Loki: Studying MARL Collusion using LLMs in a Kuhn Poker Environment

Calvin Qin

Track: Main Track

Keywords: gametheory, dpo, reasoning, collusion, advantage

TL;DR: This paper introduces a system that uses Large Language Models to study how bidirectional communication enables collusion between players in a three-player Kuhn Poker environment, analyzing the effectiveness of different communication strategies.

Abstract: Collusion and communication have long been underexplored aspects of strategic decision-making in games with hidden information. While systems like Libratus have demonstrated superhuman performance in games such as poker, little research has examined the role of table talk and how players might collude during gameplay. This paper investigates collusive behavior using Large Language Models (LLMs) in a three-player variant of Kuhn Poker. The proposed system enables two LLM agents to communicate bidirectionally, sharing private information to coordinate strategies against a non-colluding opponent. By employing prompt engineering to vary the degree of collusion, this study analyzes how different levels of communication influence the emergence, dynamics, and effectiveness of collusion. The findings reveal distinct stages of collusive behavior, characterized by the agents' communication strategies and gameplay decisions, offering insights into the intersection of language, strategy, and ethical considerations in AI systems.

Submission Number: 13