The $\textit{Silicon Society}$ Cookbook: Design Space of LLM-based Social Simulations
Choice of base LLM is the most critical factor shaping AI society outcomes.
A new paper from researchers at McGill University, Mila, Université de Montréal, and Ubisoft La Forge offers a comprehensive 'Cookbook' for designing LLM-based social simulations. The team systematically analyzed how different design choices—such as the base language model powering individual agents and the network topology connecting them—influence the outcomes of simulated 'Silicon Societies'. Using agent surveys as a proxy for opinions, they discovered that the choice of base LLM is by far the most impactful variable, dominating other parameters. The study also reveals that the design space is highly non-trivial: some parameters combine additively, while others exhibit complex interactions that can drastically alter simulation behavior.
This work arrives as LLM-only social networks begin appearing outside controlled settings, yet the field lacks systematic validation of model realism. The 'Silicon Society Cookbook' (under review at COLM 2026) provides a much-needed framework for researchers to make informed design decisions. By documenting how different configurations affect agent behavior, the paper helps future studies build more credible and reproducible social simulations. With 20 pages and 12 detailed tables, it serves as a practical reference for anyone building or evaluating AI-powered social network models, from academic labs to game studios like Ubisoft.
- Choice of base LLM is the single most important variable affecting simulation outcomes.
- Parameter interactions are non-trivial: some combine additively, others create complex effects.
- Study uses agent surveys as a proxy for opinions to validate model realism and design decisions.
Why It Matters
This framework enables more realistic and validated LLM social simulations for research and industry applications.