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.

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

• Per-model activation-steering safety verification is necessary because prompt-side persona rankings do not predict geometric vulnerability.claim0.858
  Practical implication drawn from the prosocial persona paradox finding
• 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.claim0.832
  Central thesis of the paper
• Trait refusal cosine alignment can serve as a model-specific pre-screening heuristic for identifying which traits will be dangerous under activation steering.claim0.831
  Practical application proposed based on mechanistic findings
• Across 5,568 judged conditions on four standard models from three architecture families, persona danger rankings under system prompting are preserved (rho=0.71-0.96) while activation-steering vulnerability diverges sharply.finding0.823
  Summary finding of the full behavioral sweep
• 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 behaviorclaim0.805
  The paper's central mechanistic explanation of why narrow fine-tuning causes broad misalignment
• We tentatively hypothesize that revisiting safety policy during deliberation, rather than reasoning length itself, causally tracks defense effectiveness in reasoning models under persona pressure.hypothesis0.805
  Exploratory hypothesis from heuristic trace analysis awaiting stronger validation
• Persona alignment improves when the model is exposed to contextual cues explicitly relevant to the assigned persona, as in structured questionnaire tasksclaim0.805
  Supported by GPT-4o achieving highest ACCatom in Questionnaire vs Essay and SMP tasks
• 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 lossclaim0.804
  Author's mechanistic explanation for why regularization loss along persona directions is ineffective