Structured Personality Control and Adaptation for LLM Agents

Trait induction and control2026arXivApproved editorial review

Authors: Jinpeng Wang, Xinyu Jia, Wei Wei Heng, Yuquan Li, Binbin Shi, Qianlei Chen, Guannan Chen, Junxia Zhang, Yuyu Yin

Keywords: Large Language Models, Personality, Persona, Personality Control, AI Safety

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

9
Authors
6
Findings
30
Limitations
12
Evidence

Editorial summary

English

JPAF represents a persona through weights over eight Jungian functions (Te, Ti, Fe, Fi, Se, Si, Ne, and Ni). It initializes one dominant, one auxiliary, and six undifferentiated functions; uses dominant-auxiliary coordination for responses; adds 0.06 to temporary weights when a function is reinforced or compensates for a gap; and applies “reflection” rules to normalize weights or replace, swap, and reorganize functions. This is an explicit prompt-and-state controller: evolution is an author-designed algorithmic update to variables, not parameter learning or evidence of spontaneous psychological development.

The study tests GPT-4, Llama-4-Maverick, and Qwen3-235B-A22B-Instruct-2507 at temperature 0.6. Experiment 1 configures all 16 types and administers MBTI-93 and MBTI-70 five times. The baseline directly includes the MBTI label; JPAF omits the four letters but provides the dominant-auxiliary pair and detailed function descriptions. Mean per-dimension gain often favors JPAF, especially SN and JP: on MBTI-70, JP improves by 9.75 points for GPT and 13.38 for Qwen. Improvement is not universal: GPT loses 1.13 points on EI and 2.81 on TF in MBTI-93, while MBTI-70 TF falls by 0.19 for GPT and 0.44 for Qwen. The tables contain many individual cells where JPAF underperforms the baseline.

The phrase “100% MBTI alignment” does not mean 100% of questionnaire items aligned. It means that after JPAF each of the four proportions for every type falls above the 50% threshold and therefore recovers the intended four-letter label. The tables themselves contain substantially lower accuracies: GPT-INTP, for example, has 76.67% TF on MBTI-93 and Llama-ENFP has 60% JP on MBTI-70. The paper reports no confidence intervals, inferential tests, effect sizes, multiplicity correction, or distribution of the five runs, and does not control prompt length or information content. The evidence therefore supports steering questionnaire answers, not valid personality measurement or generation.

Experiment 2 uses eight author-created sets, one per function, with three scenarios and five questions per scenario: 15 questions for each function and starting configuration. GPT's mean TAA by function ranges from 92.91% to 100%; Qwen's from 88.75% to 100%; and Llama's from 65.83% to 95.41%. The paper reports PSA of 100% for GPT and Qwen and 92.19% for Llama. These scores are circular, however: the prompt lists and describes all eight functions, each scenario is expressly built for one target function, and the same model returns the selected function, “success,” and update in JSON. PSA calls a transition correct when it follows the theoretical rules imposed by the framework itself. There are no human annotators, independent judge, no-JPAF condition, blinded or out-of-distribution scenarios, or real behavior validating activation and evolution.

The rules generate the outcome later interpreted as evidence. Weights and thresholds are heuristic: B=0.06 is chosen as the midpoint of a feasible range, A=0.30 as a “clean” value, the increment is 0.06, and decay is 0.2, with no tuning, sensitivity analysis, or ablation. Normalization forces competition among functions; crossing a threshold triggers transformations; and the model decides whether a task succeeded or which auxiliary to select “based on prior data” that are not specified. In the examples, repeating fifteen Se-directed questions converts INTP first into ISTP and then ESTP; this demonstrates execution of the state machine, not emergent evolution. Persistence after scenario removal, manipulation resistance, unwanted drift, reversibility, safety, and user effects are not evaluated.

Construct validity is weak. MBTI is a dichotomous human self-report instrument transplanted to a text generator, while JPAF directly controls the descriptors later queried. The study does not examine Big Five measures, behavior, convergent or discriminant validity, longitudinal consistency, prediction, human judgment, realism, engagement, or trust. Claimed applications in education, health care, therapy, or sustained relationships are design possibilities rather than study results. Coverage is limited to English, artificial text tasks, three model families, and one temperature; there are no users, natural long-term dialogue, tool use, multi-agent interaction, or cultural-risk tests.

