SozioMimic
- Partner
- Ministry of Economic Affairs, Labour and Tourism
- Period
- Feb 2025 – Oct 2025
- Focus
- Digital Human Simulation
Key outcomeResearch platform for digital human simulation
Context
Traditional market and user research does not scale well: interviews and surveys are expensive to repeat, and human availability limits how quickly a study can adapt. SozioMimic asks a more fundamental question: how much of a person can an AI agent represent using their own qualitative data, rather than a generic persona?
Inputs
- In-depth interviews
- Diary studies and longitudinal surveys
- Demographic and contextual metadata
Approach
- Designed LLM-based digital human representations using structured prompting and retrieval mechanisms
- Built the twin-generation pipeline: preprocessing, context extraction, and a retrievable participant context store
- Integrated multi-agent reasoning to simulate individual perspectives
- Evaluated alignment, consistency, and temporal variation in simulated responses against real survey answers
Outcomes
- Research platform for digital human simulation
- Adding more personal context does not automatically produce a better representation; how that context is organized and retrieved determines whether an agent represents an individual or produces a convincing stereotype
- Published research: Simulation of Human Survey Responses for Market Research (CHI EA '26)
- Published research: Virtual User Lab: An Interactive System for Simulating User Perspectives with AI Personas for Ideation and Evaluation (UMAP 2026)