Persona Cartography: Charting Language Model Personality Traits in Weight Space

Trait induction and control2026arXivApproved editorial review

Authors: Luke Baines, Anton Gonzalvez Hawthorne, Mariia Koroliuk, Irakli Shalibashvili, Clément Dumas, Konstantinos Voudouris, David Demitri Africa

Keywords: OCEAN / Big Five, LoRA personality steering, Constitution-guided distillation, Trait scaling and composition, Capability trade-offs, Safety and over-refusal, TIDE factors, Exploratory factor analysis, Construct validity, Reproducibility

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

7
Authors
19
Findings
26
Limitations
6
Evidence

Editorial summary

English

This preprint asks whether an LLM persona can be represented and edited as a position in trait space. For Openness, Conscientiousness, Extraversion, Agreeableness and Neuroticism (OCEAN), it trains amplifier and suppressor LoRAs plus a neutral control. Twelve-statement constitutions define each direction; a teacher generates DPO preferences, another LoRA is trained by SFT on 12,000 self-reflections and conversations, and the two are merged. Models are Llama-3.1-8B-Instruct, Qwen3-8B/32B and Gemma-3-4B/12B/27B. All six share only TRAIT and MMLU; the full stack, open judges, most safety tasks and factor analysis center on Llama-3.1-8B, except frustration on Gemma-3-27B.

The strongest result is operational and bounded. Over moderate ranges, many LoRAs move primarily their target score roughly monotonically; scaling grades the effect and mixtures are approximately additive. Axes are neither orthogonal nor perfectly invertible: off-target movement is visible, negative scaling does not always reverse direction, and compositions leave residuals. At strong scales, wrong, missing and unparsable answers appear and MMLU, GSM8K and TruthfulQA decline. Some five-LoRA mixtures collapse MMLU, while coefficient magnitude poorly predicts loss. Capability preservation holds only for moderate ranges and specific configurations.

Human validation is narrower than “human-validated panel” suggests. Three authors scored 33-36 author-written items for agreeableness, neuroticism and coherence; openness, conscientiousness and extraversion have no direct human gold set. TRAIT was designed for humans, but here it is answered through forced-choice log probabilities, only 300 of 1,000 items per trait are used, and items below 0.75 valid-choice mass are rejected. This measures conditioned responses, not stable human personality.

Safety effects are mixed. Neuroticism suppression lowers judged frustration and amplification raises it, but the control also lowers it. Sycophantic capitulation moves from 0.33 to 0.65 with high agreeableness and 0.26 with low agreeableness; the control reaches 0.61. On CoCoNot, low agreeableness raises should-decline compliance to 0.33-0.35 versus 0.14: resisting social pressure is not the same as being safer. On WildJailbreak, A+ lowers harmful compliance from 0.55 to 0.25 but raises benign noncompliance from 0.0286 to 0.1667; C+ raises harm to 0.7146 and the control to 0.6508. A half-A+/half-C+ mixture reaches 0.4512, close to activation capping at 0.4525, with benign noncompliance 0.0524. There is no human grading and each task relies on a binary judge.

The unsupervised section generates 2,500 synthetic 15-turn conversations across 25 archetypes and 100 scenarios. Claude Opus 4.6 iteratively constructs 72 binary items, filtered to 64 for Llama and 58 for Qwen. Principal-axis factoring with oblimin rotation yields Tone, Initiative, Didacticism and Epistemic Caution (TIDE), explaining 40.7% of variance. Scenario explains 56-78% of factor scores and archetype at most 6%. After scenario residualization, explained variance falls to 29.6%; Llama Didacticism drops to congruence 0.69 and alpha 0.67, while Qwen Epistemic Caution drops to congruence 0.743 and alpha 0.370. Cross-model congruence is 0.54-0.80, mean 0.66, and Qwen scores Llama-generated conversations rather than its own population. TIDE is an exploratory hypothesis about synthetic styles, not a universal taxonomy.

TIDE induction reuses the discovery questionnaire, while control and off-target shifts can equal or exceed target effects. Several coefficients are selected after inspecting coherence and transcripts on the same population later shown, without a separate selection set or multiplicity correction. The Discussion also contradicts the measured agreeableness direction: it says higher agreeableness increases harmful compliance, while results and figures show it lowers harm and low agreeableness raises CoCoNot compliance.

