BehaviorSim: A Multi-Agent Framework for Social Systems Simulation via Behavioral Decision Modeling

Multi-agent simulation leveraging large language models (LLMs) offers a powerful approach for studying complex social phenomena. However, existing frameworks often lack individual heterogeneity and cognitive biases observed in real-world decision-making, and provide limited support for interpretable behavioral analysis. To address these gaps, we propose BehaviorSim, a novel multi-agent simulation framework designed to model meso-level social interactions with cognitively grounded and interpretable agents. Unlike prior LLM-based multi-agent systems (MAS) frameworks that primarily focus on task execution or coordination, BehaviorSim introduces a hybrid decision-making mechanism grounded in behavioral decision theory, combining explicit Chain-of-Thought (CoT) reasoning with a quantified Action Utility model to capture human-like trade-offs under bounded rationality. Beyond decision modeling, we present a Behavioral Logic Analysis Module that enables structured post-simulation analysis through an observation-plan-action schema and causal factor extraction guided by the behavioral decision theory. The modular architecture of BehaviorSim supports high-fidelity, interpretable multi-agent simulations, providing a conceptual foundation for advancing cognitively grounded modeling in computational social science.