Computational Methods for Multilingual Synthetic Audience Simulation Using Multi-Agent LLM Architectures

In the SynthAudience project, Prof. Rottmann is researching systematic evaluation methods for multi-agent systems designed to generate synthetic responses, as well as their cross-linguistic generalization.

The current state of research has not yet provided sufficiently robust foundations, architectures, and measurement methods for scientifically validated, cross-lingual responses from LLM-based agents in structured interviews. To address this, the project compares various approaches to persona grounding, agent memory, and the orchestration of multiple agents; develops standardized metrics; and establishes a multilingual benchmark that includes controlled human evaluation.

The project establishes a robust foundation for scientific evaluations in computer science, promotes reproducibility, open-source approaches, and the potential for transfer across languages and cultures.