Researchers present the first systematic application of persona vectors to audit open-weight language models, compiling a 53-trait inventory across four behaviorally distinct domains. They label every trait in two models as natural (expressed at baseline), steerable latent but amplifiable, or intractable (resistant to standard extraction).
- Both models default to helpful, task-oriented behavior, with all nine agentic traits classified as natural.
- Default clinician behavior matches a board-certified psychologist's independent desirability judgments on 16 of 17 traits.
- Steering produces its largest gains on traits excluded by defaults, specifically hyperbole, hallucination, and sycophancy.
- An asymmetry exists across all 171 generic-trait pairs: two steerable traits can collapse the composition, but pairs involving a default never do.
- Vectors transferred from fine-tuned variants recover intractable traits like "evil," with residual refusals appearing inside the model's chain-of-thought.
The study concludes that persona vectors serve primarily as a probe of behavioral organization rather than a set of controls.