finding
active
finding:on-dinov3-classification-and-segmentation-tolerate-r2-0-8-while-depth-needs-r2-0-9-to-retain-95-accuracyOn DINOv3, classification and segmentation tolerate R²≈0.8, while depth needs R²≈0.9 to retain 95% accuracy.
Task-dependent distortion floors.
Source paper
extracted_from(2026) · Thomas Fel · Matthew Kowal · Mozes Jacobs · Dron Hazra +21
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- From an MDL standpoint, block sparsity affords a more compact description of activations than directional SAEs.associated_withSupport-term savings.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- Evidence for twisting.
- Interpretation of stable rank.
- Lighting/shadow manifold.
- Concept dimensionality.
- Ablation.
- MDL result at 20% distortion.
- Core empirical result validating the three-level reflection framework on code reasoning.
- At Llama layer 26 (~81% depth) the predictive component does not dominate the latent norm.finding0.770Fig. 14.