Zero-shot Image Classification on 14 Robustness Benchmark Datasets (test)
80.11ImageNet AccuracyTeCoA
Evaluation results
| Method | Links | ||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| TeCoAModel Backbone=ViT-L/14, Training Constraint (epsilon_t)=2/255, Evaluation Protocol=Clean, Reference Text Embeddings=ImageNet-21k labels2026.03 | 80.11 | 80.67 | 50.08 | 87.53 | 60.69 | 44.36 | 26.06 | 14.04 | 51.8 | 80.12 | 58.43 | 49.89 | 80.02 | 96.08 | 61.42 | — | |
| ET3Model Backbone=ViT-L/14, Training Protocol=TeCoA, Training Constraint (epsilon_t)=2/255, Evaluation Protocol=Clean, Reference Text Embeddings=ImageNet-21k labels2026.03 | 77.79 | 78.98 | 40.74 | 78.5 | 50.55 | 42.5 | 29.94 | 15.21 | 47.99 | 75.4 | 57.29 | 49.97 | 75.91 | 95.79 | 58.33 | -3.09 | |
| TeCoAModel Backbone=ViT-L/14, Training Constraint (epsilon_t)=4/255, Evaluation Protocol=Clean, Reference Text Embeddings=ImageNet-21k labels2026.03 | 74.91 | 78.36 | 37.83 | 79.61 | 50.26 | 38.03 | 22.48 | 11.76 | 38.41 | 74.35 | 54.22 | 49.95 | 76.07 | 93.44 | 55.69 | — | |
| FAREModel Backbone=ViT-L/14, Training Constraint (epsilon_t)=2/255, Evaluation Protocol=Clean, Reference Text Embeddings=ImageNet-21k labels2026.03 | 74.48 | 84.77 | 70.53 | 89.52 | 69.13 | 50.05 | 25.39 | 26.7 | 70.6 | 85.52 | 59.72 | 50.01 | 91.06 | 98.47 | 67.57 | — | |
| ET3Model Backbone=ViT-L/14, Training Protocol=TeCoA, Training Constraint (epsilon_t)=4/255, Evaluation Protocol=Clean, Reference Text Embeddings=ImageNet-21k labels2026.03 | 74.21 | 77.95 | 35.79 | 73.41 | 45.09 | 37.61 | 23.15 | 12.54 | 39.29 | 72.81 | 54.85 | 50 | 75.06 | 94.15 | 54.71 | -0.98 | |
| ET3Model Backbone=ViT-L/14, Training Protocol=FARE, Training Constraint (epsilon_t)=2/255, Evaluation Protocol=Clean, Reference Text Embeddings=ImageNet-21k labels2026.03 | 72.9 | 83.75 | 63.93 | 78.19 | 53.22 | 46.7 | 30.61 | 24.81 | 63.05 | 80.68 | 58.18 | 50.02 | 88.31 | 95.59 | 63.57 | -4 | |
| FAREModel Backbone=ViT-L/14, Training Constraint (epsilon_t)=4/255, Evaluation Protocol=Clean, Reference Text Embeddings=ImageNet-21k labels2026.03 | 70.78 | 84.7 | 63.84 | 77.67 | 56.53 | 43.83 | 18.28 | 21.96 | 58.07 | 80.24 | 56.74 | 50.02 | 87.14 | 96.04 | 61.85 | — | |
| ET3Model Backbone=ViT-L/14, Training Protocol=FARE, Training Constraint (epsilon_t)=4/255, Evaluation Protocol=Clean, Reference Text Embeddings=ImageNet-21k labels2026.03 | 69.86 | 83.68 | 60.34 | 66.23 | 42.89 | 42.98 | 18.2 | 21.48 | 54.82 | 76 | 56.25 | 50.02 | 84.93 | 92.56 | 58.59 | -3.26 | |
