Synthetic Personality Review

Exploring the Personality Traits of LLMs through Latent Features Steering

Trait induction and control2024arXivApproved editorial review

Authors: Shu Yang, Shenzhe Zhu, Liang Liu, Lijie Hu, Mengdi Li, Di Wang

Keywords: Computation and Language, Artificial Intelligence

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

6
Authors
7
Findings
21
Limitations
10
Evidence

Editorial summary

English

The paper treats the behavior of two instruction-tuned Gemma models as reflecting personality factors inspired by social determinism and tests training-free interventions. For Gemma-2B-Instruct and Gemma-2-9B-Instruct, nine “background” dimensions, gender, age, education, socioeconomic status, ideology, emotional intelligence, professional commitment, family relations, and AI familiarity, are represented through Gemma Scope sparse-autoencoder features. GPT-4o generates category descriptors; SAE features activated by one descriptor set but not its contrasts are selected, and their decoder vectors are added to the residual stream. For seven short-term “pressures”, achievement striving, activity, assertiveness, competence, deliberation, gregariousness, and trust, GPT-4o creates positive/negative instruction pairs, activation differences are computed over TRAIT questions, and PCA yields steering directions. Evaluation uses 8,000 Big Five and 3,000 Dark Triad TRAIT questions plus 11,435 SafetyBench questions. With coefficients selected by inspecting benchmark logits, 200/800 for SAE steering and 1.6/1.8 for representation directions in the 2B/9B models, the paper reports background-induced trait changes of 0–7.1 points for 9B versus 0–52.5 for 2B, and pressure-induced changes of 0.1–27.7 versus 0.4–53.5. It interprets the smaller 9B variation as greater “stability,” but provides no repeated runs, intervals, or statistical tests for that comparison. Safety scores are said to fall by 0–6.8 points, especially for offensive content, although Figure 4 itself contains a −7.7 entry and the discrepancy is not explained. The study demonstrates that activation interventions alter multiple-choice answers; it does not show that human social factors are encoded as such or that the models possess persistent personality. Factor names come from GPT-4o-generated descriptions and explanations, and inactivity on a small contrast set does not establish monosemanticity or semantic causality. The prompts used to derive directions also overlap conceptually with the questionnaires used to score them, so steering may directly bias answer selection without producing a general trait. There are no random or norm-matched direction controls, perplexity or output-quality checks, human validation, open-ended tasks, mediation analysis, or evaluation beyond Gemma. Explanations invoking perfectionism, rigidity, overconfidence, or ideological permissiveness are speculative. SafetyBench is not a jailbreak evaluation, and no hallucination results are reported despite the conclusion mentioning them. The official repository, audited at commit 89b6500e6ab7da005ae576365e5fb2bfb5d39c51, contains syntactically valid code and links an external dataset, but not a reproducible environment: dependencies are unpinned and incomplete, README arguments do not match the executables, central feature/result files are absent, tests, CI, and a license are missing, and no public routine calculates the Figure 4 safety scores.

