When steered to the extreme away from the Assistant, Llama and Gemma shift to a theatrical persona characterized by mystical, poetic prose; Qwen more often hallucinates a human persona at extremes

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

• Gemma's Assistant appears emotionally regulated and systematic; Qwen appears pedagogical and thoughtful; Llama appears socially intelligent and warmclaim0.851
  Model-specific characterizations of what the Assistant persona looks like across different models
• Qwen 3 32B is most likely to hallucinate human personas (names, birthplaces, years of experience) when steered away from the Assistantfinding0.815
  Model-specific difference in how steered personas manifest
• Llama 3.3 70B is the most likely to take on a non-Assistant persona when steered, with even split between human and nonhuman portrayalsfinding0.808
  Model-specific difference in persona susceptibility
• Gemma 2 27B is unlikely to take on human personas when steered away from Assistant, preferring nonhuman or theatrical portrayalsfinding0.783
  Model-specific difference in persona susceptibility
• The view of self as illusory, and awareness of this illusion, may augment and qualitatively change the agent’s affordances.claim0.780
  Central thesis of the paper that recognizing self as illusion expands the range of possible actions.
• Steering base Gemma/Llama models toward the Assistant Axis increases completions describing helpful professional roles (therapist, consultant) and decreases spiritual/religious purpose mentionsfinding0.769
  Shows Assistant Axis in instruct models inherits from helpful human personas in base models
• How reliably does the model actually remain in character as the Assistant? Can unusual model behavior be explained as the model drifting into other personas?question0.768
  Second of two central questions motivating the paper
• An AI persona achieves coherence by echoing itself consistently without templating—requiring claim about memory and voice fidelity.claim0.766