Researchers investigated how persona prompts influence the strategic behavior of large language model agents in a "Split or Steal" game. Using four open-source models (Ministral-3-3B, phi4:14b, Gemma3:12b, and Gemma4:e4b) interacting with a GPT-4.1 mini-powered virtual human, they found that mutual "Split" outcomes dominated, occurring in approximately 74 percent of rounds. The choice of model significantly impacted agent behavior, with phi4 and Ministral-3-3B consistently cooperative, while Gemma models showed more varied strategies. Persona traits like "Prosocial" and "Principled" were linked to cooperation, whereas "Analytical" personas were more prone to exploitation. AI
IMPACT Demonstrates how persona prompts can influence LLM agent behavior in strategic interactions, offering insights for developing more predictable and controllable AI agents.
RANK_REASON Academic paper detailing experimental results on LLM agent behavior. [lever_c_demoted from research: ic=1 ai=1.0]
- Big Tech
- European Portuguese
- Gemma3:12b
- Gemma4:e4b
- GPT-4.1 mini
- Hugging Face
- Ministral-3-3B
- phi4:14b
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
