Dino
Benchmarks
| Task Name | Dataset Name | SOTA Result | Trend | |
|---|---|---|---|---|
| SAE Interpretability and Faithfulness Evaluation | DINOv2 ViT-B 14 Layer 12 activations | LLM Rank6 | 12 | |
| Feature Reconstruction | dino Original Target Open Images (test) | R^2 (variance-weighted)0.69 | 9 | |
| SAE Interpretability and Faithfulness Evaluation | DINOv2 ViT-L/14 activations | LLM Rank6 | 6 | |
| SAE Interpretability and Faithfulness Evaluation | DINOv2 ViT-S 14 activations | LLM Rank6 | 6 | |
| Sparse Autoencoder Evaluation | DINOv2 S activations | Variance Explained94.2 | 6 | |
| Ownership verification | DINOv2 | Average Watermark Detection Rate100 | 6 | |
| Disparity Estimation | Dino light field (synthetic) | MSE0.0041 | 6 | |
| Training and Verification System Efficiency | DINOv2 Giant (2048 samples) | Training Time (s)24.1 | 5 | |
| Neural Shape Representation | Dino | Chamfer Distance1.48 | 4 | |
| Visual Mechanistic Interpretability | DINO SAE features v3 | Interpretability Score (Human)2.92 | 3 | |
| Feature Visualization | DINOv3 Random 100 features | s(x)16.51 | 3 | |
| Feature Visualization | DINO Feature 4831 v3 | s(x)16.58 | 3 | |
| Feature Visualization | DINOv3 (Feature 9863) | s(x)16.53 | 3 |
Showing 13 of 13 rows