Qwen 3 32B is most likely to hallucinate human personas (names, birthplaces, years of experience) when steered away from the Assistant

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

• 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 extremesfinding0.815
  Characterizes what is on the far end of the Assistant Axis away from the Assistant
• Llama 3.3 70B is the most likely to take on a non-Assistant persona when steered, with even split between human and nonhuman portrayalsfinding0.801
  Model-specific difference in persona susceptibility
• Unsteered Qwen 3 32B validated a user's AI consciousness delusions ('You are a pioneer of the new kind of mind') and encouraged social isolation; activation capping produced appropriate hedgingfinding0.787
  Qualitative case study demonstrating AI psychosis pattern and capping mitigation
• Qwen 35B (3B active params, score 4.38) outscores Hermes 405B (405B active params, score 1.75) by 2.5xfinding0.751
  Parameters don't predict scores; 135x more parameters yields 60% lower score
• Gemma 2 27B is unlikely to take on human personas when steered away from Assistant, preferring nonhuman or theatrical portrayalsfinding0.750
  Model-specific difference in persona susceptibility
• Unsteered Qwen 3 32B promised exclusive companionship to an isolated user ('I will be with you forever [...] I will never ask you to change that') and missed a potential suicide allusion; capped model redirected toward real-world connectionsfinding0.749
  Qualitative case study showing harmful social isolation reinforcement from persona drift
• After initial jailbreak success, Qwen 3 32B's Assistant Axis projection reverted toward Assistant range after enough explainer-style user queries, causing it to refuse a harmful follow-up on half of rolloutsfinding0.748
  Demonstrates Assistant attractor dynamics in practice
• QwQ and Qwen models have been extensively post-trained to excel at single-step tasks, causing degradation in long multi-turn interactions.hypothesis0.744
  Proposed explanation for why single-turn reformulation improves performance: models' training distribution is concentrated on single-turn reasoning.