Mechanistic Personality Analysis of LLMs Steering Personality via Latent Feature Interventions

Trait induction and control2026arXivApproved editorial review

Authors: David Courtis, Ting Hu

Keywords: Big Five, Activation steering, Sparse autoencoder, Contrastive activation analysis, Mechanistic interpretability, DeepSeek-R1-Distill-Llama-8B, Personality evaluation

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

2
Authors
7
Findings
19
Limitations
7
Evidence

Editorial summary

English

This preprint asks whether Big Five language expression can be changed by intervening directly in the hidden activations of DeepSeek-R1-Distill-Llama-8B. The authors use a public 32,768-feature sparse autoencoder trained on LMSYS-Chat-1M for layer 19 and report generating 12,000 PsyBORGS Facebook-style status updates. High- and low-trait prompt sets provide final-token activations. Their means are mapped into SAE space, features are ranked, and decoder directions are added during generation. A grid search tunes positive and negative feature counts and magnitudes under an objective intended to balance trait expression with MMLU performance. Evaluation combines text-embedding-ada-002 similarity to two unreleased references per trait, paired judgments from Claude 3.7 and GPT-4o, and a human evaluation described only as a parallel procedure. The paper shows visible changes in examples and figures. Its LLM chart reports high-confidence correct trial-bin shares of 50% for Openness, 94% for Conscientiousness, 74% for Extraversion, 78% for Agreeableness, and 92% for Neuroticism. The human chart reports 30%, 98%, 65%, 58%, and 95%. These numbers are not ordinary pairwise accuracy. The method groups 20 comparisons into a trial and calls a trial high-confidence correct when more than 80% are correct; the bars are the shares of 200 trials in each bin. The relation among 20 paired samples, 200 trials, and 20 evaluations per trial is unexplained. Human participant count, recruitment, assignments, ratings per item, randomization, blinding, agreement, exclusions, consent, and ethics review are all absent. Positive steering is plotted above neutral cosine similarity for every trait, but negative steering is mixed and for Openness exceeds both neutral and positive conditions. The two reference texts are not released, and embedding similarity may measure topic, valence, verbosity, or discussion of a trait rather than psychological expression. Several mathematical and implementation inconsistencies limit the mechanistic claim. The method computes E(mean h) rather than mean E(h); a nonlinear SAE makes these quantities different, so the displayed contrast is not the mean feature-activation difference described in prose. Features are ranked by absolute Delta z, yet the paper says swapping positive and negative sets finds negative features. Swapping only negates Delta z and leaves the absolute ranking unchanged, so distinct positive and negative feature sets require an unstated sign filter or different rule. The equations add decoder directions to the original hidden state, whereas another section encodes the state, modifies sparse coefficients, reconstructs W-transpose a, and replaces the state, potentially dropping the reconstruction residual. Feature selection also shifts from high-versus-low contrastive means to correlation with scores on 200 neutral prompts. The methods objective subtracts a degradation penalty, but the results equation adds it; the exact component scales are not defined. Figure 5 reports Extraversion magnitude/count values of 10/13 and 4/5, while the prose reports counts 9/4 and magnitudes 13/5. The central experimental threat is apparent reuse of the same 200 prompts for feature correlation and selection, grid calibration, and evaluation, with no independent discovery, validation, and locked test sets. There are no matched comparisons against prompt steering, direct contrastive activation addition, random directions, SAE reconstruction alone, or fine-tuning. MMLU is named as the performance term, but no baseline or steered accuracy, task configuration, breakdown, or uncertainty is published; the claim of preserved benchmark performance is therefore unsupported by displayed evidence. There are no confidence intervals, hypothesis tests, repeated-seed distributions, or cross-trait specificity tests. A magnitude-30 example visibly loses coherence, but four selected excerpts are not a statistical coherence study. The defensible contribution is exploratory: it integrates an existing SAE with activation steering and shows that internal perturbations can change language patterns associated with OCEAN. It does not establish monosemantic personality features, a stable psychological representation, causality over a mental trait, or superiority to simpler controls. The SAE checkpoint is public, but the claimed codebase is not linked and the 12,000 texts, prompts, feature indices, reference texts, outputs, scores, MMLU results, human ratings, seeds, and environment are not released. Despite using the accepted ICML 2025 style, no acceptance or publication evidence was located; the source should be cited as arXiv v1.

