Language Games Among LLMs: Modeling Social Norm Formation with AI Agents in Open-Ended Games

Papers like Sun et al. (2024) and Mao et al. (2023) highlight the use of LLMs for simulating strategic behavior, but mostly within the confines of classical or benchmark games. This research would push the envelope by using LLM agents in open-ended, language-rich environments—akin to “language games”—to observe how social norms, linguistic conventions, and even ethical codes can emerge and evolve. The framework could blend evolutionary game theory with natural language processing, allowing agents to invent, negotiate, and enforce rules in real time. By systematically varying initial conditions, communication constraints, or reward structures, the research could uncover the game-theoretic principles underlying the formation and stability of social conventions—not just among humans, but in future AI societies as well. This has deep implications for AI alignment, human-AI interaction, and the study of social complexity.