This study emphasizes the importance of evaluating LLM's performance and interactions in dynamic, multi-agent environments, given its application in diverse human activities such as business negotiation and group collaboration. Unlike existing statistical agents that excel in well-defined environments, LLM can generalize across a variety of real-world scenarios. This study compares and analyzes the outcomes and behavioral dynamics of humans (N=216), LLM (GPT-4o, Gemini 1.5 Pro), and Bayesian agents in a dynamic negotiation environment. Bayesian agents generate high surplus value through aggressive optimization but suffer from frequent deal rejections. While humans and LLMs achieve similar total surplus value, LLMs prefer conservative and concessional trades, while humans exhibit more strategic, risk-taking, and fairness-seeking behavior. Therefore, performance equivalence can mask fundamental differences in process and alignment, which is crucial for applying LLM to real-world collaborative tasks. This study establishes a baseline behavioral baseline under consistent conditions, laying the foundation for future research.
