Value-based large language model agent simulation for mutual evaluation of trust and interpersonal closeness

Area of Science:

• Artificial Intelligence
• Social Simulation
• Computational Social Science

Background:

• Large language models (LLMs) simulate complex social phenomena using agents with defined traits.
• Value similarity is crucial for trust and relationships in human societies but unexplored in artificial LLM societies.

Purpose of the Study:

• To investigate the influence of value similarity on relationship-building among LLM agents.
• To assess LLM agent simulations as a testbed for social science theories.

Main Methods:

• Two experiments were conducted: a preliminary study on LLM value controllability and a main experiment on agent interactions.
• Pairs of LLM agents with controlled values engaged in dialogue to assess mutual trust and interpersonal closeness.
• Experiments were performed in English and Japanese to examine language dependence.

Main Results:

• LLM agents with higher value similarity demonstrated significantly greater mutual trust and interpersonal closeness.
• The study confirmed the positive correlation between value alignment and relationship quality in artificial agents.

Conclusions:

• LLM agent simulations are a valid platform for testing social science theories.
• Value similarity mechanisms influencing relationship building are elucidated.
• Findings provide a foundation for new social science theories and insights.