DeepPersona: Synthetic Persona Engine

• DeepPersona is a taxonomy-guided generative engine that synthesizes narrative-complete synthetic personas with hundreds of structured attributes.
• It constructs the largest human-attribute taxonomy by mining extensive conversational datasets and using progressive sampling to balance coherence and novelty.
• The system achieves high fidelity in agent-based simulations and LLM personalization, improving attribute uniqueness by up to 44% over prior benchmarks.

DeepPersona is a two-stage, taxonomy-guided generative engine designed to synthesize narrative-complete synthetic personas, providing unprecedented attribute depth and diversity for agentic behavioral simulation, LLM personalization, and human-AI alignment research. DeepPersona systematically constructs a large-scale human-attribute taxonomy by mining extensive conversational corpora, then employs progressive attribute sampling and conditional LLM generation to produce coherent, privacy-preserving persona profiles averaging hundreds of structured attributes and MB-scale free-text narratives. This approach yields synthetic citizens whose distributions and behavioral fidelity approximate authentic human populations across diverse measures (Wang et al., 10 Nov 2025).

1. Taxonomy Construction and Attribute Extraction

DeepPersona’s pipeline commences with the construction of the largest human-attribute taxonomy to date. The process employs three conversational datasets: 1,000 dialogues from the Puffin dataset, 1,000 from prefeval_implicit_persona, and 60,000 user–GPT-4.1-mini interactions from Llama-3.2-3B-HiCUPID, totaling 62,224 "personalizable" QA turns. Each QA pair is automatically labeled by GPT-4.1-mini as Non-personalizable, Partially Personalizable, or Personalizable, selecting high-quality self-disclosure content.

Attribute extraction leverages manual seeding of 12 root node categories (e.g., Demographics, Core Values, Hobbies & Interests). Personalized QA pairs are processed by GPT-4.1-mini to extract attribute “paths” up to three hierarchical levels deep (e.g., Lifestyle → Food Preference → Vegan). These paths are merged using semantic-similarity clustering (70% threshold) and sequentially filtered (pre- and post-merge) to eliminate redundancy and over-specificity. The resulting taxonomy comprises 8,496 unique nodes organized hierarchically within 12 top-level domains, such as Demographics & Identity, Physical & Health, Psychological & Cognitive, and Media Consumption & Engagement.

2. Progressive Generative Persona Synthesis

The persona-generation stage targets sampling a profile

$P=\{\langle a_i, v_i \rangle\}_{i=1}^k$

of depth $k$, optionally conditioned on user-supplied seed $S$. The sampling process is defined by: $$P \sim \mathcal{F}_{\theta,T}(\cdot\mid S, k) = \prod_{i=1}^{k} \Pr(a_i\mid S, P_{<i}, T) \;\; \Pr_{\theta}(v_i \mid a_i, S, P_{<i})$$ where $T$ is the taxonomy and $\theta$ denotes LLM parameters.

• Stable core anchors: Initial attributes (e.g., age, location, career, life attitude, personal story, hobbies) are fixed using either seed $S$ or bias-free sampling tables to prevent cultural-majority defaulting.
• Bias-free value assignment: Categorical values (gender, country) are sampled from curated tables; abstract concepts (core values, life attitude) are elicited with targeted LLM prompts for cross-profile consistency.
• Balanced diversification: The full attribute set is embedded in vector space and partitioned by cosine similarity to the core set ("near," "middle," "far" thirds). Attributes are sampled in a 5:3:2 ratio among these strata, balancing coherence and novelty.
• Progressive LLM filling: A stochastic breadth-first traversal of $T$ selects unexplored child nodes, favoring long-tail branches, iterating until $k$ attributes are chosen. For each attribute $a_i$, value $k$0 is LLM-generated conditioned on $k$1. A dedicated pass then produces a free-text narrative summary, $k$2.

A high-level pseudocode of this algorithm is as follows: $k$8

3. Model Architecture and Persona Output Properties

DeepPersona adopts a model-agnostic approach but employs GPT-4.1-mini for both attribute extraction and value generation. No parameter-efficient fine-tuning is conducted; all control is implemented through engineered prompting and taxonomy-guided sampling.

Persona generation factorizes attribute selection and value generation: $k$3 Resulting personas average $k$4 structured attributes per profile. Explicit attributes extracted by GPT-4o judges average 50.92 per persona (compared to 3.98 for PersonaHub and 38.50 for OpenCharacter). Narrative summaries typically occupy approximately 1 MB of text, representing two orders of magnitude greater detail than prior synthetic persona benchmarks.

4. Evaluation Metrics and Comparative Performance

Evaluation encompasses both intrinsic and extrinsic dimensions, with GPT-4o employed as a judge.

Intrinsic metrics (Table 1, below) demonstrate significant improvements:

Metric	PersonaHub	OpenCharacter	DeepPersona
Mean # Attributes	3.98	38.50	50.92
Uniqueness	2.50	2.86	4.12
Actionability	3.60	4.78	5.00

Relative to OpenCharacter, DeepPersona achieves:

• +32% mean attribute count
• +44% uniqueness
• +5% actionability

Extrinsic evaluation includes:

• LLM Personalization: Conditioning GPT-4.1-mini on DeepPersona profiles increases response accuracy by an average of +11.6% across ten metrics. Attribute coverage and justification both improve by approximately 10–12%.
• Population Simulation: Synthetic national populations (n=100 per country) responding to World Values Survey questions show a reduction of 32–43% in Jensen-Shannon divergence and Wasserstein distance, and a 31.7% narrowing of the human–LLM gap compared to Cultural Prompting; a further ≈7% gain versus OpenCharacter.
• Big Five Personality Test: On IPIP items, DeepPersona-simulated citizens close the deviation from real data by 17% relative to LLM-only simulated citizens, and a KS improvement of 0.215 versus OpenCharacter.

5. Data Handling, Privacy, and Scalability

DeepPersona's datasets exclusively comprise anonymized user–ChatGPT dialogues, with no storage or release of real PII. All generated personas are synthetic and privacy-preserving, supporting public distribution of code, taxonomies, and synthetic profiles.

The decoupling of attribute sampling and LLM narrative generation ensures high scalability. The computational cost is linear in $k$5, dominated by LLM calls, with no additional large parameter sets beyond $k$6. This framework can deepen millions of pre-existing persona sketches into attribute-rich, uniquely structured profiles.

6. Applications, Limitations, and Research Trajectory

Key domains for DeepPersona include agentic behavioral simulation (cultural analysis, policy impact), personalized AI assistants and recommendation systems, social survey and population modeling, as well as stress-testing LLM alignment and fairness.

Limitations center on diminishing returns and attribute noise for profiles exceeding $k$7, potential for LLM hallucination on long-tail attributes, and continued labor intensity of manual filtering stages.

Notable directions for future research include dynamic taxonomy refinement from ongoing interactions, integration of temporal and multimodal attributes, parameter-efficient fine-tuning to further reduce hallucinations, and systematic controls over cost versus output quality for large deployments.

Conclusion

DeepPersona represents a taxonomy-guided generative paradigm that systematically exposes and leverages the breadth of human-attribute space to construct synthetic personas of exceptional depth and utility for both intrinsic and applied AI research. By combining rigorous attribute taxonomy mining, balanced progressive sampling, and narrative LLM synthesis without sacrificing privacy, it establishes a new baseline for high-fidelity population simulation and personalized AI development (Wang et al., 10 Nov 2025).