DeepPersona

Simulating human profiles by instilling personas into large language models (LLMs) is rapidly transforming research in personalization, social simulation, and human-AI alignment. However, most existing synthetic personas remain shallow and simplistic, capturing minimal attributes and failing to reflect the rich complexity and diversity of real human identities.

We introduce DeepPersona, a scalable generative engine for synthesizing narrative-complete synthetic personas through a two-stage, taxonomy-guided method. First, we algorithmically construct the largest-ever human-attribute taxonomy, comprising over hundreds of hierarchically-organized attributes, by systematically mining thousands of real user-ChatGPT conversations. Second, we progressively sample attributes from this taxonomy, conditionally generating coherent and realistic personas, averaging hundreds of structured attributes and roughly 1 MB of narrative text, two orders of magnitude deeper than prior works.

Intrinsic evaluations confirm significant improvements in attribute diversity (32% higher coverage) and profile uniqueness (44% greater) compared to state-of-the-art baselines. Extrinsically, our personas enhance GPT-4.1-mini's personalized Q&A accuracy by 11.6% average on ten metrics, and substantially narrow (by 32%) the gap between simulated LLM "citizens" and authentic human responses in social surveys.

DeepPersona thus provides a rigorous, scalable, and privacy-free platform for high-fidelity human simulation and personalized AI research.

Stage 1: Human-Attribute Taxonomy Construction. We systematically mine 3,000 real-world user-ChatGPT dialogues from the Puffin dataset, identifying 1,224 high-quality personalized Q-A pairs. Using GPT-4o, we extract and hierarchically organize fine-grained attributes (e.g., Lifestyle → Food Preference → Vegan), then merge semantically similar branches. The resulting taxonomy contains 4,676 unique nodes organized under 12 broad categories (Demographics, Health, Core Values, etc.).

Stage 2: Progressive Attribute Sampling. With the comprehensive taxonomy \(T\) in place, persona generation reduces to iteratively sampling from the structured distribution \(\mathbb{P}_{\theta,T}(P \mid S,k) = \prod_{i=1}^{k} \Pr(a_i \mid S, P_{\lt i}, T)\, \Pr_{\theta}(v_i \mid a_i, S, P_{\lt i}, T)\), where \(P = \{(a_i, v_i)\}_{i=1}^{k}\) and \(P_{\lt i} = \{(a_j, v_j)\}_{j \lt i}\). The attribute selector chooses the next node \(a_i\) from \(T\) and the LLM \(\theta\) generates its value \(v_i\). To achieve realistic depth and diversity, we first anchor a stable core of non-negotiable attributes (age, location, career, personal values, life attitude, hobbies), then use bias-free value assignment from predefined tables for demographics to avoid majority-culture defaults. The selector performs stochastic breadth-first traversal of \(T\), biasing toward long-tail branches to maximize diversity, while the LLM progressively fills each selected attribute conditioned on the evolving profile \(P_{\lt i}\) to ensure global coherence. A narrative summary \(\operatorname{Narr}(P)\) is produced as a byproduct of this sampling process.

We benchmark DeepPersona on three complementary axes to verify that profiles are deep, distinct, and useful: (i) intrinsic quality, (ii) LLM personalization, and (iii) social simulation. Together, these evaluations assess whether DeepPersona truly advances synthetic users from verbose text to research-ready human proxies.

Result Figure 1. Comparison of DeepPersona with Cultural Prompting and OpenCharacter across six countries.Lower values on four statistical metrics (KS, Wasserstein, JS Divergence, Mean Diff.) indicate closer alignment with real human survey responses.DeepPersona consistently achieves the lowest distances (highlighted in green), showing superior realism and cultural fidelity.

Result Figure 2. Comparison across countries shows that DeepPersona consistently attains the lowest KS, Wasserstein, JS Divergence, and Mean Difference values, indicating superior alignment with real human personality distributions and enhanced psychological fidelity.

Result Figure 3. Intrinsic comparisons show that DeepPersona substantially outperforms prior baselines (PersonaHub and OpenCharacter) across all metrics. It achieves the highest mean number of attributes (50.92), indicating far greater persona depth, along with markedly higher uniqueness (4.12) and actionability potential (5.00), reflecting richer diversity and stronger utility for downstream personalization tasks.

Result Figure 4. Comparison of persona modeling methods on World Values Survey responses from Germany. Lower scores across KS, Wasserstein, JS Divergence, and Mean Difference indicate better alignment with real human distributions. DeepPersona consistently achieves the lowest values across all backbone models (DeepSeek-v3, GPT-4o-mini, GPT-4.1, Gemini-2.5-Flash), confirming its robustness, generalizability, and superior fidelity in simulating authentic social behaviors.

Result Figure 5.Radar plots compare DeepPersona with PersonaHub and OpenCharacter across ten personalization metrics, including Personalization Fit, Attribute Coverage, Depth, Justification, and Engagement. Evaluations were conducted under multiple responder–evaluator settings (GPT-4.1-mini, GPT-4.1, GPT-4o, and Gemini-2.5-Flash). DeepPersona consistently achieves the highest overall scores, demonstrating stronger personalization fit, richer contextual grounding, and higher response diversity and actionability across all model configurations.

If you have any questions regarding DeepPersona, feel free to reach out to us via email at zhouyufan365@gmail.com or directly submit a GitHub issue.