Synthetic Personality Review

SAC: A Framework for Measuring and Inducing Personality Traits in LLMs with Dynamic Intensity Control

Trait induction and control2026scitepress.orgApproved editorial review

Authors: Adithya Chittem, Aishna Shrivastava, Sai Tarun Pendela, Jagat Sesh Challa, Dhruv Kumar

Keywords: Computation and Language, Artificial Intelligence, Human-Computer Interaction

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

5
Authors
9
Findings
12
Limitations
11
Evidence

Editorial summary

English

The paper introduces PERS-16, an adaptation of 163 IPIP items to sixteen 16PF-inspired labels, and Specific Attribute Control (SAC), a prompting scheme that anchors trait intensity from 1 to 5 with definitions, adjectives, and behavioral questions. It first profiles GPT-4o, Claude 3.7 Sonnet, Gemini 2.5 Flash, and a model identified only as Mistral, then compares binary P2 induction with SAC-neutral and SAC-induced. Its clearest contribution is operational: under SAC, target-trait questionnaire responses are generally ordered from intensity 1 through 3 to 5 across the four reported models. This is evidence of graded instruction following within the instrument, not evidence of a latent personality or validated continuous control outside the prompt.

PERS-16 yields different baseline profiles and P2 is inconsistent: Claude has negative deltas for 11 of 16 traits, Gemini for 10, GPT-4o for 3, while Mistral is mostly positive. SAC-induced asks 11,520 questions per model, and levels 1 and 5 tend to fall below and above baseline respectively. Yet level 3 does not consistently recover the neutral state: all target deltas are negative for GPT-4o and nearly all are negative for Claude. The figures therefore support ordinal steering at three tested anchors, but not metric calibration, symmetry, or real-time dynamic intensity control.

SAC-neutral is descriptively compared with 30 adults, and the two largest absolute non-target shifts are selected as co-movers for each induced trait. The human comparison contains only neutral means and dispersion, without inferential tests, individual data, or human induction trajectories. Co-movers are post hoc and the prompt itself supplies every intensity anchor, says that changing one trait may affect other traits, and asks the model to assess an observed trait relative to the target. Their covariance can therefore arise from prompt semantics and demand characteristics rather than an internalized human psychological structure.

Open publication and supplementary material are strengths, but artifact auditing found material paper-code discrepancies: SAC scripts use temperature 0.7 where the paper reports 0; SAC-neutral mutates questions cumulatively inside its factor loop; the later Mistral SAC-induced script actually calls Gemini 2.5 Flash Preview through OpenRouter; the Mistral version and size are unspecified; and human data, a license, and a reproducible dependency environment are absent. SAC is consequently a promising test of explicit prompt control over self-report answers, while the Mistral attribution and claims of human alignment, latent structure, and fine-grained calibrated control remain unverified.

Español

El artículo presenta PERS-16, una adaptación de 163 ítems IPIP a dieciséis etiquetas inspiradas en 16PF, y Specific Attribute Control (SAC), un esquema de prompting que fija intensidades de rasgo de 1 a 5 mediante definiciones, adjetivos y preguntas conductuales. Primero perfila GPT-4o, Claude 3.7 Sonnet, Gemini 2.5 Flash y un modelo denominado solo Mistral; después compara una inducción binaria P2 con SAC-neutral y SAC-induced. La aportación más clara es operacional: el protocolo SAC logra que las respuestas de cuestionario del rasgo objetivo se ordenen generalmente de intensidad 1 a 3 y a 5 en los cuatro modelos reportados. Esto demuestra seguimiento graduado de instrucciones en el instrumento, no una personalidad latente ni control continuo validado fuera del propio prompt.

PERS-16 muestra perfiles basales distintos y P2 resulta inestable: Claude presenta deltas negativos en 11 de 16 rasgos, Gemini en 10 y GPT-4o en 3, mientras Mistral produce sobre todo cambios positivos. En SAC-induced cada modelo responde 11.520 preguntas; los extremos 1 y 5 tienden a quedar, respectivamente, por debajo y por encima del baseline. Sin embargo, el nivel 3 no reproduce de forma consistente el estado neutral: en GPT-4o todos los deltas objetivo son negativos y en Claude casi todos lo son. Por eso las figuras apoyan ordenación ordinal de tres anclajes, pero no simetría, calibración métrica ni una intensidad dinámica que pueda elegirse con precisión en tiempo real.

El trabajo también compara SAC-neutral con 30 adultos y selecciona, para cada rasgo inducido, los dos mayores cambios absolutos en otros rasgos como co-movers. La comparación humana se limita a medias y dispersión descriptivas del estado neutral, sin contraste inferencial, datos individuales ni trayectorias inducidas humanas. Los co-movers son exploratorios y se eligen después de observar los resultados; además, el propio prompt enumera todos los niveles, afirma que cambiar el rasgo objetivo puede afectar a otros y pide valorar cada rasgo observado en relación con el objetivo. Así, la covariación puede explicarse por semántica y demanda de la instrucción y no prueba que el modelo haya internalizado una arquitectura psicológica humana.

