NetworkGames: Simulating Cooperation in Network Games with Personality-driven LLM Agents

Researcher Xuan Qiu has introduced NetworkGames, a novel framework that bridges Generative Agents and Geometric Deep Learning to simulate complex social dynamics. The system formalizes social simulation as a message-passing process governed by Large Language Model (LLM) policies, creating populations of AI agents each endowed with distinct MBTI personality types. Through extensive simulations of the Iterated Prisoner's Dilemma across various network structures, the research establishes baseline interaction matrices for all 16 personality pairs, revealing nuanced cooperative preferences that challenge simple predictions.

The study's key finding demonstrates that macro-level cooperation cannot be predicted from dyadic interactions alone. Network connectivity and spatial personality distribution co-determine outcomes, with small-world networks proving detrimental to cooperation while strategically placing pro-social personalities in hub positions within scale-free networks significantly boosts collective welfare. The researchers validated these findings through stress tests across multiple LLM architectures, scaled network sizes, varying random seeds, and comprehensive ablation studies, ensuring robustness of the observed patterns.

NetworkGames represents a significant advancement in social physics research, offering a systematic approach to understanding how AI societies might evolve. The framework's open-source nature allows other researchers to build upon these findings, potentially leading to better-designed online social environments and more accurate forecasting of collective human behavior in networked settings. The work bridges game theory, network science, and AI research in unprecedented ways.