Technical Summary: From Survey Personas to LLM Agents

Problem Statement

The European Union's commitment to phasing out new internal combustion engine (ICE) vehicles represents a significant policy shift, yet public responses to such changes are dynamic, shaped by interaction, feedback, and evolving socio-political contexts. Traditional survey methods are effective for measuring attitudes at specific points in time but are ill-suited for modeling how these attitudes evolve under exogenous shocks and unfolding events. While recent literature explores using Large Language Models (LLMs) as generative agents to simulate human behavior, existing studies predominantly rely on hand-crafted, stylized personas. This approach limits the empirical grounding of simulations, as persona design heavily influences how agents interpret context and make decisions. Consequently, two critical gaps remain: (1) the lack of methods to construct LLM agents from empirically grounded respondent populations rather than manual personas, and (2) the absence of frameworks to study dynamic policy preference formation using these grounded agents under changing social contexts.

Methodology

The authors propose a Survey-Grounded Generative Agent-Based Modeling (GABM) simulation framework. This approach transforms real survey respondents into generative LLM agents, bridging the gap between static empirical data and dynamic behavioral simulation.

Data and Agent Construction

The framework utilizes data from an online survey conducted in Germany (Summer 2025) with 514 respondents, designed based on the Integrative Public-Policy-Acceptance (IPAC) framework.

• Persona Encoding: Each respondent is converted into an LLM agent via a natural-language persona. These personas encode 57 attributes, including:
  • Demographics (age, gender, education, income, federal state).
  • Political orientation and party preference.
  • Mobility behavior and fuel experience.
  • Climate-related attitudes (concern, risk perception, trust in government).
• Initialization: Agents are initialized with their observed survey response regarding support for the ICE phase-out, serving as the baseline for subsequent updates.

Simulation Framework

The simulation operates in a dynamic world where agents update preferences over time through repeated rounds (16 rounds, representing quarters from Q1 2025).

• Contextual Inputs: In each round, agents receive a structured prompt containing their persona, prior response history, current national and state-level support distributions, and, in event-driven scenarios, information about external events.
• Decision Logic: Agents reason over these inputs to update their policy stance, answering the question: "In an election, to what degree would you vote for a party that supports such policies?" (Scale 1–6).
• Feedback Loop: Aggregated agent responses update the collective support distributions, which serve as the social context for the next round.
• Experimental Conditions: The study compares a baseline scenario (no exogenous shocks) against an event-driven scenario (introducing ten context-relevant events with supportive or opposing effects).
• Model Configuration: The study uses OpenAI's GPT-4o mini with a temperature of 0 to minimize stochastic variation. Experiments vary prompt language (German vs. English) and summary style (environmental vs. neutral).

Key Contributions

1. Framework Development: Introduction of a survey-grounded GABM framework that translates structured survey data into natural-language LLM agents, moving beyond hand-crafted personas.
2. Empirical Application: Application of this framework to a timely policy issue: public support for phasing out ICE vehicles in Germany, aligned with EU net-zero targets.
3. Dynamic Analysis: Provision of empirical analysis on how survey-grounded agents respond to changing social contexts and external events, enabling the examination of temporal support trajectories, preference switching, and profile heterogeneity.

Results

The simulation results highlight significant dynamics in preference formation:

• Baseline Dynamics: In the absence of external shocks, agents generally show an upward shift in support, moving from mid-level positions to the highest support level (6). This suggests the simulated environment creates an inherent upward drift rather than merely preserving initial distributions.
• Prompt Sensitivity: The framework is highly sensitive to prompt design:
  • German Prompts: Generate more decisive upward movement and convergence toward the highest support level, particularly under neutral summaries.
  • English Prompts: Result in more diffuse distributions, with responses spreading across intermediate levels (4–6) rather than collapsing into the highest support. This indicates weaker anchoring and less extreme consolidation.
• Event-Driven Dynamics: The introduction of external events amplifies the effects of prompt language and style.
  • German prompts with environmental summaries continue to drive significant shifts toward high support.
  • English and neutral prompts distribute mass into intermediate categories, creating greater heterogeneity in how agents update preferences.
  • Volatility in support levels is more visible in event-based scenarios, with English prompts showing more pronounced fluctuations.
  • Notably, the final support level in the English prompt-based scenario is slightly higher than in the German prompt-based scenario, despite the German prompts driving more decisive convergence to the top level.
• Individual Reasoning: Case studies reveal that agents adjust reasoning based on a complex interplay of personal values, trust in government, and perceived fairness. For instance, an agent might maintain high support based on climate values but drop support when events highlight distributive justice concerns for low-income households.

Significance and Limitations

The paper claims significance in advancing computational social science by connecting political problems with social acceptance analysis in a dynamic environment. It demonstrates that careful persona design enables realistic simulation of decision-making, facilitating behavioral experiments that static surveys cannot capture.

However, the authors maintain a modest stance regarding the implications:

• Data Dependency: The simulation quality is bound by the underlying survey data; measurement errors or omissions carry over into agent personas.
• Simplification: The framework simplifies complex human decision-making into structured prompts and limited response scales.
• Prompt Sensitivity: Results vary substantially based on prompt language, style, and sequence, requiring careful interpretation of simulation outcomes.
• Event Design: Events are manually designed and may not fully capture the unpredictability of real-world political developments.
• Interaction Limits: The current setup models preference updating but does not include richer social interactions like direct peer-to-peer communication or network structures.
• Model Specificity: Findings are specific to GPT-4o mini with zero temperature; different models or stochastic settings could yield different results.

The authors conclude that these outputs should be interpreted as simulations of preference dynamics rather than accurate reflections of real citizens' future political behavior, emphasizing the value of analyzing aggregated behavioral patterns over individual explanations.