Personas steer LLM agents to cooperate or exploit in Split or Steal game
74% mutual cooperation, but Analytical personas break trust 11% of the time.
A new preprint from Carlos Leon and colleagues explores how persona prompts shape strategic decision-making in LLM agents. The team ran 160 sessions of an iterated Split or Steal game (15 rounds each) in European Portuguese, pitting four open-source models — Ministral 3:3b, phi4:14b, Gemma3:12b, and Gemma4:e4b — against a Virtual Human (VH) powered by GPT-4.1 mini. Agents were tested at temperatures 0.3, 0.7, and 0 (deterministic), while the VH used a fixed prompt.
Results showed that mutual Split outcomes dominated across 74% of rounds, with exploitation occurring in fewer than 11% of rounds. Model choice strongly influenced behavior: phi4 and Ministral stayed consistently cooperative across all temperatures, while Gemma3:12b and Gemma4:e4b exhibited more varied strategies. Persona type also mattered — Prosocial and Principled personas were most reliably cooperative, whereas Analytical personas were more likely to steal. Topic analysis revealed that friendship-related dialogue correlated with Split decisions, while money- and vengeance-related language appeared in Steal outcomes. Sentiment analysis added little explanatory power beyond neutral/happy labels. The study provides a baseline for future virtual reality experiments with human participants.
- Four open LLMs tested: Ministral 3:3b, phi4:14b, Gemma3:12b, Gemma4:e4b at three temperatures (0, 0.3, 0.7) against GPT-4.1 mini.
- Over 160 sessions, mutual Split occurred in ~74% of rounds; exploitation in <11%.
- Prosocial and Principled personas cooperated most; Analytical personas were more likely to steal.
Why It Matters
Persona prompts significantly shape LLM strategic behavior, with implications for trust in AI-agent interactions.