LLM-Based Social Simulations: A Systematic Design Space Analysis
A new study systematically analyzes the design space of LLM-based social simulations, termed 'Silicon Societies.' The research examines how key design choices—such as the base LLM model and agent connectivity—affect simulation outcomes. Using surveys as a proxy for agent opinions, the findings reveal a non-trivial geometry in the design space, with some parameters acting additively and others displaying complex interactions. The choice of base LLM emerges as the most critical variable impacting simulation realism. The study aims to enable more informed design decisions and validate model realism in this growing field.
Key facts
- The study is titled 'The Silicon Society Cookbook: Design Space of LLM-based Social Simulations.'
- It focuses on simulations of human behavior using LLMs, called Silicon Societies.
- The design space of these networks is under-studied, creating a gap in validating model realism.
- Key design choices analyzed include the base LLM model and agent connectivity.
- Surveys were used as a proxy for agent opinions.
- The design space geometry is non-trivial, with additive and interactive parameter behaviors.
- The base LLM is the most important variable impacting simulation outcomes.
- The study aims to help future works make more informed design decisions.
Entities
Institutions
- arXiv