Lu Yikang, Aleta Alberto, Du Chunpeng, Shi Lei, Moreno Yamir
School of Statistics and Mathematics, Yunnan University of Finance and Economics, Kunming, 650221, China.
Institute for Biocomputation and Physics of Complex Systems, University of Zaragoza, Zaragoza, 50018, Spain; Department of Theoretical Physics, University of Zaragoza, Zaragoza, 50009, Spain.
Phys Life Rev. 2024 Dec;51:283-293. doi: 10.1016/j.plrev.2024.10.013. Epub 2024 Oct 28.
The advent of Large Language Models (LLMs) offers to transform research across natural and social sciences, offering new paradigms for understanding complex systems. In particular, Generative Agent-Based Models (GABMs), which integrate LLMs to simulate human behavior, have attracted increasing public attention due to their potential to model complex interactions in a wide range of artificial environments. This paper briefly reviews the disruptive role LLMs are playing in fields such as network science, evolutionary game theory, social dynamics, and epidemic modeling. We assess recent advancements, including the use of LLMs for predicting social behavior, enhancing cooperation in game theory, and modeling disease propagation. The findings demonstrate that LLMs can reproduce human-like behaviors, such as fairness, cooperation, and social norm adherence, while also introducing unique advantages such as cost efficiency, scalability, and ethical simplification. However, the results reveal inconsistencies in their behavior tied to prompt sensitivity, hallucinations and even the model characteristics, pointing to challenges in controlling these AI-driven agents. Despite their potential, the effective integration of LLMs into decision-making processes -whether in government, societal, or individual contexts- requires addressing biases, prompt design challenges, and understanding the dynamics of human-machine interactions. Future research must refine these models, standardize methodologies, and explore the emergence of new cooperative behaviors as LLMs increasingly interact with humans and each other, potentially transforming how decisions are made across various systems.
大语言模型（LLMs）的出现有望改变自然科学和社会科学领域的研究，为理解复杂系统提供新的范式。特别是基于生成代理的模型（GABMs），它集成了大语言模型来模拟人类行为，由于其在广泛的人工环境中模拟复杂交互的潜力，已引起越来越多的公众关注。本文简要回顾了大语言模型在网络科学、进化博弈论、社会动力学和流行病建模等领域所发挥的颠覆性作用。我们评估了近期的进展，包括使用大语言模型预测社会行为、加强博弈论中的合作以及模拟疾病传播。研究结果表明，大语言模型可以再现类似人类的行为，如公平、合作和遵守社会规范，同时还具有成本效益、可扩展性和伦理简化等独特优势。然而，结果显示它们的行为存在与提示敏感性、幻觉甚至模型特征相关的不一致性，这表明在控制这些人工智能驱动的代理方面存在挑战。尽管大语言模型具有潜力，但要将其有效地整合到决策过程中——无论是在政府、社会还是个人层面——都需要解决偏差、提示设计挑战以及理解人机交互的动态。未来的研究必须改进这些模型，规范方法，并探索随着大语言模型越来越多地与人类以及彼此互动而出现的新合作行为，这可能会改变各个系统的决策方式。
