Identifying and Manipulating Personality Traits in LLMs Through Activation Engineering

Trait induction and control2024arXivApproved editorial review

Authors: Rumi Allbert, James K. Wiles, Vlad Grankovsky

Keywords: Computation and Language, Artificial Intelligence, Large Language Models, Activation Engineering, Personality Traits

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

3
Authors
13
Findings
40
Limitations
17
Evidence

Editorial summary

English

The paper proposes controlling styles described as personality traits through an intervention in the internal activations of an 8-billion-parameter Llama 3 model. The authors assemble 179 labels said to come from HEXACO, the Big Five, lexical analysis, synonyms, and antonyms. For each label they apply a system prompt prescribing the target behavior to 1,500 Alpaca instructions and compare it with a neutral condition. They extract Layer 18 activations and define a trait direction as the difference between the two condition means. The exact checkpoint, also described as uncensored, is not identified, and token position, pooling, generation parameters, direction normalization, and train/evaluation partitioning are not specified.

The intervention neither fine-tunes the model nor permanently changes its weights. In the stated equation, the current projection of an activation onto the trait direction is removed and replaced with the mean projection observed in trait-conditioned examples, multiplied by a factor α. The authors identify 1.3–1.4 as a useful α range through empirical observation: lower values allegedly yield small changes and higher values incoherent text. They publish no sweep, personality metric, effect sizes, or aggregate results supporting that range. The behavioral demonstration is mainly a table with five responses, baseline, shy, passionate, narcissistic, and paranoid, to one market-strategy question. There is no blinded evaluation, human rating, psychometric questionnaire, held-out prompt set, prompt-only baseline, or general-capability measurement.

The remainder explores the geometry of the 179 directions. It displays PCA, t-SNE, and UMAP projections for a random sample of 100 labels, runs k-means with 20 groups on the original vectors, studies principal-component reconstruction, and visually compares a greedy basis selection with random vector sets. Seeds, reduction hyperparameters, the rationale and stability of k, repeated runs, confidence intervals, and complete numerical results are not reported. The clusters mix ordinary traits with clinical diagnoses, paraphilias, crimes, affective states, and moral judgments. One group contains, among other labels, pedophilic, murderous, and psychopathic; another places autistic alongside organized, perfectionist, and rigid thinker. Proximity in reduced spaces is interpreted as a possible pathway or precursor to harmful behavior, but the geometry of prompt-induced activations cannot establish psychological or clinical causality.

The paper shows screenshots of two interfaces: base-personality selection and modular design using principal-component sliders. The authors withhold them for safety reasons. They also link no code, data, complete prompt or trait list, checkpoint, demo, or reproducible results. Proposed applications in games, artificial companions, customer service, and education are not evaluated. The ethics section recognizes risks of manipulation, apparent empathy, concealed limitations, and toxic generation, but its concrete mitigation is limited to urging responsible use; no safeguard is tested.

The defensible contribution is an exploratory demonstration that a mean-difference direction at Layer 18 can visibly change the style of some responses from an incompletely identified model. It does not show that the direction represents a stable personality construct, that the effect generalizes, that model capabilities are preserved, or that the clusters reflect internal psychological structure. The statement that Llama 3 8B has 31 layers conflicts with Meta's official configuration, which declares 32 transformer blocks; the authors may be referring to zero-based indices 0–31, but the paper says 31 layers in total. With no artifacts or quantitative evaluation, the method's efficacy, robustness, reproducibility, and safety cannot be estimated.

Español

El artículo propone controlar estilos descritos como rasgos de personalidad mediante una intervención en las activaciones internas de un Llama 3 de 8 mil millones de parámetros. Los autores reúnen 179 etiquetas procedentes, según indican, de HEXACO, los Cinco Grandes, análisis léxico, sinónimos y antónimos. Para cada etiqueta aplican a 1.500 instrucciones de Alpaca un prompt de sistema que prescribe el comportamiento deseado y lo comparan con una condición neutral. Extraen las activaciones de la capa 18 y definen la dirección del rasgo como la diferencia entre las medias de ambas condiciones. El checkpoint exacto del modelo, descrito también como «sin censura», no se identifica, y tampoco se especifican posición de token, pooling, parámetros de generación, normalización de la dirección ni partición de entrenamiento y evaluación.

