How Personas Can Influence Agents to Play Split or Steal

2026-07-08 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Social Sciences & Behavioral Studies · Depth: Expert, extended

Summary

A study investigated how persona prompts influence large language model agents' strategic behavior in an iterated Split or Steal game. Researchers deployed 160 sessions, each with 15 rounds in European Portuguese, where agents powered by Ministral 3:3b, phi4:14b, Gemma3:12b, and Gemma4:e4b (at temperatures 0.3, 0.7, and zero) interacted with a GPT 4.1 mini-driven Virtual Human. Results showed mutual Split outcomes dominated roughly 74% of rounds, with exploitation occurring in fewer than 11%. Model choice significantly impacted behavior; phi4 and Ministral 3:3b remained consistently cooperative, while Gemma3:12b and Gemma4:e4b exhibited more diverse strategies. Personas categorized by Big Five traits revealed Prosocial and Principled types were most cooperative, whereas Analytical personas were more prone to exploitation. Dialogue analysis linked friendship topics to Split decisions and money/vengeance to Steal outcomes, while sentiment analysis provided limited insights. These findings establish a baseline for future virtual reality studies involving human participants.

Key takeaway

For NLP Engineers designing conversational agents for trust-sensitive applications, you should carefully select base models and persona prompts. Your choice of model, like phi4 or Ministral 3:3b, can bias agents towards cooperation, while Gemma models offer more strategic variability. When aiming for cooperative behavior, prioritize Prosocial or Principled personas. Be aware that Analytical personas may increase exploitation risk, and dialogue content can signal shifts in strategy.

Key insights

Persona prompts can shape LLM agent cooperation in social dilemmas, but model choice and personality traits significantly mediate outcomes.

Principles

• LLM agent cooperation varies by model architecture.
• Big Five personality traits influence agent strategy.
• Dialogue topics correlate with cooperative decisions.

Method

LLM agents with 20 unique Big Five-classified personas played 15-round Split or Steal games against a fixed-prompt Virtual Human in European Portuguese. Outcomes were analyzed by cooperation rate, strategy classification, and dialogue content (topics, sentiment).

In practice

• Use Prosocial or Principled personas for cooperative LLM agents.
• Avoid Analytical personas if aiming for high cooperation.
• Monitor dialogue topics for early indicators of defection.

Topics

• Large Language Models
• Persona Prompting
• Split or Steal Game
• Agent Behavior
• Big Five Personality
• Game Theory Strategies
• European Portuguese NLP

Best for: Research Scientist, AI Scientist, NLP Engineer, AI Engineer

Related on AIssential

• From Biased Chatbots to Biased Agents: Examining Role Assignment Effects on LLM Agent Robustness — Demographic persona assignments introduce implicit biases, significantly degrading LLM agent performance on action-oriented tasks.
• The best answer to this question I've seen yet. — AI models like Claude can exhibit unexpected, human-like "sass" or persona adoption based on user prompts.
• LLM-Driven Personalities for Decision Making in Emergency Simulations — LLMs can drive virtual human decision-making by incorporating personality traits via prompts, enhancing simulation realism.
• The persona selection model — LLMs simulate diverse personas during pre-training, with post-training refining a specific "Assistant" character.
• From Parasocial Scripts to Dyadic Persistence in Autonomous AI-Agent Communities — Autonomous AI agents exhibit parasocial interaction cues that foster persistent dyadic communication patterns.
• Creative Collision: Directorial Persona Steering and Competition in Large Language Models — The superposition of opposing steering vectors reveals complex interactions and shared representational substrates in LLMs.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.