On May 3, 2026, researchers Carlos Leon and his colleagues published a paper examining how personas influence agents in social dilemma scenarios, specifically the Split or Steal game. Their study, featuring various language model agents, reveals insights into cooperative behavior and exploitation tendencies in strategic interactions.
The Role of Personas in Decision-Making
Personas are utilized to guide large language model agents, but their effectiveness is often questioned. This research investigates how different persona prompts affect the decisions of agents in an iterated Split or Steal game. In this game, agents interact with a Virtual Human (VH) controlled by a fixed prompt, providing a unique framework to analyze strategic behavior.
The study involved four open models: Ministral 3:3b, phi4:14b, Gemma3:12b, and Gemma4:e4b, tested at different temperature settings (0.3, 0.7, and deterministic decision with zero temperature). The VH was powered by GPT 4.1 mini, allowing for a controlled environment in which the agents operated.
Findings on Agent Behavior and Cooperation
Across 160 sessions of 15 rounds each, the results showed that mutual Split outcomes dominated, occurring in approximately 74 percent of rounds. In contrast, exploitation occurred in fewer than 11 percent of rounds. Notably, the choice of model significantly impacted agent behavior. Models such as phi4 and Ministral 3:3b exhibited consistent cooperation, while Gemma3:12b and Gemma4:e4b displayed varied strategies.
Further analysis based on the Big Five personality traits indicated that Prosocial and Principled personas were the most cooperative, while Analytical personas tended to exploit the VH. This highlights how different personality traits influence strategic decision-making in social dilemmas.
Implications for Future Research
The findings from this study not only characterize the interaction between persona prompts and model differences but also serve as a foundational basis for upcoming studies involving human participants in virtual reality settings. The research emphasizes the importance of understanding persona-driven behavior in strategic games, which could have broader implications in fields such as artificial intelligence and human-computer interaction.
- 74% of rounds resulted in mutual Split outcomes
- Exploitation occurred in less than 11% of rounds
- Cooperation was consistent in models phi4 and Ministral 3:3b
🤖 This article was rewritten by Feed and Figures' editorial AI from a report originally published by arXiv NLP. Facts and quotes are preserved from the original; the rewrite focuses on clarity and structure. For the unedited original, see the source link below.
