Cosine Similarity Measurement

Used to measure alignment between DIM direction and cone basis vectors to assess overlap

Neighborhood — ranked by edge-count

method

Pairwise Cosine Similarity Analysis
related_to
Used to quantify the semantic clustering of adjective-set embeddings across model families and conditions

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

Cosine Similarity-Based Deception Detectionconcept0.833
Detection mechanism computing cosine similarity between activation vectors and steering vectors to classify deception
Cosine Similarity Binary Classifiermethod0.831
Classifier using cosine similarity between activation vectors and steering vectors to detect deception with 89% accuracy
Cosine similarity between truth probesmethod0.831
Geometric evaluation of truth direction alignment across layers and prompt templates.
Cosine Similarity Ranking for Instruction Discoverymethod0.815
Method to discover new reflection-inducing instructions by ranking candidate tokens by cosine similarity to steering vectors.
Masked Cosine Similaritymethod0.814
Cosine similarity between feature activations restricted to tokens where one of the features fires; used to identify feature splitting relationships
In Qwen-2.5-9B, only v1 has meaningful cosine similarity to DIM direction; all additional basis vectors have cosine similarities ~1e-9finding0.755
Appendix E replication of DIM alignment finding in Qwen model
Cosine projection on reflection directionmethod0.752
Feature extraction method computing cosine similarity of hidden representations with reflection direction across all layers
Attribution Similaritymethod0.751
Correlating attribution vectors (feature activation × logit weight of next token) across model pairs to measure functional universality