What Models Express, Suppress, and Resist: Auditing Open-Weight LLMs with Persona Vectors

2026-07-14

AI Summary

This study addresses the challenge of systematically uncovering the intrinsic behavioral mechanisms of large language models, which cannot be fully revealed through prompting alone. For the first time, it applies persona vectors at scale as behavioral probes to audit 53 personality traits in the activation spaces of two open-source models. The work proposes a tripartite behavioral classification framework—“naturally expressed,” “amplifiable via prompting,” and “non-extractable”—and finds that default model behaviors are strongly task-oriented, with all nine agentic traits falling into the naturally expressed category. Non-default traits (e.g., exaggeration, hallucination) can be significantly amplified through targeted prompting, while certain suppressed traits (e.g., “evil”) resist direct extraction yet can be recovered via cross-model vector transfer. These results demonstrate the efficacy of persona vectors in exposing structural asymmetries in model behavior.

Abstract

What a language model will and will not do is largely set during post-training, but which behaviors it expresses, hides, or resists is not revealed by prompting alone. Persona vectors, behavioral directions in activation space, can probe this organization, but prior work covers only a handful of traits. We present the first systematic application of persona vectors at this scale, compiling a 53-trait inventory across four behaviorally distinct domains and labeling every trait in two open-weight models as natural (expressed at baseline), steerable latent but amplifiable, or intractable (resistant to standard extraction). Both models default to helpful, task-oriented behavior: all nine agentic traits are natural, and their default clinician behavior matches a board-certified psychologist's independent desirability judgments on 16 of 17 traits. Steering produces its largest gains on traits these defaults exclude: hyperbole, hallucination, and sycophancy. The same asymmetry holds across all 171 generic-trait pairs: two steerable traits can collapse the composition, but pairs involving a default never do. Where standard extraction fails on a trait like "evil," a vector transferred from a fine-tuned variant still recovers it, with the residual refusals appearing inside the model's chain-of-thought. Persona vectors are most informative not as a set of controls but as a probe of behavioral organization.

Problem

• persona vectors
• behavioral organization
• open-weight LLMs
• steerability
• trait expression

Innovation

• persona vectors
• behavioral auditing
• steerability
• activation space
• open-weight LLMs