How Personas Can Influence Agents to Play Split or Steal
A new arXiv paper tests how assigning personas to AI agents affects their strategic decisions in an iterated Split or Steal game against a fixed-prompt virtual human.
- Personas assigned to AI agents can influence their strategic decisions in social dilemma games like Split or Steal.
- The study tested four open models (Ministral 3:3b, phi4:14b, Gemma3:12b, Gemma4:e4b) with varying temperature settings to assess behavior consistency.
- Agents interacted with a Virtual Human controlled by a fixed prompt, isolating the effect of persona prompts on decision-making.
- Findings highlight the need for careful consideration of persona design in AI agents for reliable and predictable strategic behavior.
Researchers from an unreleased institution published a study on arXiv examining how personas assigned to large language model agents impact their behavior in an iterated Split or Steal game. The experiment involved agents instantiated from four open models (Ministral 3:3b, phi4:14b, Gemma3:12b, and Gemma4:e4b) at two temperature settings (0.3 and 0.7) and a deterministic mode. Each agent interacted with a Virtual Human controlled by a fixed prompt, revealing how persona-driven prompts influence strategic decision-making in social dilemmas.
The study addresses a gap in understanding how personas, commonly used to guide AI agents, affect their behavior in competitive or cooperative scenarios. By testing multiple models and temperature settings, the research provides insights into the reliability and predictability of persona-driven agents in strategic interactions.
Source: How Personas Can Influence Agents to Play Split or Steal. Read the full piece at the source.
Provides insights into designing personas for AI agents to ensure predictable and strategic behavior in competitive or cooperative scenarios.
Offers a case study on how prompt engineering and model parameters influence AI decision-making in social dilemmas.
Demonstrates how subtle changes in AI prompts can significantly alter behavior in real-world-like scenarios.
- Split or Steal
- A social dilemma game where two players must decide how to split a resource, often used to study cooperation and competition.
- Temperature setting
- A parameter in AI models that controls the randomness of outputs; lower values make responses more deterministic.
AI Research"We cannot choose to become idiots": The AI cheating scandal roiling Brown University
AI ResearchI Built a Self-Improving AI, and So Can You
Biohub researchers use artificial intelligence to uncover new psoriasis targets - News-Medical
USC researchers develop low-cost AI tool to help cities map urban tree canopy - USC Today
‘We also need to talk about artificial intelligence's positive potential,' says AI for Good founder - Geneva Solutions
AI ToolsWhy AI Infrastructure must evolve for Agent Experience — Akshat Bubna, Modal CTO
Modal's cofounder Akshat Bubna argues that AI agent infrastructure has matured enough to support reliable agent experiences, two years after the company's initial coverage.
FundingLovable reportedly in talks to double its valuation to $13.2B
AI assistant startup Lovable is reportedly in talks to raise a $300 million funding round, potentially doubling its valuation to $13.2 billion.
Police use of artificial intelligence grows as rules lag behind - timesdaily.com
Police departments increasingly adopt AI tools while oversight frameworks struggle to keep pace, according to a new report.
Alphabet's Artificial Intelligence (AI) Spending Spree Is Great News for Nvidia - The Motley Fool
Alphabet's increased AI spending is expected to benefit Nvidia's business, according to The Motley Fool.
AI ToolsNetflix AI Team Cuts Wide-Partition Read Latency from Seconds to Milliseconds by Splitting Cassandra Partitions Per ID
Netflix’s AI team reduced Cassandra partition read latency from seconds to tens of milliseconds by dynamically splitting oversized partitions per TimeSeries ID.
SecurityGoogle’s deepfake detector system used to debunk McConnell hoax pic
Google's deepfake detection system successfully identified and debunked a fabricated image of Senator Mitch McConnell circulating online.