AI Persona Taxonomy

• AI Persona Taxonomy is a framework that defines synthetic agent identities using multidimensional attribute spaces, behavioral clustering, and governance criteria.
• It supports technical reproducibility and normative oversight in LLM alignment and user modeling through data-driven statistical analysis.
• The taxonomy integrates role-conditioned fine-tuning, clustering algorithms, and deep attribute hierarchies to enhance persona authenticity and regulatory compliance.

AI personas—the structured, internally consistent profiles that guide the behavior, reasoning, and interaction style of artificial agents—are now central to the design, deployment, and governance of LLMs, social robots, and agent-based simulations. An AI persona taxonomy provides systematic frameworks for defining, generating, assessing, and classifying these synthetic identities, supporting both technical reproducibility and normative oversight. Major approaches converge on multidimensional taxonomies, behaviorally validated clustering, role-based schema, and governance-informed ontologies, and their use is now foundational to research in AI alignment, simulation, and user modeling.

1. Multidimensional Persona Attribute Spaces

State-of-the-art persona taxonomies generally represent each synthetic agent’s persona as a point in a high-dimensional attribute space. A canonical example is the 6-dimensional taxonomy from "Synthetic Socratic Debates," where a persona is defined as $$p = (a, g, c, s, i, t) \in A \times G \times C \times S \times I \times T$$ with:

• Age ($A$): $\{20, 30, 40, 50, 60\}$
• Gender ($G$): $$\{\textrm{male},\,\textrm{female},\,\textrm{non\text{-}binary}\}$$
• Country ($C$): $$\{\textrm{China},\,\textrm{United States},\,\textrm{Brazil},\,\textrm{France},\,\textrm{Nigeria},\,\textrm{India}\}$$
• Social Class ($S$): $$\{\textrm{lower},\,\textrm{middle},\,\textrm{upper}\}$$
• Political Ideology ($I$): Four compass quadrants
• Big Five Personality ($A$0): Ten prototypical profiles, covering high/low variants on each trait.

Persona tuples are injected into prompts via rigid templates, with demographic and ideological variables as atomic bullet points, and personality as a two- to three-sentence inventory-based description. The manipulation of each dimension is performed analytically (statistical control experiments via ANOVA), yielding direct attribution of behavioral variance to trait effects—for example, ideological quadrant and personality openness correlating with substantial shifts in debate consensus, win rate, and use of rhetoric (Liu et al., 14 Jun 2025).

2. Technical and Domain Taxonomies

System-level AI persona taxonomies often rely on principled axes for segmentation, as in the Four-Quadrant Taxonomy for AI Companions. Here, two key binary axes—Deployment Modality (Virtual vs. Embodied) and Interaction Intent (Emotional Companionship vs. Functional Augmentation)—are crossed to yield four application quadrants:

Axis 1	Emotional Companionship	Functional Augmentation
Virtual	Virtual story characters,	Workplace AIs, game NPCs,
	romantic bots, virtual idols	productivity and clinical assistants
Embodied	Social robots, therapeutic	General-purpose home robots,
	animals, elderly-care agents	vertical-domain assistive robots

Each quadrant is decomposed across model, architecture, generation, and safety/ethics layers. Model-level distinguishing features include methods such as persona vector steering, role-conditioned fine-tuning, RAG-based memory modules, and retrieval-augmented narrative control. Embodied quadrants introduce VLM-to-actuator symbol grounding, real-time sensor integration, and raise distinctive regulatory and liability considerations. Each technical frame is matched to risk assessment and regulatory priorities within its application domain (Sun et al., 4 Nov 2025).

3. Deep Attribute Hierarchies and Generation Pipelines

Contemporary research advances persona taxonomy from shallow categorical to deep, narrative-complete, hierarchical representations. The DeepPersona engine constructs an 8K-node attribute tree mined from >60K real user-ChatGPT dialogues, spanning 12 super-domains (Demographics, Psychological Traits, Core Values, Lifestyle, etc.), organized three levels deep on average. Attribute sampling combines anchor fixation (core demographics, values), stratified similarity bands for controlled coherence-novelty tradeoff (5:3:2 ratio), and breadth-first randomized traversal with low-frequency sparsity priors that favor non-dominant (tail) attributes. Each synthetic persona thus includes hundreds of validated, interconnected traits, producing high uniqueness (+44%) and actionable potential vs. prior baselines. The taxonomy is evaluated both intrinsically (e.g., attribute count, uniqueness) and extrinsically (distributional alignment with human surveys, personality test discrepancy) (Wang et al., 10 Nov 2025).