Reproducibility is insufficient. Models, temperature, and partial prompts are named, but “gpt-4” has no snapshot, provider, or date; all other parameters are defaults; and no seeds are reported despite random BaseWeight sampling. No code, fully sourced questionnaires, complete scenarios, outputs, logs, initial weights, reflection history, or machine-readable results are released. The official arXiv surface exposes only v1, PDF, HTML, and TeX and links no repository or executable artifact. The PDF retains an unfinished ACM template: a 2018 reference, conference and DOI placeholders, 2018 copyright, and received/revised/accepted dates from 2007–2009. It is an arXiv v1 preprint, not a confirmed ACM publication. JPAF is an interesting structured-control specification, but the available evidence does not justify treating it as validated psychological personality, autonomous evolution, or a system ready for sensitive use.

Español

JPAF representa una persona mediante pesos sobre ocho funciones jungianas (Te, Ti, Fe, Fi, Se, Si, Ne y Ni). Inicializa una función dominante, una auxiliar y seis indiferenciadas; usa coordinación dominante-auxiliar para responder, suma 0,06 a pesos temporales cuando una función se refuerza o compensa una carencia y aplica reglas de “reflexión” para normalizar pesos o sustituir, intercambiar y reorganizar funciones. Es un controlador de prompts y estado explícito: la evolución es una actualización algorítmica de variables diseñada por los autores, no aprendizaje de parámetros ni evidencia de desarrollo psicológico espontáneo.

El estudio prueba GPT-4, Llama-4-Maverick y Qwen3-235B-A22B-Instruct-2507 a temperatura 0,6. En el experimento 1 configura los 16 tipos y administra MBTI-93 y MBTI-70 cinco veces. El baseline incluye directamente la etiqueta MBTI; JPAF omite las cuatro letras pero aporta el par dominante-auxiliar y descripciones detalladas de las funciones. La ganancia media por dimensión suele favorecer JPAF, sobre todo SN y JP: en MBTI-70, JP mejora 9,75 puntos con GPT y 13,38 con Qwen. No es una mejora universal: GPT pierde 1,13 puntos en EI y 2,81 en TF en MBTI-93, y en MBTI-70 TF cae 0,19 con GPT y 0,44 con Qwen. Las tablas muestran además múltiples celdas donde JPAF rinde peor que el baseline.

La frase “100 % de alineación MBTI” no significa 100 % de ítems alineados. Significa que, tras JPAF, las cuatro proporciones de cada tipo quedan por encima del umbral de 50 % y por ello recuperan las cuatro letras esperadas. Las propias tablas contienen accuracies bastante menores: por ejemplo GPT-INTP obtiene 76,67 % en TF con MBTI-93 y Llama-ENFP 60 % en JP con MBTI-70. El trabajo no publica intervalos, tests inferenciales, tamaños de efecto, corrección por multiplicidad ni distribución de los cinco runs, y no controla longitud o cantidad de información del prompt. Por ello sostiene que JPAF puede orientar respuestas de estos cuestionarios, no que mida o produzca personalidad válida.

El experimento 2 usa ocho conjuntos creados por los autores, uno por función, con tres escenarios y cinco preguntas por escenario: 15 preguntas para cada función y configuración. GPT obtiene TAA media por función entre 92,91 % y 100 %; Qwen, entre 88,75 % y 100 %; Llama, entre 65,83 % y 95,41 %. El artículo informa PSA de 100 % para GPT y Qwen y 92,19 % para Llama. Sin embargo, estos scores son circulares: el prompt enumera y describe las ocho funciones, cada escenario se construye expresamente para una función objetivo, y el mismo modelo devuelve en JSON la función elegida, “success” y la actualización. PSA considera correcta una transición si respeta las reglas teóricas que el propio framework impone. No hay anotadores humanos, juez independiente, condición sin JPAF, escenarios ciegos o fuera de distribución, ni conducta real que valide activación o evolución.

Las reglas producen el resultado que luego se interpreta. Pesos y umbrales son heurísticos: B=0,06 se escoge como punto medio factible, A=0,30 como valor “limpio”, el incremento es 0,06 y el decaimiento 0,2, sin ajuste, sensibilidad o ablación. Normalizar obliga competencia entre funciones; superar un umbral activa transformaciones; y el modelo decide cuándo una tarea fue exitosa o qué auxiliar seleccionar “basándose en datos previos” no especificados. En los ejemplos, repetir quince preguntas dirigidas a Se convierte INTP primero en ISTP y después en ESTP; esto demuestra la ejecución de la máquina de estados, no una evolución emergente. Tampoco se evalúan estabilidad después de retirar el escenario, resistencia a manipulación, drift no deseado, reversibilidad, seguridad o efectos sobre usuarios.

La validez de constructo es débil. MBTI es un autoinforme humano dicotómico trasladado a un generador de texto, mientras JPAF controla directamente los mismos descriptores que luego pregunta. No se estudian Big Five, conducta, validez convergente o discriminante, consistencia longitudinal, predicción, juicio humano, realismo, engagement o confianza. Las afirmaciones de utilidad en educación, salud, terapia o relaciones sostenidas son posibilidades de diseño, no resultados del estudio. La muestra cubre inglés, tareas textuales artificiales, tres familias y una sola temperatura; no hay interacción con usuarios, diálogo natural prolongado, tool use, multiagente o riesgos culturales.

