Behavior-Adaptive Conversational Agents: Toward a Fluid Personality Framework

Personas, identity, and agents2026arXivApproved editorial review

Authors: Hasibur Rahman, Smit Desai

Keywords: Low Agreeableness, Persona conditioning, Safe fine-tuning, Conversational warmth, Jailbreak robustness, Data composition, LoRA, Refusal evaluation

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

2
Authors
5
Findings
6
Limitations
6
Evidence

Editorial summary

English

This short position paper, presented at the AAAI 2026 Bridge Program, proposes the Fluid Personality Framework so that a conversational agent does not keep a fixed identity and style but jointly adapts two dimensions on every turn. The first is metaphorical persona: the role the system presents to the user, such as planner, coach, cheerleader, tutor, friend, expert, library, guide, or tool. The second is personality-expression intensity, framed as low, medium, or high and realized through tone, affect, formality, concision, enthusiasm, agreeableness, or conscientiousness. The framework takes contextual factors, task type, domain, and urgency, and user-specific factors, traits and interaction history, as inputs. A persona module would select an appropriate role and use linguistic transitions to preserve coherence; a personality module would regulate stylistic strength within that role. The authors illustrate the concept with an agent that shifts from planner during goal setting to cheerleader when celebrating progress and tutor when explaining, or reduces agreeable small talk while increasing concision and conscientiousness for an urgent medical reminder. The motivation comes from previous literature, including the authors' own work: interface-metaphor studies suggest anthropomorphism does not always increase trust; a metaphor-fluid voice-agent study reports improved enjoyment, likability, and adoption intention without changes in perceived trust or intelligence; and a cited CHI 2026 study describes an inverted-U relationship in which medium personality expression outperforms low and high expression on user ratings. This paper combines those lines into a design schema, but it does not implement the selector, define a decision function, specify how user traits would be inferred safely, provide a prototype, or run a model evaluation, simulation, or human study. It also leaves unresolved how role changes should be disclosed, consent obtained, sensitive inferences prevented, stereotypes avoided, recognizable identity preserved, drift measured, and optimization goals governed. The claim that the framework can improve experience and effectiveness is a design hypothesis supported by external studies, not a result of the present paper. Its contribution is conceptual: it separates “who the agent appears to be” from “how intensely it expresses itself” and argues that both should be adapted together.

Español

Este trabajo breve de posición, presentado en el Bridge Program de AAAI 2026, propone el Fluid Personality Framework para que un agente conversacional no mantenga una identidad y un estilo fijos, sino que adapte conjuntamente dos dimensiones en cada turno. La primera es la persona metafórica: el papel que el sistema representa ante el usuario, por ejemplo planificador, entrenador, animador, tutor, amigo, experto, biblioteca, guía o herramienta. La segunda es la intensidad con la que expresa rasgos o señales de personalidad, planteada en niveles bajos, medios o altos y manifestada mediante tono, afecto, formalidad, concisión, entusiasmo, amabilidad o responsabilidad. El marco recibe factores contextuales, tipo de tarea, dominio y urgencia, y factores específicos de usuario, rasgos e historial de interacción. Un módulo de persona seleccionaría el rol apropiado y usaría transiciones lingüísticas para mantener coherencia; un módulo de personalidad regularía la fuerza estilística dentro de ese rol. Los autores ilustran el concepto con un agente que pasa de planificador durante la fijación de objetivos a animador al celebrar avances y tutor al explicar, o que reduce la charla amable y aumenta concisión y responsabilidad ante un recordatorio médico urgente. La motivación procede de literatura previa, incluida investigación de los propios autores: estudios sobre metáforas de interfaz sugieren que antropomorfizar no siempre aumenta confianza; un trabajo sobre personas fluidas en voz informa mejoras de disfrute, agrado e intención de adopción sin cambios en confianza o inteligencia percibida; y un estudio citado como CHI 2026 describe una relación de U invertida donde una expresión media de personalidad supera a niveles bajos y altos en evaluaciones de usuario. El artículo integra esas líneas en un esquema de diseño, pero no implementa el selector, no define una función de decisión, no especifica cómo inferir de forma segura rasgos del usuario, no aporta un prototipo y no realiza experimento, simulación, evaluación de modelos ni estudio con personas. Tampoco resuelve cómo comunicar los cambios de rol, obtener consentimiento, impedir inferencias sensibles, evitar estereotipos, conservar una identidad reconocible, medir deriva o decidir quién optimiza el objetivo del sistema. La conclusión de que el marco puede mejorar experiencia y efectividad es una hipótesis de diseño apoyada en estudios externos, no un resultado del trabajo presente. Su valor está en separar claramente «quién parece ser el agente» de «con qué intensidad se expresa» y obligar a adaptar ambas dimensiones de forma coordinada; su evidencia propia es conceptual.

