Event-Driven Belief Updating with LLM Agents

Forschungsthema/Bereich

Information Systems/Computational Social Science

Typ der Abschlussarbeit

Master

Startzeitpunkt

22.04.2026

Bewerbungsschluss

15.05.2026

Dauer der Arbeit

6 months

Beschreibung

Public opinion does not stand still. When major news events occur such as elections, economic shocks, policy changes and crises, survey panels capture how people's views shift over time. But running repeated panel waves is costly and slow. This thesis investigates whether LLM based persona agents conditioned on real panel data can simulate such belief updates after sequential exposure to news events and whether those simulated shifts match patterns observed in actual panel follow-up waves.

The core idea builds on the Silicon Sampling paradigm (Argyle et al., 2023), in which LLMs are conditioned on sociodemographic backstories to reproduce survey response distributions. This thesis extends that approach into a dynamic, event driven framework: each agent is initialised with a participant's wave 1 survey responses and demographic profile, then exposed sequentially to real news events drawn from an event database. After each event batch, the agent is re-surveyed, and the resulting belief trajectory is compared against the actual panel follow-up wave.

Key open questions include: How much wave 1 context does an agent need to stay consistent across waves? Which event types such as crises, elections and scandals produce the strongest alignment with real panel shifts? Does simulation accuracy vary systematically by persona characteristics such as age, education, or political orientation? And how can framing effects and sycophancy be disentangled from genuine belief updating?

The thesis will develop and evaluate a persona agent pipeline using real panel data and an event database, comparing simulated belief trajectories against actual panel follow-up waves across multiple topics and subgroups.

Literature:

Argyle et al. (2023). Out of One, Many: Using Language Models to Simulate Human Samples. Political Analysis, 31(3), 337 351.

Bisbee et al. (2024). Synthetic Replacements for Human Survey Data? Political Analysis, 32(4).

Santurkar et al. (2023). Whose Opinions Do Language Models Reflect? ICML 2023, 29971 30004.

Voraussetzung

Voraussetzungen an Studierende

• Students should have:
• Good working knowledge of large language models and NLP including prompt engineering, fine tuning and agent frameworks.
• Basic understanding of survey methodology and panel study design, or willingness to acquire it.
• Familiarity with Python and ability to work with APIs such as the OpenAI API.
• Interest in computational social science and the use of AI for studying opinion dynamics.
• Ability to work independently with real research data.

Studiengangsbereiche

• Ingenieurwissenschaften
  Informatik
  Information System Engineering and Management
• Naturwissenschaften und Technik
  Wirtschaftsmathematik
• Sonstiges z.B. Lehramt
  Informatik
• Wirtschafts- und Rechtswissenschaften
  Wirtschaftsinformatik
  Wirtschaftsingenieurwesen

Betreuung

Titel, Vorname, Name

M.Sc. Amal Labbouz

Organisationseinheit

Institut für Wirtschaftsinformatik

E-Mail Adresse

amal.labbouz@kit.edu

Link zur eigenen Homepage/Personenseite

Website

Bewerbung per E-Mail

Bewerbungsunterlagen

• Anschreiben
• Lebenslauf
• Notenauszug

E-Mail Adresse für die Bewerbung
Senden Sie die oben genannten Bewerbungsunterlagen bitte per Mail an amal.labbouz@kit.edu


Zurück