finding
active
finding:optimal-block-dimension-is-3-for-grassmannian-but-nearer-k-1-2-3-for-the-untied-block-featurizer-group-lasso-tolerates-largest-blocks-at-higher-residual-costOptimal block dimension is ≈3 for Grassmannian but nearer k∈{1,2,3} for the untied Block featurizer; Group Lasso tolerates largest blocks at higher residual cost.
Featurizer-specific optima.
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
- Support-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.
- Block and Grassmannian featurizers keep utilization ≈0.6–0.66 at k=16, while Group Lasso falls to ≈0.17.finding0.855Per-feature occupancy.
- What is the optimal block dimension and what is the typical dimensionality of a concept?question0.779Motivates stable-rank study.
- MDL result at 20% distortion.
- Interpretation of stable rank.
- Reconstruction cannot rank featurizers.
- Corroborates role space findings using traits; shows PC1 also captures Assistant-ness in trait space
- Establishes generalizability of the core difficulty-boundary finding across model families.
- On DINOv3, classification and segmentation tolerate R²≈0.8, while depth needs R²≈0.9 to retain 95% accuracy.finding0.737Task-dependent distortion floors.