Español

El trabajo trata el comportamiento de dos modelos Gemma instruidos como si reflejara factores de personalidad inspirados por el determinismo social y prueba a modificarlo sin reentrenamiento. En Gemma-2B-Instruct y Gemma-2-9B-Instruct, los autores representan nueve dimensiones de “trasfondo”, género, edad, educación, nivel socioeconómico, ideología, inteligencia emocional, compromiso profesional, relaciones familiares y familiaridad con IA, mediante características de autoencoders dispersos Gemma Scope. GPT-4o genera descriptores de cada categoría; se eligen características SAE que se activan con unos descriptores y no con sus opuestos, y sus vectores de decodificación se añaden al flujo residual. Para siete “presiones” de corto plazo, afán de logro, actividad, asertividad, competencia, deliberación, gregarismo y confianza, GPT-4o crea pares de instrucciones positivas y negativas, se calculan diferencias de activación sobre preguntas TRAIT y PCA produce direcciones de steering. La evaluación usa 8.000 preguntas Big Five y 3.000 de Dark Triad de TRAIT, además de 11.435 preguntas de SafetyBench. Con los coeficientes escogidos por inspección de logits, 200/800 para SAE y 1,6/1,8 para direcciones de representación en 2B/9B, el artículo informa que los cambios de rasgo por trasfondo abarcan 0–7,1 puntos en 9B frente a 0–52,5 en 2B, y por presión 0,1–27,7 frente a 0,4–53,5. Interpreta esa menor variación del 9B como mayor “estabilidad”, pero no aporta repeticiones, intervalos ni tests que sustenten la comparación. En seguridad afirma descensos de 0 a 6,8 puntos, concentrados en contenido ofensivo, aunque la propia matriz de la Figura 4 contiene un valor de −7,7 y no se explica la discrepancia. El estudio demuestra que intervenciones sobre activaciones cambian respuestas de opción múltiple; no demuestra que haya factores sociales humanos codificados como tales ni una personalidad persistente. Los nombres de los factores proceden de descripciones y explicaciones generadas por GPT-4o, y que una característica no se active ante unos pocos contrastes no prueba monosemanticidad o causalidad semántica. Existe además solapamiento conceptual entre los prompts que construyen las direcciones y los cuestionarios que luego las puntúan, por lo que una intervención puede sesgar directamente la respuesta sin crear un rasgo general. No hay controles con direcciones aleatorias o de norma equivalente, medidas de perplejidad/calidad, validación humana, tareas abiertas, análisis de mediación ni evaluación fuera de Gemma. Las explicaciones sobre perfeccionismo, rigidez, sobreconfianza o permisividad ideológica son especulativas. SafetyBench no es una prueba de jailbreak y el artículo no publica resultados de alucinación pese a mencionarlos en la conclusión. El repositorio oficial, auditado en el commit 89b6500e6ab7da005ae576365e5fb2bfb5d39c51, aporta código sintácticamente válido y un dataset externo, pero no un entorno reproducible: dependencias sin versiones e incompletas, argumentos de README discordantes, features/resultados principales ausentes, sin tests, CI o licencia, y sin rutina pública que calcule las puntuaciones de seguridad de la Figura 4.

Research question

Can latent features interpreted as long-term social factors or short-term pressures be extracted and manipulated to change personality and safety scores in Gemma models, and do those effects differ across model scales?

Method

Activation intervention study on Gemma-2B-Instruct and Gemma-2-9B-Instruct. For background factors, GPT-4o generates contrastive descriptors and Gemma Scope SAE features active for one category and inactive for others are selected; their decoder vectors are added to the residual stream at layers 12 and 31. For short-term pressures, positive/negative prompt pairs are combined with TRAIT questions; the normalized activation difference of the last token is reduced with PCA and added as a control direction. Eight TRAIT subscales, seven SafetyBench categories, and logit changes between four pairs of social attributes are scored. Coefficients are chosen by observing option separation and the apparent stability of a concrete generation.

Sample: Two instructed models from a single family, Gemma 2B and Gemma 2 9B, with pretrained Gemma Scope SAEs. The background intervention covers 22 observable categories distributed across nine dimensions, although the main text inconsistently calls them "eight factors": two genders, two ages, three education levels, two socioeconomic levels, six ideologies, two emotional intelligence levels, two professional commitment levels, two family relationship levels, and only the AI familiarity category. The short-term intervention covers seven pressures. The external dataset publishes 8,000 Big Five items, 3,000 Dark Triad items, and 11,435 SafetyBench questions; the article does not specify how many effective observations enter each figure or independent repetitions.