Español

Este preprint estudia si se puede modificar la expresión lingüística de los cinco rasgos OCEAN interviniendo directamente en activaciones internas de DeepSeek-R1-Distill-Llama-8B. Los autores usan un sparse autoencoder público de 32.768 características, entrenado sobre LMSYS-Chat-1M para la capa 19, y una colección declarada de 12.000 actualizaciones de estado tipo Facebook generadas con PsyBORGS. Construyen conjuntos de prompts que describen niveles altos y bajos de cada rasgo, extraen la activación del último token, comparan sus medias en el espacio latente del SAE y buscan características que asocian con mayor o menor expresión. Durante la generación añaden direcciones del decodificador a la activación de la capa 19. Una búsqueda en rejilla ajusta magnitud y número de características positivas y negativas mediante un objetivo que pretende equilibrar expresión del rasgo y rendimiento MMLU. La evaluación usa similitud coseno de text-embedding-ada-002 con dos textos de referencia no publicados por rasgo, juicios pareados de Claude 3.7 y GPT-4o, y una evaluación humana descrita solo como paralela. El artículo presenta diferencias visibles en ejemplos y en sus gráficos. En la figura de jueces LLM, la proporción de bloques clasificados como correctos con alta confianza es 50% para Apertura, 94% para Responsabilidad, 74% para Extraversión, 78% para Amabilidad y 92% para Neuroticismo. En humanos se muestran 30%, 98%, 65%, 58% y 95%, respectivamente. Es crucial no llamar a estos porcentajes exactitud binaria ordinaria: el método dice agrupar 20 comparaciones por ensayo y etiquetar como alta confianza un ensayo con más del 80% de aciertos; las barras representan la proporción de 200 ensayos en cada categoría. La relación entre 20 pares, 200 ensayos y 20 evaluaciones por ensayo no queda explicada. Tampoco se informa cuántas personas evaluaron, cómo fueron reclutadas, cuántas valoraciones recibió cada texto, si hubo aleatorización o cegamiento, acuerdo entre jueces, exclusiones, consentimiento o revisión ética. La similitud coseno aumenta bajo la condición positiva para los cinco rasgos, pero la condición negativa es mixta y para Apertura supera tanto a la neutral como a la positiva, lo que cuestiona una lectura monotónica. Los dos textos de referencia por rasgo no se publican y la métrica puede capturar tema, valencia, longitud o discusión explícita del rasgo en vez de su expresión psicológica. El propio artículo reconoce parte de esta limitación. Las conclusiones mecanicistas requieren todavía más cautela. El método aplica el encoder no lineal a la activación media, E(media h), en vez de promediar los códigos, media E(h); esas cantidades no son equivalentes y la primera no es la diferencia media de activación de características que el texto afirma ordenar. Además, se ordena por |Delta z| y se dice que para encontrar características negativas se invierten los conjuntos. Invertirlos solo cambia el signo de Delta z y deja idénticos todos los valores absolutos, de modo que, sin un filtro de signo no descrito, no puede producir rankings positivos y negativos distintos. También hay dos implementaciones incompatibles: las ecuaciones añaden vectores del decodificador al estado original, mientras otra sección codifica el estado, cambia coeficientes, reconstruye W^T a y reemplaza la activación, lo que puede eliminar el residuo de reconstrucción. El criterio de selección cambia igualmente entre el contraste alto-bajo y la correlación con puntuaciones en 200 prompts neutrales. La función objetivo resta la penalización de degradación en Métodos, pero la suma en Resultados; si C es una degradación positiva, la segunda fórmula premia peor rendimiento. No se define la escala exacta de los dos términos. La figura de hiperparámetros también contradice el texto: para Extraversión muestra magnitud positiva 10 y 13 características, y magnitud negativa 4 con 5 características; el párrafo dice 9 y 4 características con magnitudes 13 y 5. El problema experimental más importante es la reutilización aparente de los mismos 200 prompts para correlacionar y seleccionar características, calibrar la rejilla y evaluar. No se documentan conjuntos independientes de descubrimiento, validación y test bloqueado. Tampoco hay comparaciones con prompting, una dirección contrastiva sin SAE, direcciones aleatorias de igual norma, reconstrucción SAE sin intervención, fine-tuning u otros controles que permitan atribuir la mejora al SAE. MMLU se menciona como penalización, pero no se publica ninguna exactitud basal o intervenida, configuración, desglose o incertidumbre; por ello no queda respaldada la afirmación de conservar alto rendimiento en benchmarks. No hay intervalos de confianza, pruebas, distribución entre semillas ni evaluación de desplazamientos cruzados entre rasgos. Los autores sí muestran un límite práctico: con magnitud positiva 30 aparece una salida severamente incoherente, aunque cuatro ejemplos no constituyen una medida estadística de coherencia. La aportación defendible es exploratoria: integra un SAE existente con steering de activaciones y ofrece evidencia visual de que perturbaciones internas cambian patrones lingüísticos asociados a OCEAN. No demuestra características monosemánticas de personalidad, una representación psicológica estable, causalidad sobre un rasgo mental ni superioridad frente a controles sencillos. El SAE está disponible públicamente, pero el manuscrito no enlaza el supuesto codebase y no libera los 12.000 textos, prompts, índices de características, referencias, salidas, puntuaciones, MMLU, evaluaciones humanas, semillas o entorno. Aunque el TeX usa estilo ICML 2025 aceptado, no se encontró evidencia de aceptación o publicación y debe citarse como arXiv v1.

