📄 中文摘要
随着大型语言模型（LLMs）在社会和战略场景中的广泛应用，理解其行为与人类行为的差异变得至关重要。行为博弈理论（BGT）为分析行为提供了框架，但现有模型未能充分捕捉人类的特异性行为或像LLMs这样的黑箱非人类代理的行为。通过使用前沿的程序发现工具AlphaEvolve，直接从数据中发现人类和LLM行为的可解释模型，从而实现对驱动人类和LLM行为的结构性因素的开放式探索。在对迭代石头剪刀布的分析中，发现前沿的LLMs能够展现出比人类更深层次的战略行为。
发现人类与大型语言模型之间战略行为的差异
出处: Discovering Differences in Strategic Behavior Between Humans and LLMs
发布: 2026年2月12日
📄 English Summary
Discovering Differences in Strategic Behavior Between Humans and LLMs
As Large Language Models (LLMs) are increasingly utilized in social and strategic contexts, understanding the divergence in their behavior compared to humans becomes crucial. Behavioral Game Theory (BGT) offers a framework for analyzing behavior, yet existing models fail to fully capture the idiosyncratic behaviors of humans or the black-box nature of non-human agents like LLMs. Utilizing AlphaEvolve, a state-of-the-art program discovery tool, interpretable models of both human and LLM behavior are directly discovered from data, facilitating open-ended exploration of the structural factors influencing these behaviors. Analysis of iterated rock-paper-scissors reveals that frontier LLMs can exhibit deeper strategic behavior than humans.