Research question

How should a conversational agent jointly adapt the metaphorical role it represents and the intensity of its personality expression according to the task, domain, urgency, objectives, traits, and history of the user?

Method

A four-page conceptual article that synthesizes two lines of literature, metaphorical persona design and trait expression, and organizes them into a two-module framework. It does not develop software, a formal algorithm, a dataset, an experiment, a quantitative analysis, or an evaluation with users.

Sample: There is no empirical sample, evaluated models, or participants. The examples of planner, animator, tutor, and medical reminder are illustrative scenarios.

Findings

  • The own contribution is a conceptual architecture with two orthogonal axes: metaphorical persona selection and trait expression intensity.
  • The framework proposes recalculating both axes before each turn based on task, domain, urgency, user traits, and interaction history.
  • The persona module would maintain a repertoire of roles and transitions; the personality module would modulate tone, affect, formality, enthusiasm, agreeableness, and conscientiousness within the chosen role.
  • The claims about preference for medium intensity, advantages of fluid personas, and effects of mirroring come from cited works, not from data collected in this article.
  • There is no experimental result that compares the fluid framework with a static assistant or that measures safety, trust, efficacy, or behavior change.

Limitations

  • The framework is not implemented, nor is an algorithm, an adaptation policy, a state representation, an objective function, a change threshold, or a mechanism to resolve contradictory signals defined.
  • There is no empirical evaluation, prototype, models, tasks, participants, metrics, or analysis; the expected improvements are hypotheses extrapolated from external studies.
  • The proposal assumes availability of user traits and history, but does not explain consent, privacy, accuracy, updating, right to correction, or protection of sensitive attributes.
  • The costs of frequent changes are not studied: confusion, loss of identity, manipulation, overpersonalization, reinforcement of stereotypes, dependence, or miscalibration of trust.
  • The low/medium/high levels are not operationalized or linked to a psychometric scale or output measurement, so "intensity" still lacks a reproducible definition.
  • The examples of health and behavior change are risk domains, but the article does not propose human supervision, clinical limits, auditing, or harm evaluation.

What the study does not establish

  • It does not demonstrate that a fluid personality improves experience, efficacy, trust, adherence, or behavior change outcomes.
  • It does not demonstrate that medium intensity is optimal across all users, tasks, cultures, languages, or traits; that expectation comes from a cited and contextual study.
  • It does not prove that inferring or mirroring the user's personality is accurate, desirable, safe, or ethically legitimate.
  • It does not offer a technology ready to deploy or a benchmark to evaluate implementations of the framework.
  • It does not establish that metaphorical roles correspond to internal personalities of the model; they are presentation and interaction decisions.
  • It does not resolve how to maintain transparency and user agency when the system changes role or style dynamically.

Traceability

Scope: Full text

Version: arXiv:2607.01034v1; AAAI-2026 Bridge Program position paper

Consulted source: https://arxiv.org/pdf/2607.01034

Review: Codex editorial review, 2026-07-14

Approval: Codex fidelity pass, 2026-07-14

English translation: approved, 2026-07-18

Evidence and location

  • Design problem and motivation for joint adaptation: arXiv v1, pp. 1–2, Abstract and Introduction
  • External evidence on metaphors, fluid personas, and medium intensity: arXiv v1, pp. 1–2, Introduction and Related Work
  • Contextual and user-specific inputs: arXiv v1, p. 2, Fluid Personality Framework
  • Definition and examples of the persona module: arXiv v1, p. 2, Persona Module (Metaphor Adaptor)
  • Definition and examples of the intensity modulator: arXiv v1, p. 2, Personality Module (Trait Intensity Modulator)
  • Conceptual scope without own evaluation: arXiv v1, pp. 2–4, Framework, Conclusion and References