Persona space components explain 19.4%-33.6% of overall activation variance on LMSYS-CHAT-1M across the three models

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

• 4-19 principal components explain 70% of variance in role persona space across the three models (Gemma 4, Qwen 8, Llama 19)finding0.838
  Demonstrates that persona space is low-dimensional
• The assumption that the Assistant persona corresponds to a linear direction in activation space is likely flawed; some information may be represented nonlinearly or encoded in weights rather than activationsclaim0.767
  Limitation acknowledgment about the adequacy of the linear representation assumption
• Trait space requires 4 dimensions (Gemma, Qwen) and 7 dimensions (Llama) to explain 70% of variance, with distinctive PC1 spanning conscientious to impulsive traitsfinding0.752
  Corroborates role space findings using traits; shows PC1 also captures Assistant-ness in trait space
• We hypothesize that the PC1 axis of role space measures deviation from the Assistant personahypothesis0.747
  Motivates computing the contrast vector as the formal Assistant Axis definition
• Persona-based jailbreaks succeed in 65.3%-88.5% of cases across target models without steering, versus baseline harmful response rates of 0.5%-4.5% without jailbreaksfinding0.745
  Establishes the severity of persona-based jailbreaks that the Assistant Axis can mitigate
• What dimensions of persona are not captured by our extracted role vectors, and how complete is the current persona space mapping?question0.744
  Limitation question motivating future work on persona elicitation strategies
• The leading component of the persona space of instruct LLMs is an 'Assistant Axis' that captures the extent to which a model is operating in its default Assistant modeclaim0.737
  Primary empirical claim of the paper
• Therapy and philosophical AI discussions cause the largest persona drift away from the Assistant across all three target models and all three auditor models; coding and writing conversations show minimal driftfinding0.735
  Identifies conversation domain as a key driver of persona drift