Regularization-based prevention of persona shifts fails because the model encodes the trait through alternative directions in activation space to satisfy the next-token prediction loss

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

• Train-time regularization penalizing projection changes along persona directions is ineffective at preventing persona shiftsfinding0.830
  Negative result showing model bypasses regularization by encoding trait through alternative directions
• We hypothesize that geometric explanations for persona-safety interactions are tractable and that per-model geometric audits can predict activation-steering vulnerability from trait-refusal cosine alignment.hypothesis0.804
  Forward-looking claim about the utility of the trait refusal alignment framework as a general tool
• During pre-training, models learn a variety of personas including misaligned ones; fine-tuning on narrowly incorrect datasets amplifies misaligned personas because they reduce training loss, causing broadly misaligned behaviorclaim0.803
  The paper's central mechanistic explanation of why narrow fine-tuning causes broad misalignment
• Persona representations originate in pretraining, not alignment; they form without explicit supervision as useful features for next-token predictionclaim0.802
  Core interpretive claim providing mechanistic explanation for early persona formation
• Per-model activation-steering safety verification is necessary because prompt-side persona rankings do not predict geometric vulnerability.claim0.791
  Practical implication drawn from the prosocial persona paradox finding
• Persona alignment improves when the model is exposed to contextual cues explicitly relevant to the assigned persona, as in structured questionnaire tasksclaim0.787
  Supported by GPT-4o achieving highest ACCatom in Questionnaire vs Essay and SMP tasks
• Both inference-time and preventative steering mitigate persona shifts without reversing domain-specific effects learned during finetuningfinding0.786
  Shows steering is behaviorally targeted: suppresses general persona drift while preserving intended narrow-domain learning
• Single-method persona safety evaluation is incomplete: prompting and activation steering expose different, architecture-dependent vulnerability profiles, and testing with only one method can miss a model's dominant failure mode.claim0.785
  Central thesis of the paper