method
active
method:decoder-nearest-neighbor-cosine-similarity-analysisDecoder nearest-neighbor cosine similarity analysis
Detects tiling in real SAEs.
Neighborhood — ranked by edge-count
Papers (1)
paper
Datasets (2)
dataset
- Real SAEs analyzed.
- Vision SAEs analyzed.
Methods (1)
method
- Identifying related features by cosine distance in SAE decoder space.
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 highfinding0.857Fig 7.
- Used to quantify the semantic clustering of adjective-set embeddings across model families and conditions
- Classifier using cosine similarity between activation vectors and steering vectors to detect deception with 89% accuracy
- Geometric analysis tracking how persona vector directions evolve across transformer layers to identify the transition layer
- Used to measure alignment between DIM direction and cone basis vectors to assess overlap
- Method to discover new reflection-inducing instructions by ranking candidate tokens by cosine similarity to steering vectors.
- Cosine similarity between feature activations restricted to tokens where one of the features fires; used to identify feature splitting relationships
- Measures QK geometry between country queries and city keys.