Presentation Information
[3Yin-A-46]Leadership Emergence in Personality Assigned LLM-based Multi-Agent Systems:Simulation and Evaluation via Automated Dialogue
〇Chie Yamanishi1, Yuki Nozaki2, Yoshinobu Kano1 (1. Shizuoka University, 2. Konan University)
Keywords:
Large Language Model,Leadership Emergence,Multi Agent Simulation
Optimizing team formation and leader selection is a critical management issue. Recent leadership research has shifted from a "leader-centric" approach to a "relational approach" where leadership emerges from member interactions. However, validating this approach with human experiments is costly and difficult to observe internal psychological states. This study proposes a multi-agent simulation using Large Language Models (LLMs) to reproduce and evaluate the leadership emergence process. We constructed agents based on Yukl's taxonomy (Task/Relations/Change) and separated their "Inner Thought" from "Utterance" to simulate psychological conflicts.
Experiments showed that a "Shared Leadership" pattern (PatternF), where no specific leader exists, achieved consensus comparable to a strong leader pattern. Furthermore, in comparison with human experiments, while objective evaluation by LLM showed a positive correlation with self-evaluation (r=0.512), human peer evaluation showed a negative correlation (r=-0.202). This reveals a discrepancy between "objective behavior" and "human perception," suggesting the efficacy of LLM-based organizational diagnosis.
Experiments showed that a "Shared Leadership" pattern (PatternF), where no specific leader exists, achieved consensus comparable to a strong leader pattern. Furthermore, in comparison with human experiments, while objective evaluation by LLM showed a positive correlation with self-evaluation (r=0.512), human peer evaluation showed a negative correlation (r=-0.202). This reveals a discrepancy between "objective behavior" and "human perception," suggesting the efficacy of LLM-based organizational diagnosis.