finding
active
finding:bsf-codes-improve-on-the-basic-probe-for-classification-and-approach-it-for-segmentation-and-depthBSF codes improve on the basic probe for classification and approach it for segmentation and depth.
Linear probing.
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 surpass the SAE in best single-concept F1 on ImageNet-1k at every sparsity, gap widening with b.finding0.777Class selectivity.
- Core idea summary from Figure 1.
- If BSFs capture concepts, moving within a block should manipulate behavior according to concept geometry.hypothesis0.755Motivates diffusion steering.
- Featurizers whose atoms are blocks of directions rather than single directions.
- Motivation for causal evaluation over purely behavioural probing accuracy
- At k=64, BSF concept maps are roughly five times smoother (total variation) at b=4 than at b=1.finding0.737Spatial coherence.
- Key interpretive claim from Case Study II distinguishing probe accuracy from causal relevance
- On DINOv3, classification and segmentation tolerate R²≈0.8, while depth needs R²≈0.9 to retain 95% accuracy.finding0.727Task-dependent distortion floors.