method
active
method:feature-neighborhood-exploration-via-cosine-similarity-of-decoder-weightsFeature neighborhood exploration via cosine similarity of decoder weights
Identifying related features by cosine distance in SAE decoder space.
Neighborhood — ranked by edge-count
Papers (1)
paper
Methods (1)
method
- Detects tiling in real SAEs.
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.
- Gemma Scope decoder nearest-neighbor cosine similarities skewed above random but few extremely highfinding0.816Fig 7.
- Classifier using cosine similarity between activation vectors and steering vectors to detect deception with 89% accuracy
- Method to discover new reflection-inducing instructions by ranking candidate tokens by cosine similarity to steering vectors.
- Geometric analysis tracking how persona vector directions evolve across transformer layers to identify the transition layer
- Larger rotations between shared features skew decoder cosine similarities downward in the toy modelfinding0.777Supports 'same features, used differently' interpretation.
- Third empirical pattern.
- Many latents tiling a low-dimensional manifold would show decoder cosine similarities near 1hypothesis0.761Motivates Appendix E.
- Detection mechanism computing cosine similarity between activation vectors and steering vectors to classify deception