This paper highlights that despite advances in social simulation using large-scale language models (LLMs), teamwork performance in goal-oriented social situations remains underexplored. This ability is crucial for LLMs to effectively mimic human-like social behavior and form effective teams for task-solving. To this end, we present MetaAgents, a social simulation framework comprised of LLM-based agents. MetaAgents facilitates agents' participation in conversations and decision-making in social contexts, serving as a suitable platform for investigating agents' interactions and interpersonal decision-making. Specifically, we examine the team formation and skill-matching behaviors of LLM-based agents using a job fair as a case study. Quantitative metrics and qualitative text analysis are used to assess teamwork performance in the job fair setting. The evaluation results reveal that LLM-based agents skillfully make rational decisions for developing effective teams, but also identify Limitations, which hinders their effectiveness in more complex team formation tasks. This study provides valuable insights into the role and evolution of LLMs in goal-oriented social simulation.
