LLMs don’t represent “everyone”—they represent a default human. Most models are tuned to a WEIRD (Western, Educated, Industrialized, Rich, Democratic) norm, making them fragile or misleading when used in global or diverse settings. In this workshop, participants will explore Synthetic Cultural Agents (SCAs): AI agents designed to simulate the values, risk preferences, and decision-making styles of real human populations.

Learn how to spot hidden bias, stress-test ideas, and build LLM-powered solutions that actually scale beyond a single assumed user. Perfect for hackathon teams building for real-world impact.

Challenge

Teams will put theory into action by comparing five powerful techniques for narrowing an
LLM’s probability space into a specific cultural manifold:

• Demographic Steering — Lock priors using structured JSON configs
• Recursive Memory & Reflection — Build agents with stable self-schemas
• Social RAG — Ground agents in localized, real-time epistemic bases
• Polyglot Persona — Treat language as a hidden cultural parameter
• Latent Space Fine-Tuning — Hard-code alignment via weights and reward models

Facilitators

Kseniia Biriukova and Monica Capra from the EconLLM Lab

Key Takeaways

• A Social Science Lens on AI: See how technical design choices reshape model behavior, fidelity, and decision-making
• SCA Methods, Compared: Understand the real trade-offs between scalability, precision, and realism
• Research-Driven Innovation: Engage with foundational work (e.g., Argyle et al.) and cutting-edge advances in Polyglot Persona Drift and Synthetic Ethnography