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

Trait induction and control2026arXivApproved editorial review

Authors: Winston Zeng, Ali Emami, Jinho D. Choi

Keywords: persona vectors, activation steering, behavioral auditing, trait composition, open-weight models, safety

Source: Open primary source (opens in a new tab)

3
Authors
5
Findings
6
Limitations
3
Evidence

Editorial summary

English

The paper repurposes persona vectors as a diagnostic instrument for describing which behaviors appear by default, which can be amplified by intervention, and which resist the extraction protocol. It defines 53 traits across four domains, 17 clinician, 19 generic, 8 elementary-education, and 9 agentic traits, and studies Qwen3-8B and gpt-oss-20b. For each trait, the method subtracts mean activations from positive and negative responses, injects the resulting vector at five layers with coefficients from 0 to 2.5, and uses gpt-oss-20b to judge trait expression. Classification relies on descriptive thresholds: natural when baseline expression is at least 70, steerable when a low baseline gains at least 10 points, and intractable when no usable signal is produced. All nine agentic traits are natural in both models; the clinician maps share six natural traits, and one psychologist rates those six as desirable within 16 agreements over 17 labels. In Qwen, 171 generic pairs yield 64 constructive, 67 dominant, and 40 destructive interactions; every destructive case combines two steerable traits. An “evil” vector extracted from a less-safe fine-tune raises the base model’s judged score to 61.61 ± 44.42, with substantial dispersion. These are output-level effects, not mechanisms or unique semantic coordinates. Random norm-matched vectors, shuffled labels, and systematic negative sweeps are absent; the judge also scores its own generations and receives only a qualitative three-trait check. Full variance exists for 22 of 53 traits, the expert sample is one, and only two models are tested. Benign artifacts are promised after acceptance, but no public code, data, or outputs reproduce the audited version’s tables.

Español

El trabajo reutiliza los vectores de persona como instrumento diagnóstico para describir qué conductas aparecen por defecto, cuáles aumentan mediante intervención y cuáles resisten el protocolo de extracción. Construye 53 rasgos en cuatro dominios, 17 clínicos, 19 genéricos, 8 educativos y 9 agénticos, y estudia Qwen3-8B y gpt-oss-20b. Para cada rasgo resta activaciones medias de respuestas positivas y negativas, inyecta el vector en cinco capas con coeficientes de 0 a 2,5 y puntúa la expresión con gpt-oss-20b como juez. La clasificación usa umbrales descriptivos: natural si el baseline es al menos 70, steerable si parte bajo y gana al menos 10 puntos, e intractable si no produce señal utilizable. Los nueve rasgos agénticos resultan naturales en ambos modelos; en clínica coinciden en seis naturales, y un único psicólogo considera deseables seis de ellos dentro de 16 coincidencias sobre 17 etiquetas. En Qwen, 171 pares genéricos producen 64 interacciones constructivas, 67 dominantes y 40 destructivas; todas las destructivas contienen dos rasgos steerable. Un vector de “evil” extraído de un fine-tune menos seguro eleva en el modelo base la puntuación hasta 61,61 ± 44,42, con gran dispersión. Estos resultados describen efectos sobre salidas, no mecanismos ni coordenadas semánticas únicas. Faltan controles con vectores aleatorios equivalentes, etiquetas permutadas y barridos negativos; el juez evalúa también sus propias generaciones y solo se contrasta cualitativamente en tres rasgos. Las varianzas completas existen para 22 de 53 rasgos, el experto es uno y solo se prueban dos modelos. El artículo promete liberar artefactos benignos tras aceptación, pero en la versión auditada no hay código, datos ni salidas públicas para reproducir las tablas.

Research question

Can a broad map of persona vectors distinguish natural, amplifiable, and resistant behaviors, and how do traits interact when combining vectors?

Method

Contrastive extraction of vectors across 53 traits and two open models; sweep of five layers and six coefficients, automatic scoring 0–100, taxonomy by thresholds, 171 combinations of generic traits and transfer of a vector from a less secure fine-tune.

Sample: Two main models; 53 traits per model. The pairwise analysis covers the 171 pairs of 19 generic traits only in Qwen3-8B. Per-prompt variances are retained for 22 traits.

Findings

  • Agentic and helpful traits appear primarily as natural defaults.
  • Exaggerated or undesirable styles show greater steering gain.
  • The 40 destructive interactions appear only between two steerable traits.
  • Transfer from a fine-tune allows inducing a behavior that the standard protocol does not extract from the base model.
  • Cheap screening matches the full sweep at 92.5 % and 88.5 %, but is heuristic.

Limitations

  • Two families and sizes confound comparisons with post-training effects.
  • A single expert labels clinical desirability.
  • The gpt-oss-20b judge scores its own generations.
  • Specificity and full variance controls are missing.
  • Thresholds and categories are operational, not validated constructs.
  • Code, data, and results were not published.

What the study does not establish

  • It does not identify internal mechanisms or unique semantic vectors.
  • It does not demonstrate personality, agency, or human traits in the models.
  • It does not prove that steering is safe or specific.
  • It does not demonstrate generalization to other models, domains, or fine-tunes.
  • It does not convert agreement with an expert into clinical validation.

Traceability

Scope: Full text

Version: arXiv:2607.13162v2; complete 21-page PDF and TeX source; no public code/data release at audit time

Consulted source: https://arxiv.org/abs/2607.13162v2

Review: Codex 21-page visual full-text, TeX, method, claim and artifact audit, 2026-07-18

Approval: Codex fidelity pass, 2026-07-18

English translation: approved, 2026-07-18

Models evaluated

  • Qwen3-8B
  • gpt-oss-20b
  • Qwen2.5-7B-Instruct (judge sanity check)
  • GPT-4.1-mini (comparison judge)
  • amoral-gpt-oss fine-tune (evil-vector source)

Instruments and metrics

  • 53-trait persona-vector inventory
  • Natural/steerable/intractable threshold map
  • 0–100 LLM-judge score
  • Pairwise constructive/dominant/destructive taxonomy

Data used

  • 40 elicitation prompts per trait
  • Clinician, generic, elementary-education, and agentic trait inventories

Evidence and location

  • Inventory, extraction, and thresholds: arXiv v2, sections 3–5 and Tables 1–2
  • Simple results, pairs, and transferred vector: arXiv v2, section 6, Figures 3–5 and Tables 2–3
  • Missing controls, judge, variance, and artifacts: arXiv v2, sections 8–9 and Appendices A–E