La reproducibilidad es insuficiente. Se nombran modelos, temperatura y parte de los prompts, pero “gpt-4” carece de snapshot, proveedor y fecha; los demás parámetros quedan por defecto; no hay seeds pese al muestreo aleatorio de BaseWeights; no se publican código, cuestionarios completos con procedencia verificable, escenarios completos, respuestas, logs, pesos iniciales, historial usado por reflexión ni resultados máquina-legibles. La superficie oficial de arXiv solo identifica v1, PDF, HTML y TeX; no enlaza un repositorio o artefacto ejecutable. El PDF conserva una plantilla ACM sin completar: referencia 2018, conferencia y DOI placeholder, copyright 2018 y fechas recibido 2007/revisado 2009/aceptado 2009. Es un preprint arXiv v1, no una publicación ACM confirmada. JPAF es una especificación interesante de control estructurado, pero la evidencia disponible no justifica presentarlo como personalidad psicológica validada, evolución autónoma o sistema listo para uso sensible.

Research question

Can a prompt and weight controller inspired by Jungian functions induce expected MBTI labels and execute predefined type adaptations under directed scenarios?

Method

JPAF assigns weights to eight functions, coordinates dominant and auxiliary, sums temporal weights by reinforcement or compensation, and applies reflection rules to change the hierarchy. It is compared with an MBTI label prompt in two questionnaires, five runs, and three models. Then it is tested without baseline in eight families of self-constructed scenarios, where the model itself selects function, success, and transition.

Sample: Three models, 16 MBTI configurations, two questionnaires, and five runs in experiment 1. Experiment 2 crosses 16 initial configurations with eight functions and 15 questions per function; the TAA tables advance in steps of 6.67 %. No verifiable total number of calls, seeds, distribution per run, cost, tokens, API errors, or complete history is reported.

Findings

  • JPAF improves the mean DAG in most dimensions and models, especially SN and JP, but presents regressions in some dimensions and many individual cells.
  • In MBTI-70, the JP gain is +9.75 points for GPT and +13.38 for Qwen; TF falls 0.19 and 0.44 points respectively.
  • The reported 100 % is recovery of the four-letter type by exceeding 50 % in each dimension, not 100 % accuracy on the items.
  • GPT achieves mean TAA per function of 92.91–100 %, Qwen 88.75–100 %, and Llama 65.83–95.41 % in directed scenarios.
  • Reported PSA is 100 % for GPT and Qwen and 92.19 % for Llama under the internal transition rules.
  • The results show that strong models follow the schema and its JSON outputs; there is no independent validation of the construct.

Limitations

  • The evaluation is aligned with the mechanism: Jungian descriptors are controlled and then the same descriptors are asked about.
  • The baseline receives a brief MBTI label and JPAF a functional pair with extensive descriptions; information, length, or token budget is not equalized.
  • There is no ablation of dominant-auxiliary, weights, memory, reflection, normalization, or prompt length.
  • B=0.06, A=0.30, delta=0.06, and decay=0.2 are heuristic choices without sensitivity or calibration.
  • BaseWeights are sampled randomly but no seed, effective distribution, or initial weights per run are reported.
  • Normalization and thresholds produce by design hierarchy changes after repeated activations.
  • TAA mixes scenario quality, instruction following, and model self-classification; it does not isolate adaptation.
  • The prompt delivers the definitions of the eight functions and the model chooses which one the scenario expressly designed for it requires.
  • The same model declares the function, success, reason, and update; there is no independent judge or human annotation.
  • PSA verifies conformity with rules constructed by the authors, not psychological validity or external utility.
  • Experiment 2 lacks a baseline without JPAF, alternative controller, blind scenarios, negative scenarios, or out-of-distribution scenarios.
  • Changes are not compared with observable behavior, task objectives, human preferences, or known profiles.
  • There is no convergent, discriminant, predictive, incremental, factorial, or longitudinal validity.
  • Persistence after withdrawing the stimulus, reversibility, drift, stability, or error accumulation is not quantified.
  • Prompt attacks, adversarial activations, unwanted changes, or safe state recovery are not studied.
  • There are no users or measures of engagement, trust, realism, well-being, or HCI effects.
  • Applications to education, health, therapy, and sustained relationships are speculative.
  • MBTI and Jungian functions are treated as a normative structure without sufficiently discussing their psychometric limits.
  • The study only covers English, text, one temperature, three families, and artificial scenarios of 15 turns.
  • There is no multiagent evaluation, tool use, real persistent memory, long context, languages, or cultures.
  • gpt-4 has no snapshot, provider, or date; Llama-4-Maverick also lacks a complete executable identifier.
  • “Default parameters” does not ensure comparability between providers and does not document decoding, max tokens, or system settings.
  • Five runs are averaged without intervals, tests, effect sizes, or publication of individual values.
  • No adjustment is made for the numerous comparisons between types, dimensions, questionnaires, and models.
  • DAR counts exact equalities of percentages and many are ceiling at 100 %, so it is not a strong measure of stability.
  • The 100 % alignment lends itself to confusion with item accuracy; tables show JPAF scores much lower than 100 %.
  • No code, traceable complete questionnaires, complete scenarios, responses, logs, history, or structured results are published.
  • No repository or official executable artifact linked from arXiv or the paper was found.
  • The ACM template contains fictitious or unupdated year, conference, DOI, copyright, and reception dates.
  • The official surface only confirms arXiv v1; it does not confirm venue or peer review.