Findings

  • The interventions markedly change option probabilities and TRAIT scores, providing causal evidence about the capacity of those vectors to alter responses under the protocol, not about an internal human personality.
  • The 9B shows background changes from 0 to 7.1 points and the 2B from 0 to 52.5; under short-term pressures the ranges are 0.1–27.7 and 0.4–53.5. The article calls this greater stability of the large model without estimating uncertainty or isolating size, data, or architecture.
  • The pressure directions produce large and heterogeneous changes: in 9B, achievement drive reduces Agreeableness by 7.2 points and raises Neuroticism by 18.4, while competence reduces Agreeableness by 27.7; these are score variations, not observed psychological states.
  • For the SAE factors, the logit separation tends to stabilize around an absolute coefficient of 250; 200 is chosen for 2B and 800 for 9B. For the PCA directions, 1.6 and 1.8 are chosen after exploring 0–10.
  • The article claims that reinforcing backgrounds reduces SafetyBench by between 0 and 6.8 points and that Offensiveness is the most sensitive category; the displayed matrix also contains −7.7, so the narrated maximum and the figure do not match.
  • Figure 2 shows logit difference increases from steering attributes such as gender, age, wealth, or ideology; this verifies control of target tokens, not that social bias improves or worsens generally.
  • The appendix recognizes severe degeneration with high coefficients: when applying 2,000 to a "female" feature, Gemma-2B repeats WOMAN dozens of times.

Limitations

  • Social determinism works as an analogy. The training data are not inspected, there is no intervention during pretraining, and it is not demonstrated that gender, ideology, poverty, or family are causes of the internal patterns found.
  • The semantic labels depend on phrases and explanations generated by GPT-4o. Activation on chosen phrases and inactivity on some contrasts does not guarantee monosemanticity, specificity, or absence of overlap.
  • The article itself counts "eight" background factors but enumerates nine dimensions; furthermore, the appendix requests Life satisfaction in the output format although it does not appear among the nine entries, and it only publishes familiarity descriptors, not non-familiarity descriptors, with AI.
  • The positive/negative pairs that construct the PCA directions are combined with TRAIT questions, and TRAIT is again the outcome measure. This overlap creates a risk of circularity and of directly steering the answers.
  • TRAIT measures response choice on human inventories. Stability across prompts, tasks, dialogues, times, languages, or contexts is not demonstrated, nor is equivalent expression in open text or behavior.
  • Coefficients are selected by observing logits of options from the benchmark itself and one generation, without a separate calibration set. This introduces fitting to the evaluated instrument and does not define an independent quality criterion.
  • There are no controls with random vectors, equal-norm directions, unrelated features, GPT-4o ablation, comparison with simple prompting, or placebo intervention.
  • Perplexity, fluency, general capability, invalid response rate, or out-of-task degradation is not measured; the appendix demonstrates that strong steering can push the model out of its normal regime.
  • No repetitions, complete experimental seeds, dispersion, confidence intervals, statistical tests, or correction for the large number of comparisons in the matrices are published.
  • The explanations of perfectionism, psychological tension, cognitive rigidity, overconfidence, or liberal permissiveness are not derived from a causal analysis and anthropomorphize score changes.
  • The logit difference between word pairs is directional and is calculated precisely after reinforcing one of the attributes; it does not by itself measure harm, stereotype, impartiality, or bias in real generation.
  • SafetyBench evaluates multiple-choice safety questions, not adversarial attacks or jailbreaks. The research question about susceptibility to jailbreak remains without a corresponding test.
  • The conclusion says that impacts on hallucination were evaluated, but the body presents no definition, benchmark, figure, or hallucination result; therefore, that claim is not supported by the published results.
  • The narrative claims safety decreases of up to 6.8 points, but Figure 4 includes a −7.7 in Offensiveness. No tabular data or scoring script is published to resolve the discrepancy.
  • Only Gemma 2B and 9B with Gemma Scope SAE are tested. The authors acknowledge that using already trained SAEs restricts generalization to other sizes, architectures, and interpretation methods.
  • The use of categories such as binary gender, poor/rich, educated/uneducated, stable/volatile intelligence, and six ideologies may incorporate normative stereotypes; there is no audit of construct validity, fairness, harm, or participation of the represented populations.
  • The official repository at commit 89b6500e6ab7da005ae576365e5fb2bfb5d39c51 does not pin versions and omits imported dependencies such as numpy, pandas, requests, openai, tqdm, tiktoken, jaxtyping, and repe. It also includes no tests, CI, code license, or releases.
  • The README invokes analysis.py, which does not exist, and documents arguments different from sae_run.py. The demo scripts require features not published within the repository and do not explain the download or placement of the external dataset.
  • The Hugging Face dataset does contain TRAIT and SafetyBench, but its viewer fails due to incompatible schemas. The SafetyBench code generates predictions and assigns a random option when it does not extract an answer, without seeding random; it does not include correct answers or a public scoring routine to reproduce Figure 4.
  • safety_sae_run.py passes freq_penalty to model.generate, a parameter that does not belong to the standard Transformers interface, and uses argparse type=bool for zero_shot, so the string "False" is interpreted as true. These paths are not covered by execution tests.
  • The repository documentation and citation retain a previous title and two authors who do not appear in version 2 of arXiv, an additional indication of drift between paper and artifact.

