackbone=ViT-B-32, Number of tasks=8 tasks2026.04 | 64.51 | — | 20.1 | |
| LoRA-ATT+OrthoRegBackbone=ViT-B-32, Number of tasks=8 tasks2026.04 | 64.43 | — | 19.19 | |
| ATT-FT+OrthoRegBackbone=ViT-B-32, Number of tasks=8 tasks2026.04 | 64.16 | — | 15.67 | |
| Non-lin. FT+OrthoRegBackbone=ViT-B-32, Number of tasks=8 tasks2026.04 | 64.07 | — | 18.55 | |
| TTA+OrthoRegBackbone=ViT-B-32, Number of tasks=8 tasks2026.04 | 64.07 | — | 11.39 | |
| Tangent Task ArithmeticBackbone=ViT-B-32, Number of tasks=8 tasks2026.04 | 63.99 | — | 11.47 | |
| Non-linear FinetuningBackbone=ViT-B-32, Number of tasks=8 tasks2026.04 | 63.91 | — | 25.05 | |
| Pre-trainedBackbone=ViT-B/322026.02 | 63.3 | 48.4 | — | |
| Pre-trainedBackbone=ViT-B/322026.05 | 63.3 | 48.4 | — | |
| TAKBackbone=ViT-B/32, Dataless=true2026.02 | 62.4 | 3.4 | — | |
| DELTABackbone=ViT-B/322026.05 | 62.1 | 9.6 | — | |
| TAKBackbone=ViT-B/322026.05 | 61.5 | 2.8 | — | |
| TaLoSBackbone=ViT-B/322026.05 | 61.1 | 18.4 | — | |
| τJpBackbone=ViT-B/32, Dataless=false2026.02 | 60.8 | 6.7 | — | |
| Attn. Only FTBackbone=ViT-B/322026.05 | 60.8 | 16.7 | — | |
| TaLoSBackbone=ViT-B/32, Dataless=true2026.02 | 60.7 | 11 | — | |
| τJpBackbone=ViT-B/322026.05 | 60.6 | 3 | — | |
| Non-linear FTBackbone=ViT-B/322026.02 | 60.5 | 20.4 | — | |
| Linear FTBackbone=ViT-B/322026.02 | 60.5 | 9.3 | — | |
| Linear FTBackbone=ViT-B/322026.05 | 60.5 | 9.3 | — | |
| Non-Linear FTBackbone=ViT-B/322026.05 | 60.5 | 20.4 | — |