Social Deduction Game

Social deduction games, like Werewolf and Among Us, are increasingly used as testbeds for evaluating the social intelligence and strategic reasoning capabilities of large language models (LLMs). Current research focuses on developing and benchmarking LLMs within these game environments, employing techniques like reinforcement learning and entropy-based search algorithms to improve agent performance and analyze LLM decision-making processes. These studies provide valuable insights into LLM limitations in understanding deception, persuasion, and complex social dynamics, ultimately contributing to the development of more sophisticated and socially aware AI systems. The findings also offer a novel approach to evaluating various aspects of social intelligence, such as intention understanding and opinion leadership.