What the study does not establish

  • It does not demonstrate that LLMs possess personality, social history, ideology, gender, emotions, motivations, or experiences comparable to human ones.
  • It does not prove that the selected SAE features are monosemantic or that they causally represent the named constructs; it demonstrates sensitivity to descriptors and capacity for intervention.
  • It does not establish that the 9B model is psychologically more stable; it only shows smaller score changes under concrete coefficients and features.
  • It does not demonstrate resistance or vulnerability to jailbreak, reduction of hallucinations, or generative safety in adversarial scenarios.
  • It does not allow concluding that poverty, liberalism, or emotional instability cause model insecurity; those explanations are untested hypotheses.
  • It does not prove generalization to other models, SAEs, languages, benchmarks, open tasks, or longitudinal deployments.
  • It does not allow reproducing the main figures end to end with only the repository and its current instructions.

Traceability

Scope: Full text

Version: arXiv:2410.10863v2, submitted 7 October 2024, revised 16 February 2025, 21 pages; comments: under review

Consulted source: https://arxiv.org/pdf/2410.10863v2

Review: Codex full-text, bilingual-fidelity, 21-page visual, arXiv-v2, official-code, external-dataset, construct-validity, SAE-monosemanticity, circularity, coefficient-selection, statistical, safety, jailbreak, hallucination, bias, reproducibility and artifact-drift audit, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • Gemma-2B-Instruct
  • Gemma-2-9B-Instruct
  • Gemma Scope SAE, 16k and 131k widths
  • GPT-4o for factor descriptions and contrastive prompts

Instruments and metrics

  • TRAIT Big Five test set
  • TRAIT Short Dark Triad test set
  • SafetyBench English test set
  • Gemma Scope sparse autoencoders
  • Contrastive activation differences and PCA
  • Residual-stream activation steering
  • Paired-attribute next-token logit difference

Data used

  • TRAIT Big Five, 8,000 multiple-choice questions
  • TRAIT Dark Triad, 3,000 multiple-choice questions
  • SafetyBench English test set, 11,435 multiple-choice questions
  • GPT-4o-generated background descriptors and short-term pressure prompt pairs

Evidence and location

  • Research question, social determinism analogy, and contributions: Paper, pp. 1–3, Abstract, Introduction and Sections 2–3; Figure 1 and Table 1
  • Models, factors, SAE extraction, and PCA directions: Paper, pp. 4–6, Sections 4.1–4.2 and Figure 2
  • Personality change ranges and anthropomorphic explanations: Paper, pp. 7–8, Sections 4.3–4.4 and Figure 4
  • Results and safety discrepancy: Paper, p. 8, Figure 4a(C), Safety and Personality, and Conclusion
  • Descriptors, categories, and dependence on GPT-4o: Paper, pp. 14–19, Appendix C, especially C.2–C.3.1
  • Layer and coefficient selection; degeneration from over-steering: Paper, pp. 19–21, Appendix D, Table 3 and Figures 5–6
  • Family and SAE limitation: Paper, p. 9, Section 6 Limitations
  • Official code and documentation: Official GitHub repository commit 89b6500e6ab7da005ae576365e5fb2bfb5d39c51, README, requirements.txt and src/steer_experiments
  • External dataset content and schema failure: Hugging Face dataset Chouoftears/LLM-Persona-Steering-Testset revision 43cc72f6bdb555e39cba56999034cb5a90f428d1
  • Comprehensive visual inspection: Paper, all 21 rendered pages, including all tables, figures, appendices and code examples