Using Large Language Models to Implement Normative Reasoning Capabilities of Autonomous Agents

Normative reasoning is vital for autonomous agents to train and develop human-like social behaviours through established norms in human society. Researchers have proposed various models and frameworks to represent the process of norm emergence among a population of individuals or agents. However, these attempts do not concentrate on handling the agents that defect and oppose norms, prompting non-cooperation among other agents. Most studies in the field of metanorm punishments for penalizing agents that defect during social dilemmas tend only to focus on mathematical representations and abstract agent-agent interactions. However, social human interactions cannot be presented in mere rigid mathematical models. The development of Large Language Models (LLMs) trained over massive amounts of data written by humans has opened avenues to model human behaviour onto autonomous agents. Thereby, LLM agents have shown promise in normative reasoning and communicating norms to other agents. Our objective is to model the emergence of punishment-based metanorms previously represented as mathematical theories and conditional rules, using the natural language capabilities in LLM agents within a multi-agent society. This research proposal provides our proposed implementation to introduce a natural-language-centred approach to flexibly implement punishment to persuade normative LLM agents to adopt cooperative behaviours that can stabilize the multi-agent system.