用大语言模型模拟囚徒困境中的人类行为
原标题：Modeling Human Behavior in the Prisoner's Dilemma with Large Language Models
AI 摘要
该研究评估了五款开源大模型（Falcon 7B Instruct、Gemma 7B、Llama 3.1 8B Instruct、Phi Mini MoE Instruct、Qwen 2.5 7B Instruct）在重复囚徒困境博弈中模拟人类合作行为的能力，并设置固定对手（有记忆）与随机换对手（被告知对手会变）两种实验框架。结果显示所有模型与人类合作率均存在差异，开源模型普遍过度合作，随机换对手场景下人类平均合作率28.5%而多个模型超过80%；Qwen-2.5-7B-Instruct表现出确定性倾向，Phi-mini-MoE-Instruct波动更大。作者认为LLM有作为行为模拟代理的潜力，但当前开源模型尚不能可靠替代人类被试。
正文节选
Modeling Human Behavior in the Prisoner’s Dilemma with Large Language Models Abstract The Prisoner’s Dilemma has long served as a foundational framework for studying human decision-making and cooperative outcomes across numerous social scenarios, yet human behavior consistently deviates from predictions of fully rational economic theory. Previous works have extensively studied human and LLM outcomes in the Prisoner’s Dilemma; however, these studies have lacked at least one key feature: they eith