Cosine Similarity of Persona Vectors

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

• Cosine similarity between earliest extractable checkpoint persona vector and final-pretraining direction starts at approximately 0.3finding0.816
  Quantifies geometric distance of early persona directions from final direction
• Differences in persona vector geometry across checkpoints correspond to meaningfully different versions of the personaclaim0.813
  Connects geometric refinement findings to semantic/facet changes in persona expression
• In Qwen-2.5-9B, only v1 has meaningful cosine similarity to DIM direction; all additional basis vectors have cosine similarities ~1e-9finding0.803
  Appendix E replication of DIM alignment finding in Qwen model
• Pairwise cosine similarities between Description, Narration, and Dialogue evil vectors are all below 0.5finding0.799
  Shows different elicitation strategies recover qualitatively distinct persona directions
• Cosine similarity between truth probesmethod0.797
  Geometric evaluation of truth direction alignment across layers and prompt templates.
• Masked Cosine Similaritymethod0.794
  Cosine similarity between feature activations restricted to tokens where one of the features fires; used to identify feature splitting relationships
• There may exist a natural 'persona basis' characterizing the space of all model personas, with correlations between persona vectors predicting co-expression of traitshypothesis0.793
  Open question proposed by authors for future work on the dimensionality and structure of persona space
• Persona Vector Arithmeticconcept0.790
  Compositional combination of persona vectors via algebraic operations on activation directions