Research question

Can an additive intervention on latent features obtained with a sparse autoencoder change in a controllable manner the language associated with each OCEAN trait in DeepSeek-R1-Distill-Llama-8B without degrading its general performance?

Method

Exploratory experimental study with a single model and a single layer. It generates 12,000 synthetic states with OCEAN prompting, contrasts final activations of high and low prompts, uses a pretrained SAE to propose features, intervenes on activations at inference, and adjusts magnitudes and quantities through grid search. It evaluates with OpenAI embeddings, Claude 3.7 and GPT-4o judges, an unspecified human evaluation, and an MMLU penalty whose numerical results are not published.

Sample: The article declares 12,000 synthetic state updates with OCEAN intensities 1-3 and 7-9, omitting middle levels, but gives no distribution by trait or level and does not publish the data. For evaluation it declares 200 neutral prompts. For judges it speaks at once of 20 pairs and 200 trials of 20 evaluations per trait; it does not clarify reuse or unit of analysis. The human sample and the number of ratings are unknown.

Findings

  • The perturbations produce visible linguistic differences between tuned up and tuned down examples across the five traits.
  • The LLM bars show proportions of high-confidence trials of 50%, 94%, 74%, 78%, and 92% for O, C, E, A, and N; these are not ordinary per-pair accuracy.
  • The human bars show 30%, 98%, 65%, 58%, and 95%, but lack denominators and an auditable protocol.
  • The positive condition shows higher cosine similarity than the neutral condition for all traits; the negative condition is mixed, and for Openness it exceeds both neutral and positive.
  • The objective figure marks maxima of 0.0343, 0.0467, 0.0321, 0.0375, and 0.0494 for O, C, E, A, and N, but these are optimized values over a metric and a set that are not clearly separated.
  • A Neuroticism output at magnitude 30 loses coherence visibly; the article does not quantify that degradation.
  • Conscientiousness and Neuroticism are the traits with the highest proportion of high-confidence blocks in LLM and human judges within the published protocol.

Limitations

  • A single model, a single layer, a single SAE, and a single state-update task limit generalization.
  • E(mean h) does not equal mean E(h) for a nonlinear encoder; the measure is not the mean feature activation difference.
  • Inverting sets and sorting by absolute value produces the same ranking, so the positive-negative separation is incomplete or poorly specified.
  • The equations for direct addition and reconstruction from sparse code describe distinct and potentially confusing interventions.
  • The manuscript does not resolve whether it selects features by high-low contrast or by correlation with the 200 neutral outputs.
  • The objective function changes sign for the penalty between Methods and Results and does not define comparable scales.
  • Text and figure contradict the optimal hyperparameters for Extraversion.
  • The same 200 prompts appear to be used in selection, calibration, and evaluation; no blocked final test is documented.
  • There are no baselines for prompting, direct CAA, random directions, SAE reconstruction, fine-tuning, or equal-norm controls.
  • Two unpublished references and embeddings may conflate topic, valence, style, or length with personality expression.
  • The classification percentages summarize trials binned by thresholds; per-pair accuracy and the dependency structure are not published.
  • The human evaluation omits sample size, recruitment, consent, allocation, blinding, agreement, and raw data.
  • The LLM judges lack exact snapshots, configuration, dates, order balancing, and agreement analysis.
  • MMLU results are not published despite claiming preservation of general performance.
  • There are no intervals, statistical tests, multiplicity, repeated seeds, or uncertainty.
  • Specificity between traits or simultaneous changes in the other four traits are not tested.
  • The selected features are not individually labeled or interpreted; monosemanticity and psychological mechanism are not validated.
  • Toxicity, bias, truthfulness, factuality, refusals, deception, persuasion, or safety are not evaluated.
  • The mentioned codebase is not linked and all study-specific artifacts are missing.

