EASE framework standardizes LLM social simulations for reproducibility

LLMs are increasingly used to simulate social interactions, but most existing simulators are ad hoc and monolithic, lacking architectural standardization. This makes reproducibility nearly impossible and complicates downstream evaluation. A team of researchers from academia (including Sneheel Sarangi, Maximilian Puelma Touzel, and others) has proposed EASE—a modular framework that breaks down LLM-based multi-agent simulations into four core components: Environments, Agents, Simulation engines, and Evaluation metrics.

To demonstrate EASE in practice, the team developed SiliSocS (Silicon Society Sandbox), an open-source, research-ready platform that implements the EASE configuration. SiliSocS enables researchers to define explicit research questions, configure simulation parameters, and run highly reproducible experiments. The paper, under review at NeurIPS 2026, presents three case studies using SiliSocS: a comprehensive assessment of existing questions, a deeper dive into complex social dynamics, and an elaboration of prior studies.

These case studies reveal how subtle design choices—such as agent personality prompts, environment rules, or evaluation metrics—can dramatically alter outcomes. The framework isolates these impacts, helping researchers identify limitations in current modeling approaches. By modularizing the simulation pipeline, EASE and SiliSocS pave the way for a more rigorous, scientific approach to studying LLM behavior in social contexts. This is a critical step for fields like computational social science, where reproducibility is paramount.