German General Personas: A Survey-Derived Persona Prompt Collection for Population-Aligned LLM Studies

TL;DR

This work introduces German General Personas (GGP), a survey-derived, population-aligned collection of 5,246 persona prompts grounded in the ALLBUS German General Social Survey. By combining a fixed core demographic block with a data-driven TOP-k attribute set, GGP enables plug-and-play prompting of diverse LLMs to simulate population response distributions across 27 outcome variables. Across a range of models and data regimes, GGP-based prompts often outperform baseline distribution-prediction methods, especially in data-scarce settings, and exhibit limited sensitivity to representativity. The study highlights the utility of principled attribute selection and prompt-based population alignment for scalable, interpretable NLP-based social simulations, while noting ethical and methodological caveats. Overall, GGP provides a practical, empirically grounded resource for researching population-aligned prompting in NLP and CSS contexts.

Abstract

The use of Large Language Models (LLMs) for simulating human perspectives via persona prompting is gaining traction in computational social science. However, well-curated, empirically grounded persona collections remain scarce, limiting the accuracy and representativeness of such simulations. Here we introduce the German General Personas (GGP) collection, a comprehensive and representative persona prompt collection built from the German General Social Survey (ALLBUS). The GGP and its persona prompts are designed to be easily plugged into prompts for all types of LLMs and tasks, steering models to generate responses aligned with the underlying German population. We evaluate GGP by prompting various LLMs to simulate survey response distributions across diverse topics, demonstrating that GGP-guided LLMs outperform state-of-the-art classifiers, particularly under data scarcity. Furthermore, we analyze how the representativity and attribute selection within persona prompts affect alignment with population responses. Our findings suggest that GGP provides a potentially valuable resource for research on LLM-based social simulations that enables more systematic explorations of population-aligned persona prompting in NLP and social science research.

Figures (7)

• Figure 1: Grounding the German General Personas (GGP) in the ALLBUS survey. We construct individual persona prompts for each ALLBUS participant, varying size and composition via available attributes. A global variable importance ranking informs the selection of the $k$ most important attributes (TOP-$k$). Personas comprise a fixed block of core socio-demographics and a more extensible block of TOP-$k$ attributes that allow varying information content. The GGP is available in JSON- and full-text formats. The originally German survey items and personas are here presented in English for illustration.
• Figure 2: Change in alignment with increased number of training samples. We show the JSD between the survey response distribution and response distributions generated using persona-prompting with different LLMs as well as response distributions produced using random forest classifiers with increasing training set sizes. The alignment (averaged across $27$ outcome variables) is better when using persona-prompting, particularly for small $n$— LLMs are already well-aligned, even without training samples.
• Figure 3: Change in alignment with increased number of persona attributes. We show the JSD between the survey response distribution and response distributions generated using increasingly large sets of TOP-$k$ persona attributes for different LLMs. The largest Llama model outperforms others up until the TOP-$64$ attributes are used, showing best alignment when using only the TOP-$2$ attributes. Across models, adding persona attributes does not monotonously lead to better alignment.
• Figure 4: Alignment comparison across different topics. We compare the Jensen-Shannon Distance (JSD) between the survey response distribution and the response distribution generated using the best GGP configuration as well as the response distributions from random forest baselines with different training set sizes. Using the GGP and thus no input other than a single persona description produces response distributions that are better aligned with survey responses than the best random forest baselines across five out of nine different topics.
• Figure 5: Alignment comparison with unrepresentative persona collections. We compare the average Jensen-Shannon Distance of response distributions generated with representative (GGP), unrepresentative (baselines), and no persona collections to the survey response distribution. All synthetic responses are generated using Llama-3.3-70B-Instruct. The unrepresentative baselines and the GGP cluster closely together in terms of their JSD scores, indicating that representativity only has a small influence on the alignment.