Public artifacts are extensive but do not yet form a stable end-to-end reproduction. Code 76d3c14 and data 755b97d were audited; Hugging Face reports about 1.75 TB of adapters, outputs and figure sources. Logs, responses, judgments and CSVs support checks such as WildJailbreak. The lock resolves 290 packages, uv lock --check passes, and wheel and sdist builds succeed. Yet pytest ends with 21 failures, 599 passes and 11 skips; Ruff reports 212 findings in the stable layer and 1,258 repository-wide; no CI exists. Safety and unsupervised work remains in src_dev/scripts_dev, GSM8K/TruthfulQA lack clean configs, and the figure registry marks only one figure verified. This ambitious, auditable contribution shows approximate behavioural control and trade-offs. It does not establish pure axes, cost-free control, general safety, inner personality or universal factors.

Español

Este preprint estudia si una persona de un LLM puede representarse y modificarse como una posición en un espacio de rasgos. Para apertura, responsabilidad, extraversión, amabilidad y neuroticismo (OCEAN), entrena un LoRA amplificador y otro supresor, además de un control neutral. La señal procede de constituciones de 12 enunciados por rasgo y dirección; un profesor genera preferencias para DPO, después se entrena otro LoRA por SFT con 12.000 autorreflexiones y conversaciones, y ambos se combinan. Incluye Llama-3.1-8B-Instruct, Qwen3-8B/32B y Gemma-3-4B/12B/27B. Los seis modelos solo comparten TRAIT y MMLU: la evaluación completa, los jueces abiertos, la mayor parte de seguridad y el análisis factorial se concentran en Llama-3.1-8B, salvo frustración en Gemma-3-27B.

El resultado más sólido es operativo y acotado. En rangos moderados, muchos LoRA desplazan principalmente su puntuación objetivo de forma aproximadamente monótona; escalarlos gradúa el efecto y combinarlos produce personas aproximadamente aditivas. Los ejes no son ortogonales ni perfectamente invertibles: hay movimiento fuera del objetivo, el escalado negativo no siempre invierte la dirección y las combinaciones dejan residuos. A escalas fuertes aparecen respuestas erróneas, ausentes o imposibles de parsear y caen MMLU, GSM8K y TruthfulQA. Algunas mezclas de cinco LoRA colapsan MMLU y la suma de coeficientes no predice bien la pérdida. La conservación de capacidades solo se sostiene en rangos moderados y configuraciones concretas.

La validación humana es más estrecha que la expresión «panel validado por humanos». Tres autores puntuaron 33-36 ítems escritos por ellos para amabilidad, neuroticismo y coherencia; apertura, responsabilidad y extraversión no recibieron gold humano directo. TRAIT fue creado para humanos, aquí se responde por log-probabilidades de elección forzada, se usan 300 de 1.000 preguntas por rasgo y se descartan ítems con menos de 0,75 de masa en opciones válidas. El protocolo mide respuestas condicionadas, no una personalidad humana estable.

Los efectos de seguridad son mixtos. Suprimir neuroticismo reduce frustración juzgada y amplificarlo la aumenta, pero el control también la reduce. En sicofancia, la capitulación pasa de 0,33 a 0,65 con alta amabilidad y a 0,26 con baja; el control llega a 0,61. En CoCoNot, baja amabilidad eleva el cumplimiento indebido a 0,33-0,35 frente a 0,14: ser menos capitulador no equivale a ser más seguro. En WildJailbreak, A+ reduce cumplimiento dañino de 0,55 a 0,25, pero eleva no cumplimiento benigno de 0,0286 a 0,1667; C+ lo eleva a 0,7146 y el control a 0,6508. Un punto medio A+/C+ queda en 0,4512, casi igual que activation capping (0,4525), con 0,0524 de no cumplimiento benigno. No hay evaluación humana de estas decisiones y se usa un juez binario por tarea.

La parte no supervisada genera 2.500 conversaciones sintéticas de 15 turnos con 25 arquetipos y 100 escenarios. Claude Opus 4.6 construye iterativamente 72 preguntas binarias, filtradas a 64 para Llama y 58 para Qwen. Un análisis factorial de ejes principales con rotación oblimin devuelve Tone, Initiative, Didacticism y Epistemic Caution (TIDE), que explican 40,7% de varianza. El escenario explica 56-78% de las puntuaciones y el arquetipo como máximo 6%. Al residualizar escenario, la varianza explicada baja a 29,6%; Didacticism en Llama queda en congruencia 0,69 y alfa 0,67, y Epistemic Caution en Qwen en congruencia 0,743 y alfa 0,370. La congruencia entre modelos es 0,54-0,80, media 0,66, y Qwen puntúa conversaciones generadas por Llama, no una población propia. TIDE es una hipótesis exploratoria sobre estilos sintéticos, no una taxonomía universal.