La intervención no realiza ajuste fino ni modifica de forma permanente los pesos. En la ecuación presentada se elimina de una activación su proyección actual sobre la dirección del rasgo y se sustituye por la proyección media observada en los ejemplos del rasgo, multiplicada por un factor α. Los autores sitúan el intervalo útil de α entre 1,3 y 1,4 por observación empírica: valores menores producirían cambios leves y valores mayores texto incoherente. No publican el barrido, una métrica de personalidad, tamaños de efecto ni resultados agregados que sustenten ese intervalo. La demostración conductual consiste principalmente en una tabla con cinco respuestas, base, tímida, apasionada, narcisista y paranoide, a una única pregunta de estrategia de mercado. No hay evaluación ciega, jueces humanos, cuestionarios psicométricos, prompts retenidos, comparación con prompting simple ni medición de capacidades generales.

El resto del trabajo explora la geometría de las 179 direcciones. Muestra PCA, t-SNE y UMAP de una muestra aleatoria de 100 etiquetas, aplica k-means con 20 grupos sobre los vectores originales, estudia reconstrucción mediante componentes principales y compara visualmente una selección codiciosa de vectores base con conjuntos aleatorios. No informa semillas, hiperparámetros de reducción, justificación o estabilidad de k, repeticiones, intervalos de confianza ni valores numéricos completos. Las agrupaciones mezclan rasgos ordinarios con diagnósticos clínicos, parafilias, delitos, estados afectivos y juicios morales. Un grupo reúne, entre otros, «pedophilic», «murderous» y «psychopathic»; otro incluye «autistic» junto a «organized», «perfectionist» y «rigid thinker». Proximidades en espacios reducidos se interpretan como posibles vías o precursores de conducta dañina, pero la geometría de activaciones inducidas por prompts no permite inferir causalidad psicológica o clínica.

El artículo presenta capturas de dos interfaces: selección de personalidades base y diseño modular con deslizadores de componentes principales. Los autores deciden no publicarlas por motivos de seguridad. Tampoco enlazan código, datos, lista completa de prompts o rasgos, checkpoint, demo ni resultados reproducibles. Los usos en juegos, compañeros artificiales, atención al cliente y educación son propuestas, no evaluaciones. La sección ética reconoce riesgos de manipulación, empatía aparente, ocultación de limitaciones y generación tóxica, pero la mitigación concreta se limita a pedir un uso responsable; no se prueba ningún control de seguridad.

La aportación defendible es una prueba exploratoria de que una dirección de diferencia de medias en la capa 18 puede alterar visiblemente el estilo de algunas respuestas de un modelo no identificado con precisión. No demuestra que esa dirección represente un constructo estable de personalidad, que el cambio generalice, que conserve las capacidades del modelo o que las agrupaciones reflejen estructura psicológica interna. La afirmación de que Llama 3 8B tiene 31 capas contradice la configuración oficial de Meta, que declara 32 bloques; podría deberse a indexación de 0 a 31, pero el texto dice «31 layers in total». La ausencia de artefactos y de evaluación cuantitativa impide reproducir o estimar la eficacia, robustez y seguridad del método.

Research question

Can an activation direction associated with a personality label be identified by comparing conditioned and neutral prompts, can that direction be used to dynamically modify the responses of an LLM, and can the joint geometric structure of many such directions be analyzed?

Method

Exploratory study of activation engineering. For 179 labels, the authors apply trait-conditioned system prompts and a neutral condition to 1,500 Alpaca instructions, extract activations from layer 18 of a model described as Llama 3 8B uncensored, and compute a difference of means. An equation replaces the current projection of the activation onto each direction with a mean projection scaled with α. Behavioral validation is qualitative and is complemented with PCA, t-SNE, UMAP, k-means with k=20, reconstruction via principal components, and screenshots of an unpublished interface. The editorial audit read and rendered the 24 pages, verified metadata and license, cross-checked the official Llama 3 configuration, and searched for artifacts that the article does not link.

Sample: 1,500 Alpaca prompts are processed for each of 179 labels and their neutral condition, although the effective number of pairs, failures, or exclusions is not reported. The initial visualizations use 100 randomly chosen directions. The visible behavioral evaluation contains a single instruction and five responses. There are no human participants, separate test set, or documented replications.