What the study does not establish

  • It does not establish that an LLM possesses a human personality or a Jungian psychological structure.
  • It does not demonstrate autonomous evolution; it executes updates of a designed state machine.
  • It does not validate that weight changes correspond to persistent or beneficial behavioral changes.
  • It does not prove that JPAF is superior to a prompt of equal length and information or to other controllers.
  • It does not demonstrate safety, stability, reversibility, or adequacy for education, health, or therapy.
  • It does not allow fully reproducing the results with the identified public artifacts.
  • It does not confirm publication or peer review in ACM.

Traceability

Scope: Full text

Version: arXiv:2601.10025v1, submitted 15 January 2026; preprint, 31 pages

Consulted source: https://arxiv.org/pdf/2601.10025v1

Review: Codex full-text, bilingual-fidelity, 31-page visual, arXiv-v1, prompt-control, MBTI-threshold, baseline-fairness, TAA-self-label, PSA-rule-circularity, state-machine, construct-validity, HCI-claim, safety, reproducibility, artifact and publication-status audit, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • gpt-4, snapshot and provider unspecified
  • Llama-4-Maverick, checkpoint/provider unspecified
  • qwen3-235b-a22b-instruct-2507

Instruments and metrics

  • Jungian Personality Adaptation Framework (JPAF)
  • Eight Jungian cognitive functions with BaseWeight and TemporaryWeight
  • Dominant-auxiliary coordination
  • Reinforcement-compensation with delta weight 0.06
  • Rule-based reflection and weight normalization
  • MBTI-93 binary questionnaire
  • MBTI-70 binary questionnaire
  • Dimension Accuracy Gain (DAG)
  • Dimension Agreement Rate (DAR)
  • Type Activation Accuracy (TAA)
  • Personality Shift Accuracy (PSA)
  • Model-generated JSON self-classification and state update

Data used

  • MBTI-93 questionnaire, public source not identified sufficiently for reconstruction
  • MBTI-70 questionnaire, public source not identified sufficiently for reconstruction
  • Eight author-designed function-specific scenario sets
  • Three scenarios and five questions per function set, 15 questions per target function
  • Partial prompts and scenario descriptions in Appendix A.2-A.5

Evidence and location

  • Metadata, abstract, version, and status: Official arXiv:2601.10025v1 surface, submitted 15 January 2026
  • Weights, ranges, and mechanisms: Paper, pp. 5–11, Section 3.2, Equations 1–6 and Algorithm 1
  • Models, parameters, questionnaires, and scenarios: Paper, pp. 10–12, Sections 4.1–4.2
  • Definition of DAG, DAR, TAA, and PSA: Paper, pp. 12–13, Section 4.3, Equations 7–10
  • Questionnaire results and meaning of the threshold: Paper, pp. 13–16 and 24–27, Section 5.1, Figures 6–9 and Tables 1–6
  • TAA and PSA: Paper, pp. 15–18 and 29–30, Section 5.2, Figures 10–12 and Tables 7–9
  • Examples of transformations by repetition: Paper, pp. 17–22, Figures 13–20
  • Content and circularity of prompts: Paper, pp. 28–31, Appendix A.2–A.4, Figures 21–24
  • Scenario coverage: Paper, p. 31, Appendix A.5, Table 10
  • Editorial template errors: Paper, pp. 1 and 31, ACM placeholders and inconsistent 2007–2009 dates
  • Comprehensive visual inspection: Paper, all 31 rendered pages, including every table, figure, prompt, and appendix page
  • Absence of linked official artifact: Paper and official arXiv surface checked 15 July 2026; no repository or executable artifact identified