concept
active
concept:class-centroid-difference-initializationClass-Centroid Difference Initialization
Method of initializing control vectors as difference between mean SAE codes for positive and negative samples
Neighborhood — ranked by edge-count
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.
- Novel calibration of injection strength as the distance from centroid midpoint to centroid; enables meaningful cross-layer comparison of alpha values
- Weaker geometry-steering correlation.
- Used to extract trait and style directions.
- L2 distance between centroids compared across manifolds, RSA-like.
- Steering effect between concepts correlates with My centroid distance, r=.65 (p<.001) for Emotionsfinding0.692Geometry predicts steering.
- Steering vector from mean activation difference between contrasting datasets.
- Method for extracting linear directions by subtracting mean activations of contrastive groups; used to define the Assistant Axis
- Formal consequence of Belrose et al. (2023) Theorem G.1 connecting mass-mean probing to optimal linear concept erasure