Findings

  • The authors construct one direction per label by subtracting the neutral mean activation from the mean activation under the conditioned prompt.
  • Layer 18 is selected because, according to an unquantified observation, it produced the largest effect.
  • The intervention replaces the current projection onto the direction with a mean projection of the trait scaled by α.
  • The authors place the useful empirical range of α between 1.3 and 1.4, without publishing the sweep or the metric used.
  • A qualitative table shows visible tone changes among the base, shy, passionate, narcissistic, and paranoid conditions.
  • PCA, t-SNE, and UMAP produce visual maps in which some semantically close labels appear nearby.
  • K-means divides the 179 original vectors into 20 clusters, without validation of the number of groups.
  • PCA reconstruction reduces error as components increase, as is expected, but no quantitative threshold or optimum is fixed.
  • A greedy selection of basis vectors presents lower visual error than random selections, without inferential statistics.
  • The neighbors of the cluster called socially undesirable change according to PCA, t-SNE, and UMAP.
  • The screenshots show that the authors implemented prototypes to choose base labels and combine principal components.
  • The authors do not publish the interface for safety reasons and acknowledge risks of manipulation and harmful generation.
  • The evidence supports exploratory stylistic changes in selected examples, not psychometrically validated personality modification.

Limitations

  • The article does not identify the exact checkpoint of Llama 3 8B or clarify whether it uses a base, instruct, or modified variant.
  • The description "uncensored" does not allow knowing which safeguards were removed or under what license or procedure.
  • The text states that the model has 31 layers, while the official Meta configuration defines 32; it does not clarify whether it counts indices from 0 to 31.
  • No code, configuration, derived dataset, complete list of traits, full prompts, checkpoint, demo, or structured results are published.
  • Hardware, software, dependencies, numerical precision, cost, runtime, or seed are not specified.
  • It is not clear which token position or pooling form produces each activation vector.
  • The exact context, complete templates, generated responses, and decoding parameters are not detailed.
  • The direction is treated mathematically as if it were normalized, but this normalization is not documented.
  • It is not clarified whether the difference of means is computed with prompt pairs or as two independent samples.
  • The selection of layer 18 relies on an empirical observation without a sweep, table, metric, or per-layer comparison.
  • The range α=1.3–1.4 lacks aggregated results, a success criterion, and systematic examples for other values.
  • There is no separation between prompts used to construct the directions and prompts held out to evaluate them.
  • The system prompts explicitly state linguistic markers of the trait, so the direction may capture words, style, and instruction-following.
  • No comparison is made against persona prompting without intervention, random activation, permuted labels, CAA, fine-tuning, or equivalent controls.
  • The visible behavioral validation uses a single question and chosen responses for five conditions.
  • There are no blind judges, inter-annotator agreement, human evaluation, questionnaires, trait classifiers, or automatic metrics.
  • No effect sizes, variance, confidence intervals, statistical tests, or failure rates are reported.
  • The claim of preserving general comprehension is not evaluated with capability or safety benchmarks.
  • Stability across runs, prompts, domains, languages, cultures, checkpoints, model families, or layers is not studied.
  • The provenance of the 179 labels is not provided in an auditable manner, nor is coverage of HEXACO or Big Five quantified.
  • The list mixes personality traits with diagnoses, symptoms, paraphilias, crimes, preferences, styles, and moral judgments.
  • No clinical specialists, psychometricians, ethics experts, or members of affected groups participate in the labeling.
  • No construct, convergent, discriminant, predictive, or incremental validity is provided for the directions.
  • With 179 points in 4,096 dimensions, the two-dimensional visualizations are exploratory and potentially unstable.
  • No seeds or hyperparameters for PCA, t-SNE, UMAP, and k-means are published.
  • The figure uses only 100 of the 179 directions without explaining the sampling or showing replications.
  • k=20 is not justified via silhouette, stability, theory, or comparison with other values.
  • It is not clarified whether k-means operates on normalized vectors or what distance metric is used.
  • Distances and neighborhoods of PCA, t-SNE, and UMAP do not have the same meaning and do not allow a direct causal comparison.
  • Semantic proximity may arise from the text of the instructions themselves, not from an internal psychological structure of the model.
  • Interpreting geometric neighbors as contributors or precursors of harmful behavior exceeds the observational evidence.
  • Grouping "autistic" with rigidity, obsession, and perfectionism reproduces stereotypes and confounds diagnosis with personality.
  • The socially undesirable group combines clinical categories, crimes, and moral judgments without a valid definition of the construct.
  • The comparison of greedy versus random reconstruction does not publish values, repetitions, seeds, or uncertainty.
  • The interface is not available and has no study of usability, accessibility, latency, safety, or adversarial behavior.
  • The possible uses in games, companionship, customer service, and education are not tested with users or task outcomes.
  • Ethical mitigation is largely limited to responsibility recommendations; no barriers, audits, or controls are implemented.
  • There is no formal limitations section and the conclusion uses stronger language than the evidence presented.
  • The bibliography is brief and some methodological foundations rely on secondary or community sources.
  • The lack of artifacts prevents reproducing the figures, verifying the clusters, or checking the intervention formula.

