By harnessing GPT-4’s ability to mimic consistent personality traits, SMU researchers are paving the way for NPCs that think, react, and evolve like real people, transforming how stories are told in video games.
Research: Driving Generative Agents With Their Personality. Image Credit: Shutterstock AI
*Important notice: arXiv publishes preliminary scientific reports that are not peer-reviewed and, therefore, should not be regarded as definitive, used to guide development decisions, or treated as established information in the field of artificial intelligence research.
When Jake Klinkert was growing up, his father suggested that since he loved video games, he should make them. Klinkert took those words to heart and went on to earn a Master of Interactive Technology in Digital Game Development from SMU Guildhall. Now a PhD student in the Computer Science Department at SMU's Lyle School of Engineering, he is testing large language models (like ChatGPT) to create non-playable characters (NPCs) that act and respond more like real people.
In testing that generated more than 50,500 results, Klinkert and his colleagues found that GPT-4 achieved 73.98% accuracy in maintaining consistent personality traits—a significant improvement from earlier AI models that scored below 18%.
Addressing a gaming challenge
The research addresses a longstanding challenge in the gaming industry: creating NPCs that display realistic emotional complexity and consistent behavioral patterns instead of the limited, repetitive responses that often break player immersion.
"This represents a shift in how we can approach character development," explained Klinkert. "We've moved from a world where creating believable AI characters required complex systems and extensive technical resources to one where developers can quickly prototype personality-driven characters using text-based interactions."
Evaluating AI personalities
Researchers used the International Personality Item Pool questionnaire, a 50-item test based on the Big Five personality model, to evaluate three OpenAI models: text-davinci-003, gpt-3.5-turbo-0613, and gpt-4-0613. The Big Five measures personality across five dimensions:
- Openness: High creativity, readily embraces novelty, driven by tackling new challenges, engages in abstract thought
- Conscientiousness: How organized and responsible someone is in their work and with other people
- Extraversion: How outgoing and social someone is when around other people
- Agreeableness: How willing someone is to consider different opinions and work toward shared understanding with others
- Neuroticism: How anxiously or emotionally someone tends to view and react to situations
Gaming industry impact
Video game companies have long sought to integrate affective computing (technology that recognizes, understands, and responds to human emotions) into their games. This paper suggests that advanced language models could significantly enhance their efforts by generating dialogue and decisions that authentically reflect specific personality traits.
The implications extend beyond simple dialogue systems. Klinkert envisions NPCs that could retell stories from their unique personality-driven perspectives, contribute to evolving narratives through improvisation, or solve in-game mysteries using personality-guided intuition. The advancement comes as players increasingly expect more sophisticated and emotionally engaging experiences.
Future directions
The study's complete dataset and results are publicly available through a GitLab repository, enabling other developers and researchers to build upon this work. The team is now exploring partnerships with game development studios to implement their findings in commercial projects.
Other researchers contributing to the study include Corey Clark, Deputy Director for Research at SMU Guildhall and Associate Professor of Computer Science in the Lyle School of Engineering, and Steph Buongiorno, a recent postdoctoral fellow at SMU Guildhall.
*Important notice: arXiv publishes preliminary scientific reports that are not peer-reviewed and, therefore, should not be regarded as definitive, used to guide development decisions, or treated as established information in the field of artificial intelligence research.
Source:
Journal reference:
- Preliminary scientific report.
Klinkert, L. J., Buongiorno, S., & Clark, C. (2024). Driving Generative Agents With Their Personality. ArXiv. https://arxiv.org/abs/2402.14879
AI Framework Speeds Up Relay Catalysis Design With Transparent, Literature-Linked Pathways