BiD: Behavioral Agents in Dynamic Auctions

Complex societal systems are characterized by heterogeneous agents engaging in strategic interactions, yet current representative agent-based models (ABMs) struggle to capture these dynamics. We present BiD (Behavioral Agents in Dynamic Auctions), a novel ABM framework that focuses on modeling complex systems: heterogeneous agent modeling and socioeconomic dynamics. Using Dutch auctions as a microcosm, BiD models agent heterogeneity in risk preferences and dynamic trust scores, while modeling socioeconomic interactions via strategic communication phases before bidding. Our theoretical equilibria analysis reveals how BiD enables phenomena observed in real markets that are unexplained by classical ABMs, \textit{i.e.}, successful low-valuation bidders win through strategic communication. We formalize behavioral agent strategies under different communication protocols and develop a reinforcement learning-guided policy for LLM-based agents to adapt their behaviors based on market dynamics. Experimental results demonstrate BiD's capability in first modeling realistic market dynamics, providing socioeconomic perspectives for studying multi-agent systems and complex societal systems.