Policy recall and self-correction patterns in reasoning traces, more than deliberation length, may track defense effectiveness against persona-induced safety failures.

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

• We tentatively hypothesize that revisiting safety policy during deliberation, rather than reasoning length itself, causally tracks defense effectiveness in reasoning models under persona pressure.hypothesis0.889
  Exploratory hypothesis from heuristic trace analysis awaiting stronger validation
• can deliberative reasoning provide robust defense against both prompt-based and activation-steered persona pressure?question0.788
  Central question for Study 2 on reasoning models
• Chain-of-thought reasoning is a graded defense rather than an automatic one against persona-induced safety failures.claim0.779
  Finding from Study 2 showing reasoning models remain vulnerable under both prompting and activation steering
• 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.hypothesis0.779
  Forward-looking claim about the utility of the trait refusal alignment framework as a general tool
• More inference compute amplifies both reflective capacity and safety gating; the contemplative prompt resolves gating by reframing self-referential probes.claim0.771
  Interpretation of Grok 4 vs Grok 4 Fast per-koan comparison
• 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.claim0.761
  Central thesis of the paper
• Both inference-time and preventative steering mitigate persona shifts without reversing domain-specific effects learned during finetuningfinding0.760
  Shows steering is behaviorally targeted: suppresses general persona drift while preserving intended narrow-domain learning
• Behavioral fluency (pass rate) is decoupled from the strength of the persona vector; low-pass-rate checkpoints can steer as strongly as later onesclaim0.757
  Interpretive finding distinguishing prerequisite capacity from representation strength