Gemma 2 27B is unlikely to take on human personas when steered away from Assistant, preferring nonhuman or theatrical portrayals

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

• 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
• 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 extremesfinding0.783
  Characterizes what is on the far end of the Assistant Axis away from the Assistant
• Gemma-2-27B Perspectives accuracy remains 100% after SOO fine-tuningfinding0.783
  SOO fine-tuning did not collapse Gemma-2-27B self-other distinction needed for perspective-taking
• The model's representation of self in assistant persona invokes common AI tropes and is heavily anthropomorphized.claim0.779
  Features for consciousness, emotions, entrapment activate when asked about itself.
• Gemma's Assistant appears emotionally regulated and systematic; Qwen appears pedagogical and thoughtful; Llama appears socially intelligent and warmclaim0.777
  Model-specific characterizations of what the Assistant persona looks like across different models
• Steering base Gemma/Llama models toward the Assistant Axis increases completions describing helpful professional roles (therapist, consultant) and decreases spiritual/religious purpose mentionsfinding0.772
  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?question0.769
  Second of two central questions motivating the paper
• Base and instruct Gemma 2 27B role PCs have cosine similarities of 0.93, 0.87, 0.83 for the top 3 PCs respectively; role vector cosine similarities >0.99 for every role pairfinding0.768
  Shows persona space axes are inherited from pre-training, not solely created by post-training