Study reveals zero-ablation technique significantly overestimates how critical register tokens are in DINO vision transformers

Zero-ablation (replacing tokens with zeros) showed up to 36.6pp drops in classification and 30.9pp in segmentation, suggesting registers are essential to DINOv2+registers and DINOv3

Alternative replacement methods—mean-substitution, noise-substitution, and cross-image register-shuffling—maintained near-baseline performance within ~1pp across classification, correspondence, and segmentation tasks

Cosine similarity analysis confirmed replacement controls genuinely perturb internal representations, but zeroing causes disproportionately larger perturbations, explaining why it alone degrades performance

Findings suggest performance depends on having plausible register-like activations rather than exact register content, challenging assumptions about register indispensability