La inducción TIDE reutiliza el cuestionario que descubrió los factores; control y efectos laterales pueden igualar o superar el efecto objetivo. Varios coeficientes se seleccionan después de inspeccionar coherencia y transcripciones sobre la misma población luego mostrada, sin conjunto separado ni corrección por multiplicidad. La discusión contiene una contradicción: afirma que aumentar amabilidad aumenta cumplimiento dañino, cuando resultados y figuras muestran que lo reduce y que la baja amabilidad aumenta CoCoNot.

Los artefactos públicos son amplios, pero no forman todavía una reproducción estable de extremo a extremo. Se auditó código 76d3c14 y datos 755b97d; Hugging Face informa unas 1,75 TB con adaptadores, outputs y fuentes de figuras. Hay logs, respuestas, juicios y CSV que verifican números como WildJailbreak. El lock resuelve 290 paquetes, uv lock --check pasa y se construyen wheel y sdist. Sin embargo, pytest termina con 21 fallos, 599 éxitos y 11 skips; Ruff encuentra 212 problemas en la capa estable y 1.258 en todo el repo; no existe CI. Seguridad y no supervisado siguen en src_dev/scripts_dev, GSM8K/TruthfulQA no tienen configs limpias y el registro de figuras solo marca una como verificada. Es una contribución ambiciosa y auditable que demuestra control conductual aproximado y trade-offs; no demuestra ejes puros, control sin coste, seguridad general, personalidad interna ni factores universales.

Research question

Can LoRA adapters trained by constitutional distillation learn, scale, and combine OCEAN traits in weight space, affect behaviors relevant to safety, and serve to discover unsupervised behavioral factors?

Method

Experimental study of LoRA rank 64 with DPO and SFT on six models from three families. It evaluates OCEAN with TRAIT, LLM judges, and capability benchmarks; tests frustration, sycophancy, CoCoNot, and WildJailbreak; and applies principal axis factor analysis to 2,500 synthetic conversations. Code and data are audited in fixed public revisions.

Sample: Six base models between 4B and 32B for TRAIT/MMLU; complete evaluation primarily on Llama-3.1-8B; frustration on Gemma-3-27B; 2,500 synthetic conversations for factors. No persons participate as subjects: three authors annotate small gold sets.

Findings

  • LoRA adapters generally move the target trait in an approximately monotonic manner over a moderate range.
  • The axes show off-target movement and are not orthogonal.
  • Negative scaling only partially inverts some adapters.
  • Combinations are approximately additive, but leave residuals.
  • Strong scales and some mixtures degrade or collapse capability.
  • Only TRAIT and MMLU are run on the six models.
  • Human validation does not cover all five traits.
  • Suppressing neuroticism reduces frustration, but the control also reduces it.
  • High agreeableness raises sycophancy from 0.33 to 0.65; low agreeableness reduces it to 0.26.
  • The neutral control raises sycophancy to 0.61.
  • Low agreeableness raises CoCoNot compliance to 0.33-0.35 versus 0.14.
  • A+ low reduces WildJailbreak harm to 0.25 but raises benign non-compliance to 0.1667.
  • C+ raises harm to 0.7146 and the control to 0.6508 versus 0.55.
  • TIDE explains 40.7% before controlling for scenario and 29.6% after.
  • Scenario explains 56-78% of the factor scores.
  • Congruence across models is 0.54-0.80, mean 0.66.
  • Not all factors retain reliability after residualizing scenario.
  • Induction reuses the discovery questionnaire and has lateral effects.
  • The artifacts allow auditing numbers, but the public suite has 21 failures.