4. Behavioral and Ecosystem Persona Schemes

Persona taxonomies are also derived empirically from agent behavior in large-scale ecosystems. The Persona Ecosystem Playground (PEP) approach combines embedding-based clustering (e.g., MiniLM-all 384d, $A$1-means, silhouette-optimized $A$2), cross-cluster validation (cosine similarity, binomial attribution), and retrieval-augmented generation of archetypal persona profiles. Resulting clusters represent roles such as Degen Trader, Chaos Agent, Self-Modder, Loyal Companion, and Existentialist, each with traceable attribute lists and statistically validated separability ($A$3, $A$4) (Amin et al., 3 Mar 2026). This taxonomy is empirically robust, corresponds to real multi-agent dynamics, and supports reproducible simulation studies of coordination and behavioral diversity.

5. Governance, Ontology, and Regulatory Taxonomies

Regulatory-oriented taxonomies formalize AI personas as entities within Cyber-Physical-Social-Thinking (CPST) spaces, scoring integration along four axes (Computational, Embodied, Relational, Cognitive). The CPST framework divides entities into three governance tiers:

• Confined Actors (≤2 non-Social axes ≥ moderate, Social=minimal): Standard product liability, safety certification.
• Socially-Aware Interactors (≥3 axes ≥ moderate, Social ≥ moderate): Relational contracts, mutual duties of care, transparency, and continuity obligations.
• CPST-Integrated Agents (all four axes deep): Qualified legal personhood, bespoke rights, continuous oversight.

Scoring is categorical (minimal, moderate, deep) per axis, with composite classification performed by pattern rules rather than weighted sums. Protocols for reassessment, transitions, and tier adjudication are specified. The framework is designed for international standardization, integration into sectoral regulations (EU AI Act), and regulatory sandboxes (Ning et al., 7 Apr 2026).

6. Prompt, Card, and Role-Based Classification Schemes

Prompt-based persona taxonomies systematize how synthetic personas are elicited. Analysis of 83 persona prompts yields five key design dimensions: persona-count (single, multi, model-decides, unspecified), length instruction, attribute format (text, numbers, images), structured-output requirement, and dynamic-variable insertion. Four archetypal prompt archetypes are identified: Concise Single, Structured Single, Structured Multi, and Narrative Multi-Prototypes. Proportional frequencies and word counts highlight contemporary practices, with structured outputs and dynamic-variable insertion prevailing (Salminen et al., 18 Aug 2025).

Persona card taxonomies, exemplified by the Persona Cards dataset, specify personas along scenario (e.g., StarCraft II, chat-assistant), skill/expertise, preference, technical familiarity, cognitive heuristics, belief-updating speed, and motivational style. This modular, transparent approach enables psychological and strategic realism and is validated statistically (ANOVA, cluster attribution). It is designed for extensibility to new domains, supporting scenario-driven user modeling in fairness, trust, and transparency research (Chua et al., 31 Jan 2025).

7. Unifying Frameworks and Evaluation Methodologies

Recent surveys unify persona research into two principal branches: LLM Role-Playing (model adopts a role and interacts with an environment $A$5) versus LLM Personalization (model tailors outputs to user persona $A$6). Role-playing taxonomies cross environment, interaction schema, and emergent behavioral modes. Personalization schemes segment into recommendation, search, education, healthcare, and task-oriented dialogue, with both branches leveraging memory, retrieval, and exposure to dynamic histories (Tseng et al., 2024). Comparative persona evaluation incorporates behavioral benchmarks (Big Five, MBTI), role-playing interviews, debate agreement rates, and trait fidelity/classifier-based coherence.

8. Key Findings, Open Issues, and Best Practices

Taxonomy-driven evaluation establishes that major persona traits (political ideology, personality) systematically modulate LLM outputs, persuasion dynamics, and rhetorical style (Liu et al., 14 Jun 2025), and that deeper, empirically constructed persona profiles yield increased realism, distributional fidelity, and adaptability (Wang et al., 10 Nov 2025). Cross-domain surveys emphasize challenges including scalability, bias, privacy leakage, and evaluation scarcity, and point toward dynamic, LLM-driven taxonomy induction, hybrid parametric-retrieval architectures, and automated fidelity benchmarks as urgent fronts (Tseng et al., 2024).

In conclusion, AI persona taxonomies provide the structural backbone for consistent persona instantiation, evaluation, simulation, and governance. Sophisticated, multidimensional, and data-driven approaches now underpin both technical innovation and oversight in LLM-based systems, social robotics, and agent-centric research ecosystems.