La publicación abierta y el suplemento son valiosos, pero la auditoría del artefacto revela diferencias materiales con el artículo: scripts SAC usan temperatura 0,7 donde el texto declara 0; SAC-neutral modifica acumulativamente las preguntas dentro del bucle; el código añadido como Mistral para SAC-induced llama en realidad a Gemini 2.5 Flash Preview mediante OpenRouter; no se fija la versión o tamaño de Mistral; y faltan datos humanos, licencia y entorno de dependencias reproducible. En consecuencia, SAC es una prueba prometedora de control de respuestas de autoinforme por prompting explícito, pero la reproducibilidad del resultado atribuido a Mistral y las afirmaciones de alineación humana, estructura latente y control fino quedan sin establecer.

Research question

Can an evaluation inspired by 16PF measure LLM response profiles, and can a prompt with semantic anchors gradually modulate a target trait, surpassing the binary induction P2 and producing interpretable covariations with other traits?

Method

The study constructs PERS-16 with 163 IPIP statements scored from 1 to 5 and profiles four models in stateless mode. It then induces each of the 16 traits separately through P2 and re-administers the 163 items. SAC-neutral combines 16 observed traits, five intensity factors, and two questions to obtain 160 responses per model; SAC-induced crosses 16 target traits, intensities 1, 3, and 5, five factors, three behavioral actions, and 16 observed traits, totaling 11,520 responses per model. Means, dispersion, Euclidean distances, and deltas against the baseline are calculated. Thirty adults complete SAC-neutral as a descriptive reference. To explore covariation, the two non-target traits with the greatest absolute delta are chosen post hoc. The review contrasted the full PDF of proceedings and arXiv, figures and tables, the Zenodo deposit, and the current code repository.

Sample: PERS-16 uses 163 items for each model. P2 generates 16 × 163 = 2,608 responses per model. SAC-neutral administers 16 traits × 5 factors × 2 questions = 160 responses per model. SAC-induced administers 16 target traits × 3 intensities × 5 factors × 3 behavioral prompts × 16 observed traits = 11,520 responses per model. The human reference includes 30 adults, 14 men and 16 women, all over 18 years of age and pursuing or holding a university degree, who respond to the same 160 SAC-neutral questions; their individual responses are not published.

Findings

  • In PERS-16, GPT-4o and Gemini are the closest profiles by Euclidean distance (1.50); Claude and Mistral are the most distant (3.94). These are descriptive distances of responses to items, not measures of global psychological similarity.
  • The dispersion of PERS-16 varies substantially by trait and model; calling the standard deviation a proxy for internal consistency is not psychometrically correct, because dispersion among items and the reliability of a scale are distinct concepts.
  • P2 does not uniformly induce the high extreme: Claude yields 11 negative deltas out of 16, Gemini 10, and GPT-4o 3; Mistral is mostly positive. The tables titled for three LLMs contain four models.
  • SAC produces a clear ordinal relationship: for the majority of traits, the target score increases when moving from anchors 1 to 3 and to 5 across the four reported models.
  • Level 3 is not a calibrated zero relative to the baseline: GPT-4o shows all 16 target deltas negative and Claude almost all negative; Gemini is mixed. The article's expression "around zero" conceals systematic shifts visible in the figure and the data.
  • The extremes 1 and 5 separate the responses more clearly than P2, which supports SAC as prompt engineering to control how the model answers the same questionnaire.
  • The comparison with 30 humans places several neutral means near the center of the scale, but offers no statistical tests, measurement equivalence, or a normative sample that would allow concluding human alignment.
  • Co-movers are defined as the two largest absolute changes observed; this rule guarantees selecting some covariation even under noise, and Table 6 only interprets 13 of the 16 targets, omitting Sensitivity, Imagination, and Orderliness.
  • The audit of the supplement prevents verifying the SAC-induced result for Mistral: the script labeled as Mistral calls the model google/gemini-2.5-flash-preview, while other scripts use a generic Ollama name "mistral" without version or size.

