finding
active
finding:bsf-codes-surpass-the-sae-in-best-single-concept-f1-on-imagenet-1k-at-every-sparsity-gap-widening-with-bBSF codes surpass the SAE in best single-concept F1 on ImageNet-1k at every sparsity, gap widening with b.
Class selectivity.
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
- Motivation for lifting atoms to blocks.
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.
- BSF codes improve on the basic probe for classification and approach it for segmentation and depth.finding0.777Linear probing.
- At k=64, BSF concept maps are roughly five times smoother (total variation) at b=4 than at b=1.finding0.765Spatial coherence.
- MDL result at 20% distortion.
- Authors' experience-based claim.
- Featurizers whose atoms are blocks of directions rather than single directions.
- A promising property for interpretability analysis off-distribution.
- The primary novel framework introduced in the paper for learning facet-level personality control vectors
- Figure 15.