AgoraSim: A Hybrid Framework for LLM-Agent Social Simulations.

This paper introduces AgoraSim, a novel hybrid agent-based modeling framework specifically designed for simulating social scenarios. While Large Language Model (LLM) agents simplify the instantiation of natural-language social dynamics, their outputs can be misinterpreted as predictions and are often difficult to compare with established social models. AgoraSim addresses these challenges by enabling the conversion of textual or multimodal inputs into editable agent-based model configurations. The framework supports diverse agent populations, allowing for a mix of LLM, vision-language, custom, random, and classical agents, all operating within ratio-controlled environments. A key feature is its ability to compare simulated scenarios against matched classical reference dynamics, ensuring a robust evaluation. All agents within AgoraSim emit a standardized structured decision object, facilitating common action spaces, interaction protocols, metrics, and audit trails. AgoraSim is accessible via a local UI, Python SDK/CLI, and a REST API, providing users with comprehensive tools to inspect scenario trajectories, compare different modeling assumptions, and identify situations requiring empirical validation. This framework aims to bridge the gap between flexible LLM-agent simulations and rigorous social dynamics analysis.