| ET3Model Backbone=ViT-L/14, Training Protocol=TeCoA, Training Constraint (epsilon_t)=4/255, Evaluation Protocol=Robust, Adversarial Attack Configuration=Auto-Attack at epsilon_a = 4/255, Reference Text Embeddings=ImageNet-21k labels2026.03 | 54.7 | 66 | 11.5 | 57.7 | 32.6 | 22.4 | 16 | 6 | 21.4 | 50.7 | 40.8 | 51.4 | 63.1 | 85.4 | 41.41 | 8.56 | |
| ET3Model Backbone=ViT-L/14, Training Protocol=TeCoA, Training Constraint (epsilon_t)=2/255, Evaluation Protocol=Robust, Adversarial Attack Configuration=Auto-Attack at epsilon_a = 4/255, Reference Text Embeddings=ImageNet-21k labels2026.03 | 47.4 | 63.5 | 13.1 | 51.2 | 31.4 | 21.6 | 22.6 | 8.8 | 24 | 47 | 40.5 | 46.3 | 59.8 | 82.2 | 39.96 | 12.3 | |
| TeCoAModel Backbone=ViT-L/14, Training Constraint (epsilon_t)=4/255, Evaluation Protocol=Robust, Adversarial Attack Configuration=Auto-Attack at epsilon_a = 4/255, Reference Text Embeddings=ImageNet-21k labels2026.03 | 44.5 | 60.9 | 8.5 | 37.1 | 21.5 | 16.5 | 6.4 | 2.2 | 12.6 | 41.9 | 32.8 | 45.7 | 55 | 74.3 | 32.85 | — | |
| ET3Model Backbone=ViT-L/14, Training Protocol=FARE, Training Constraint (epsilon_t)=4/255, Evaluation Protocol=Robust, Adversarial Attack Configuration=Auto-Attack at epsilon_a = 4/255, Reference Text Embeddings=ImageNet-21k labels2026.03 | 42.1 | 69 | 17.8 | 47.1 | 30.3 | 24.5 | 13.7 | 6.5 | 20.9 | 46.3 | 38.3 | 52.3 | 57.5 | 80.6 | 39.06 | 6.29 | |
| TeCoAModel Backbone=ViT-L/14, Training Constraint (epsilon_t)=2/255, Evaluation Protocol=Robust, Adversarial Attack Configuration=Auto-Attack at epsilon_a = 4/255, Reference Text Embeddings=ImageNet-21k labels2026.03 | 37 | 57.4 | 6.4 | 31 | 17.9 | 14.7 | 7.8 | 1 | 9.6 | 36.6 | 30.9 | 17.4 | 50.4 | 69.1 | 27.66 | — | |
| FAREModel Backbone=ViT-L/14, Training Constraint (epsilon_t)=4/255, Evaluation Protocol=Robust, Adversarial Attack Configuration=Auto-Attack at epsilon_a = 4/255, Reference Text Embeddings=ImageNet-21k labels2026.03 | 34.8 | 64.2 | 12.7 | 34.8 | 20.2 | 17.5 | 11.1 | 3 | 12.2 | 40.5 | 30.6 | 52.3 | 50.6 | 74.3 | 32.77 | — | |
| ET3Model Backbone=ViT-L/14, Training Protocol=FARE, Training Constraint (epsilon_t)=2/255, Evaluation Protocol=Robust, Adversarial Attack Configuration=Auto-Attack at epsilon_a = 4/255, Reference Text Embeddings=ImageNet-21k labels2026.03 | 27.1 | 56.4 | 12.7 | 49.1 | 27.8 | 20 | 19.2 | 5.8 | 15.4 | 37 | 32.1 | 36.3 | 38.8 | 72.1 | 32.13 | 11.74 | |
| FAREModel Backbone=ViT-L/14, Training Constraint (epsilon_t)=2/255, Evaluation Protocol=Robust, Adversarial Attack Configuration=Auto-Attack at epsilon_a = 4/255, Reference Text Embeddings=ImageNet-21k labels2026.03 | 17.8 | 46.4 | 5 | 25.7 | 14.2 | 11.6 | 0.4 | 0.9 | 7.1 | 25.6 | 22.1 | 19.1 | 28.1 | 61.5 | 20.39 | — |