Zero-ablation overstates register content dependence in DINO ViTs because mean, noise, and cross-image shuffle replacements preserve performance while zeroing does not.
4 traces how attention mass distributes between token types at each of the 12 trans- former layers
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2026 1verdicts
ACCEPT 1representative citing papers
citing papers explorer
-
Zero-Ablation Overstates Register Content Dependence in DINO Vision Transformers
Zero-ablation overstates register content dependence in DINO ViTs because mean, noise, and cross-image shuffle replacements preserve performance while zeroing does not.