What the study does not establish

  • It does not demonstrate that LLMs possess a human personality or a stable psychological trait.
  • It does not validate the 179 labels via HEXACO, Big Five, or another psychometric instrument.
  • It does not demonstrate that the found direction is specific to the trait and not to the prompt text or linguistic style.
  • It does not prove that the effect generalizes beyond the shown examples.
  • It does not demonstrate stability across models, checkpoints, layers, domains, languages, or runs.
  • It does not establish that α between 1.3 and 1.4 is optimal, safe, or universal.
  • It does not demonstrate that the intervention preserves comprehension, reasoning, factuality, or safety.
  • It does not prove that the clusters correspond to a valid psychological taxonomy.
  • It does not allow inferring that a nearby label causes or anticipates socially harmful behavior.
  • It does not demonstrate that PCA, t-SNE, and UMAP yield consistent or equivalent results.
  • It does not prove the efficacy of the interface or of the proposed uses.
  • It does not provide a comparison of efficacy against prompting, fine-tuning, or other steering methods.
  • It does not provide verified safeguards against manipulation, toxicity, or improper clinical use.
  • It does not offer a reproducible artifact with which third parties can confirm the results.

Traceability

Scope: Full text

Version: arXiv:2412.10427v2 (10 Jan 2025); first submitted 10 Dec 2024; CC BY 4.0

Consulted source: https://arxiv.org/pdf/2412.10427v2

Review: Codex full-text, visual, methodological, ethical and artifact audit, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • Llama 3 8B, described as uncensored; exact checkpoint and base-versus-instruct variant not reported
  • Layer 18 residual activations with a reported width of 4,096
  • Official Meta Llama 3 configuration used only for editorial verification of the 32-layer architecture

Instruments and metrics

  • Trait-conditioned and neutral system prompts
  • Mean-difference activation directions
  • Projection-replacement intervention controlled by alpha
  • PCA, t-SNE and UMAP visualizations
  • K-means clustering with k=20
  • Principal-component reconstruction error
  • Greedy versus random basis-vector selection
  • Unreleased interactive personality-control interfaces

Data used

  • 1,500 instructions from the Alpaca Dataset
  • 179 personality-related labels assembled by the authors; complete machine-readable list and provenance not released
  • Qualitative response examples for one market-strategy prompt

Evidence and location

  • Objective, motivation, and declared contribution: arXiv v2, abstract and section 1, pp. 1–2
  • Difference of means and intervention equations: arXiv v2, sections 2.2.1–2.2.2, equations 3–5, pp. 3–4
  • Alpha range and preparation of 179 labels: arXiv v2, sections 2.2.2–2.3, pp. 4–5
  • Trait prompts and qualitative responses: arXiv v2, sections 2.3.1–2.3.2 and Table 2, pp. 5–7
  • Model, 1,500 prompts, layer 18, and claim of 31 layers: arXiv v2, sections 2.3.2–2.4.1 and Figure 1, pp. 7–8
  • PCA, UMAP, t-SNE, and random sample of 100 vectors: arXiv v2, sections 2.4.1–2.4.2 and Figure 2, pp. 8–9
  • K-means with 20 groups and complete composition: arXiv v2, section 2.4.3 and Table 3, pp. 8–11
  • PCA reconstruction and basis vector selection: arXiv v2, sections 2.5.1–2.5.2 and Figures 4–5, pp. 11–13
  • Traits associated with principal components: arXiv v2, section 2.5.3 and Figures 6–7, pp. 14–15
  • Socially undesirable group and geometric neighbors: arXiv v2, section 2.6 and Figures 8–9 and Table 4, pp. 16–17
  • Ethical interpretation of precursors and proposed mitigation: arXiv v2, section 2.6.3, p. 18
  • Interfaces and decision not to publish them: arXiv v2, section 3 and Figures 10–12, pp. 18–21
  • Use cases and ethical risks: arXiv v2, sections 4–5, pp. 19–21
  • Future work and conclusion: arXiv v2, sections 6–7, pp. 21–23
  • Official Llama 3 architecture with 32 blocks: Meta Llama 3 official repository, llama/model.py, ModelArgs n_layers=32; checked 15 Jul 2026
  • Version, dates, and license: arXiv:2412.10427v2 metadata; revised 10 Jan 2025; CC BY 4.0
  • Absence of linked artifacts: arXiv v2 full paper and abstract metadata; targeted repository search audited 15 Jul 2026