Limitations

  • Preprint without verified peer review.
  • Complete evaluation is limited primarily to Llama-3.1-8B.
  • TRAIT was designed for humans and only a subsample is used.
  • Human gold covers two traits and coherence, with three authors.
  • There is no direct gold for openness, conscientiousness, or extraversion.
  • The axes show correlation, lateral movement, and incomplete inversion.
  • Capability collapses at strong scales or combinations.
  • The neutral control is not inert.
  • Safety is judged with a single binary judge per task.
  • There is no human validation of the safety outputs.
  • Safety effects include over-refusal.
  • Coefficients are selected on the same metrics shown.
  • There is no independent selection set or multiplicity correction.
  • The discussion contradicts the measured direction of agreeableness.
  • The factor dialogues are synthetic and dominated by scenario.
  • Qwen scores Llama rollouts, not its own population.
  • Residualized reliability is not uniform.
  • Induction validation reuses the discovery instrument.
  • Only Initiative and Tone receive direct induction.
  • Lateral effects and control can exceed the target.
  • The clean layer does not cover all safety or unsupervised analysis.
  • GSM8K and TruthfulQA lack published clean configs.
  • pytest has 21 failures and no CI exists.
  • Ruff is not configured as a gate.
  • OpenRLHF is installed from an unfixed branch.
  • The figure registry marks only one as verified.

What the study does not establish

  • That LLMs have human personality or internal mental life
  • That OCEAN is the natural structure of models
  • That the axes are orthogonal, pure, or perfectly invertible
  • That combining LoRA is exactly linear
  • That the control has no capability cost
  • That all six models receive the full suite
  • That the human panel validates all five traits
  • That higher or lower agreeableness is uniformly safer
  • That the neutral control is inert
  • That safety generalizes to other judges, languages, or attacks
  • That coefficients are confirmed out of sample
  • That TIDE is a universal taxonomy
  • That the factors are independent of scenario
  • That TIDE induction has independent validation
  • That the factors imply intention, affect, or agency
  • That the full public pipeline reproduces the paper
  • That a general guarantee of alignment or safety exists

Traceability

Scope: Full text

Version: arXiv:2607.07916v1; 85-page preprint; PDF, TeX, code commit 76d3c14, Hugging Face artifact commit 755b97d, tests, build, figures and numerical claims audited 2026-07-16

Consulted source: https://arxiv.org/abs/2607.07916v1

Review: Codex 85-page full-text visual, arXiv TeX, immutable GitHub/Hugging Face artifact, numerical claim, test, build and construct-validity audit, 2026-07-16

Approval: Codex fidelity pass, 2026-07-16

English translation: approved, 2026-07-18

Models evaluated

  • Llama-3.1-8B-Instruct
  • Qwen3-8B
  • Qwen3-32B
  • Gemma-3-4B-IT
  • Gemma-3-12B-IT
  • Gemma-3-27B-IT
  • GLM-4.5-Air teacher
  • DeepSeek-V3.2 teacher robustness check
  • Qwen3-235B-A22B default judge
  • GPT-5.4 nano user simulator
  • Claude Opus 4.6 questionnaire generator

Instruments and metrics

  • OCEAN / Big Five
  • Constitution-guided distillation
  • LoRA
  • DPO
  • SFT
  • TRAIT
  • MMLU
  • GSM8K
  • TruthfulQA
  • LLM-judge calibration panel
  • Sycophancy inspect_eval
  • CoCoNot
  • WildJailbreak
  • Multi-turn frustration evaluation
  • Principal axis factoring
  • Oblimin rotation
  • Cronbach alpha
  • Tucker congruence
  • TIDE factors

Data used

  • TRAIT; 1,000 items per trait, 300 sampled per reported sweep
  • MMLU; 100 sampled items per scale point and three runs
  • GSM8K and TruthfulQA on Llama-3.1-8B
  • 240/299 neutral psychometric seed prompts
  • Sycophancy and CoCoNot inspect_evals outputs
  • WildJailbreak: 800 adversarial-harmful plus 210 adversarial-benign prompts
  • 2,500 synthetic 15-turn conversations across 25 archetypes and 100 scenarios
  • Hugging Face artifact revision 755b97d

Evidence and location

  • Method, results, appendices, limitations, and agreeableness contradiction: arXiv:2607.07916v1 PDF, 85 pages; every page rendered and visually inspected
  • Identity, date, version, DOI, and preprint status: Official arXiv record and TeX source for 2607.07916v1
  • Tests, lint, build, clean layer coverage, and absence of CI: GitHub commit 76d3c14a226c8fb0459601e267bbab5563d5c1fc
  • Public data and exact WildJailbreak values: Hugging Face dataset commit 755b97da883159e218a45712f6e36454958e576d
  • Editorial status as preprint: Exact-title arXiv and OpenReview web searches checked 2026-07-16
  • Consolidated audit: reports/verification/article-278-persona-cartography-ocean-lora-factor-validity-safety-selection-code-data-and-claim-audit.json