How Personas Can Influence Agents to Play Split or Steal

Short summary

Researchers tested how persona prompts shape LLM agent behavior in an iterated Split or Steal trust game, running 160 sessions across four open models against a GPT 4.1 mini-powered virtual human. Prosocial and Principled personas consistently cooperated while Analytical personas were more likely to exploit, and model choice significantly affected strategy—phi4 and Ministral stayed cooperative across temperatures while Gemma models showed more varied behavior. The study establishes a baseline for future VR experiments with human participants interacting with embodied agents.

• •Persona prompts significantly influence LLM agent cooperation vs exploitation in trust games
• •Prosocial and Principled personas cooperate most; Analytical personas exploit more often
• •Model choice matters: phi4 and Ministral 3:3b stay cooperative while Gemma models show varied strategies

Generated with AI, which can make mistakes.