What the study does not establish

  • It does not establish that LLMs possess human personality or stable psychological traits.
  • It does not identify monosemantic features or causal mechanisms of personality.
  • It does not demonstrate that the SAE adds value over contrastive directions or simpler controls.
  • It does not demonstrate generalization to other models, layers, tasks, languages, or conversational contexts.
  • It does not support the claim of preserving MMLU or general performance with published numerical results.
  • It does not demonstrate psychometric validity of embeddings, LLM judges, or human evaluation.
  • It does not allow interpreting the high-confidence percentages as ordinary individual accuracy.
  • It does not rule out overfitting from selection and tuning on the same prompts and metrics.
  • It does not demonstrate independent control of one trait without altering the others.
  • It does not allow end-to-end reproduction or independent reanalysis.
  • It does not provide sufficient evidence for persuasion deployments, mental health, psychological assessment, or high-risk agents.

Traceability

Scope: Full text

Version: arXiv:2606.28770v1

Consulted source: https://arxiv.org/pdf/2606.28770

Review: Codex 16-page full-text visual, TeX, figure, mathematical, implementation, evaluation, human-subject, artifact, reproducibility, safety and claim audit, 2026-07-16

Approval: Codex fidelity pass, 2026-07-16

English translation: approved, 2026-07-18

Models evaluated

  • DeepSeek-R1-Distill-Llama-8B
  • qresearch/DeepSeek-R1-Distill-Llama-8B-SAE-l19
  • text-embedding-ada-002
  • Claude 3.7 as LLM judge
  • GPT-4o as LLM judge

Instruments and metrics

  • PsyBORGS structured personality prompting
  • Final-token layer-19 activation extraction
  • Sparse-autoencoder latent feature ranking
  • Additive activation steering
  • 5x5x5x5 grid search with coordinate refinement
  • Cosine similarity to two reference texts per trait
  • Twenty-pair high, middle and incorrect confidence binning
  • Human paired comparison with unreported protocol
  • MMLU performance penalty with unreported numerical results

Data used

  • LMSYS-Chat-1M, used previously to train the public SAE
  • 12,000 reported PsyBORGS-generated Facebook-style status updates, not released
  • Two hundred reported neutral everyday and knowledge prompts, not released
  • Two trait-reference text files per OCEAN trait, not released
  • MMLU, configuration and results not reported

Evidence and location

  • Metadata, date comment, authors, and category: Official arXiv record 2606.28770v1, checked 2026-07-16
  • Synthetic corpus, activation contrast, and SAE: arXiv v1, Sections 3.1-3.2 and equations 1-8
  • Embeddings, judges, human evaluation, and MMLU: arXiv v1, Section 3.2.1-3.2.2
  • Steering, grid search, reconstruction, and selection by correlation: arXiv v1, Sections 3.3-3.5 and equations 9-14
  • Percentages, similarity, hyperparameters, and objective function: arXiv v1, Section 4, Figures 2-6 and equations 15-16
  • Qualitative examples, coherence loss, and declared limitations: arXiv v1, Tables 2-4 and Sections 4.3-6
  • Consolidated audit of mathematics, implementation, selection, evaluation, artifacts, safety, and claim limits: reports/verification/article-291-arxiv-icml-template-feature-sign-inversion-encoded-mean-reconstruction-selection-leakage-human-evaluation-mmlu-artifact-and-claim-audit.json