Synthetic Personalities: How Well Can LLMs Mimic Individual Respondents Using Socio-Economic Microdata?

Market research has traditionally relied on large-scale surveys, which are costly and time-consuming. Recent developments in large language models (LLMs) like GPT-4 have demonstrated their capacity to simulate human responses, but mainly on purpose-collected datasets. Meanwhile, enterprises accumulate vast heterogeneous panel data via CRM, loyalty programs, and repeated surveys, which are underutilized for digital twin construction. Prior academic work has shown the potential of detailed individual twins, but often on curated datasets, limiting real-world applicability. This paper bridges that gap by exploring how existing panel data can be used to generate high-quality individual twins, addressing scalability and operational relevance.

The core challenge is to leverage existing heterogeneous panel data—characterized by uneven item coverage and varying response depths—to construct accurate individual digital twins. Traditional methods require dedicated data collection, which is costly and inflexible. Additionally, the impact of different construction choices—such as information depth, embedding strategies, and reasoning modes—has not been systematically studied. The question remains: how to optimize these parameters to maximize twin fidelity while minimizing costs, in a real-world enterprise setting?

The paper introduces a systematic evaluation framework combining Shannon entropy-based information regulation, dialog history embedding, and explicit reasoning modes. It demonstrates that a balanced information depth (around 75%) suffices for high accuracy, reducing costs significantly. The use of open-weight models ensures deployment feasibility. Unlike prior work limited to curated datasets, this approach leverages existing heterogeneous panel data, making digital twins more accessible for industry. The integration of multiple construction parameters into a unified evaluation grid provides actionable insights for practitioners.

• �� Data collection: Use SOEP 2023 wave, selecting 500 representative participants with rich response histories.
• �� Model setup: Employ three open-weight LLMs (Qwen 3, Gemma 4, LLaMA-3), each fine-tuned for response generation.
• �� Information regulation: Rank items by Shannon entropy, select subsets at 0%, 25%, 50%, 75%, and 100% levels.
• �� Embedding strategies: Use narrative persona summaries and raw dialog histories.
• �� Reasoning modes: Prompt models with explicit reasoning instructions or direct response generation.
• �� Response generation: Generate responses to 183 held-out questions, evaluate with accuracy, rank correlation, and Fisher-z.
• �� Performance analysis: Compare across parameter combinations, identify Pareto optimal points, and validate robustness.

The experimental setup involves generating responses for 500 participants across 60 configurations, measuring key metrics like accuracy (max 78.8%), rank correlation (max 0.590), and response variance. The models are evaluated on a large, real-world dataset, with systematic variation of information depth, embedding, and reasoning. Results show diminishing returns beyond 75% entropy, with the best performance at full information depth. Embedding choice consistently influences accuracy, while reasoning mode affects response ranking. The experiments confirm that optimized parameter combinations yield high-fidelity twins comparable to curated datasets, with practical cost-benefit trade-offs.

The study finds that increasing information depth improves twin fidelity, but with diminishing returns after 75% entropy. The best accuracy (78.8%) and correlation (0.590) occur at full depth with dialog history embedding and explicit reasoning. Switching embeddings enhances accuracy across models, while reasoning boosts rank-order fidelity. Qwen 3 outperforms others, validating the approach’s robustness. These results demonstrate that high-quality twins can be built from existing panel data, with cost-effective parameter tuning.

The methodology enables firms to leverage their existing heterogeneous data for scalable consumer response modeling, supporting targeted marketing, segmentation, and personalized service. It reduces reliance on costly bespoke surveys, offering a practical, deployable solution for real-time decision-making. Long-term, integrating multi-modal data and advanced reasoning could further enhance twin fidelity, transforming market research into a continuous, AI-driven process that adapts to consumer dynamics in real-time.

Model performance declines with minimal data or highly biased responses, risking inaccuracies. Embedding strategies require further optimization for robustness across diverse datasets. High computational costs limit real-time deployment, especially for large models and high information depths. Future work should focus on improving efficiency, bias mitigation, and multi-modal integration to broaden practical applicability.