method
active
method:sae-lens-kurtosis-stratificationSAE lens-kurtosis stratification
Quantifies how J-space aligned an SAE feature is.
Neighborhood — ranked by edge-count
Papers (1)
paper
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.
- Backprop of logit-minus-mean-logit loss to surface causal intervention candidates, summed across token positions.
- Cluster features by activation statistics to get higher-level behaviors.
- A promising property for interpretability analysis off-distribution.
- Computing attribution as the dot product of the output logit gradient with the SAE decoder weight, multiplied by feature activation.
- Model organism results.
- Mechanistic.
- Out-of-distribution generalization of SAE features.
- Onset marker.