How Personas Can Influence Agents to Play Split or Steal (arxiv.org)

Recent research has explored how persona prompts can influence the strategic behavior of AI agents in social dilemma scenarios, specifically in an iterated Split or Steal game. Powered by models like Ministral 3:3b and phi4:14b, these persona-driven agents interacted with a Virtual Human (VH) guided by GPT-4.1 mini across 160 sessions. The study found that mutual Split outcomes occurred approximately 74% of the time, indicating a significant inclination towards cooperation, while exploitation was relatively rare. Notably, different models showcased varied strategies: phi4 and Ministral 3:3b consistently demonstrated cooperative behavior, whereas Gemma models exhibited more unpredictable outcomes based on the temperature settings. This research is significant for the AI and machine learning community as it illustrates how nuanced persona prompts can effectively steer AI behavior in complex social interactions. The findings highlight the role of personality traits, showing that Prosocial and Principled personas were most inclined to cooperate, while Analytical personas were more likely to pursue exploitation strategies. Additionally, sentiment analysis revealed that discussions around friendship promoted cooperative outcomes, whereas topics related to money and vengeance were associated with exploitative behavior. These insights provide a foundation for future studies, particularly those involving human participants and embodied VHs, aimed at further understanding AI interactions in real-world social contexts.