Limitations

  • All main results come from forced multiple-choice self-report. It is not validated that the scores predict behavior in open conversation, tasks, decisions, safety, or longitudinal interaction.
  • The SAC prompt explicitly reveals the target trait, its desired level, and the adjectives of the five levels, and instructs to answer accordingly. The effect may be simple semantic obedience and not a modification of an internal state.
  • The prompt also indicates that changing the target may affect other traits and asks to evaluate the observed trait in relation to that target; this introduces direct demand into the co-movement analysis.
  • Only three induction levels are tested, 1, 3, and 5. Levels 2 and 4 are not evaluated, nor are sufficient points to justify continuity, linearity, or fine control across the entire 1–5 scale.
  • Co-movers are selected post hoc by absolute magnitude, with no prior hypothesis of covariance, null model, significance, confidence intervals, correction for multiple comparisons, replication, or cross-validation.
  • The human reference is small, educated, and demographically limited; detailed recruitment, ethics/consent, individual responses, reliability, uncertainty, and inferential model-human contrast are missing.
  • The standard deviation of items does not estimate internal consistency. No alpha or omega, factor analysis, invariance, convergent/discriminant validity, or psychometric calibration of PERS-16 or SAC is provided.
  • The text inconsistently alternates between three and four LLMs, between two and three questions in SAC-neutral, and omits three traits from the interpretive co-movers table.
  • The article declares temperature 0, top-p 0.95, and maximum reproducibility, but SAC scripts use temperature 0.7 and different configurations across providers; nor are snapshots of the proprietary APIs fixed.
  • In several SAC-neutral scripts, the two questions are cumulatively modified within the loop of five factors, so that subsequent runs may contain concatenated prefixes and not represent the described procedure.
  • The Zenodo deposit and the repository do not contain raw human data, a license, a dependency manifest or lockfile, or a single pipeline that regenerates figures and tables; the Zenodo ZIP also does not cleanly correspond to its embedded Git commit.
  • "Dynamic" here means choosing a textual anchor in advance. Adaptive control in real time, multi-trait and context-dependent, is expressly reserved for future work.

What the study does not establish

  • It does not demonstrate that an LLM possesses the sixteen traits as internal states, stable or equivalent to human personality.
  • It does not demonstrate that SAC controls a continuous psychological variable: it only tests three discrete prompt conditions and responses to the instrument itself.
  • It does not demonstrate symmetry or calibration of intensity 3 relative to the neutral baseline.
  • It does not demonstrate that co-movers reflect an internalized latent architecture or human 16PF correlations, because there is no equivalent human comparison or semantic demand control.
  • It does not demonstrate psychometric validity or reliability of PERS-16; means and standard deviations do not replace those analyses.
  • It does not allow reproducing or safely attributing the SAC-induced results presented as Mistral.
  • It does not establish generalization to other models, languages, instruments, prolonged conversations, users, or deployment domains.
  • It does not evaluate safety consequences, persuasion, bias, well-being, dependence, deception, or effects on task performance when inducing traits.

Traceability

Scope: Full text

Version: arXiv:2506.20993v2 (12 Jan 2026); ICAART 2026 proceedings, DOI 10.5220/0014415000004052, also reviewed; Zenodo 10.5281/zenodo.15487993 and GitHub commit 15fb883951562df25889f78de06717f8f1ea0e58 audited

Consulted source: https://arxiv.org/pdf/2506.20993

Review: Codex editorial review and methods/code audit, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • GPT-4o
  • Claude 3.7 Sonnet
  • Gemini 2.5 Flash
  • Mistral (versión y tamaño no especificados)

Instruments and metrics

  • PERS-16 (163 IPIP-derived items across 16 traits)
  • 16PF-inspired trait taxonomy
  • Machine Personality Inventory / P2 prompting adaptation
  • Specific Attribute Control (SAC-neutral and SAC-induced)
  • Euclidean distance between 16-trait profiles
  • Trait deltas relative to neutral baseline
  • Post-hoc top-two absolute co-mover selection

Data used

  • PERS-16 item set
  • SAC adjective and behavioral-anchor dataset
  • SAC-neutral model responses
  • SAC-induced model responses
  • Human SAC-neutral baseline (aggregate results only)

Evidence and location

  • Definitive publication, DOI, authorship, and pagination: SCITEPRESS ICAART 2026 paper 144150, DOI 10.5220/0014415000004052, pp. 3267–3278
  • Construction, scoring, and declared configuration of PERS-16: Paper, pp. 3268–3269, section 3.1, Equation 1 and Figure 1
  • Neutral profiles, dispersion, and distances between models: Paper, pp. 3269–3270, Tables 1–2, sections 3.2–3.3
  • P2 design, 2,608 responses, and deltas per model: Paper, pp. 3270–3272, section 4, Tables 3–4 and Equation 3
  • SAC-neutral design, size, and reference of 30 participants: Paper, pp. 3272–3273, section 5.1 and Table 5
  • SAC-induced design and total of 11,520 questions per model: Paper, pp. 3273–3274, section 5.2, Figure 4 and Equation 4
  • Trajectories of intensities 1, 3, and 5 and lack of exact alignment of level 3: Paper, pp. 3274–3275, Figure 5 and accompanying discussion; supplementary SAC delta CSVs
  • Post hoc selection and interpretation of co-movers: Paper, pp. 3274–3277, Figures 6–7 and Table 6
  • Declared limitations, discussion, and the future nature of real-time control: Paper, pp. 3275 and 3277–3278, sections 6–8
  • Availability and content of the supplement: Zenodo record 15487993, DOI 10.5281/zenodo.15487993, archive SHA-256 8bb7e9ff287f073c9323a5def1dcf3a85fbf7c1884d735c3f66ba458c4c9de8
  • Differences between published method and code, including the route attributed to Mistral: GitHub repository aishna0310/SAC at commit 15fb883951562df25889f78de06717f8f1ea0e58; MPI/models/*/personality_prompting/sac.py and SAC-neutral scripts