Reproduce “Dynamic Personality Adaptation in Large Language Models via State Machines” on OpenResearch

About this paper

Leon Pielage, Ole Hätscher, Mitja Back, Bernhard Marschall, Benjamin Risse

Researchers at the University of Münster developed a state-machine-based framework enabling Large Language Models (LLMs) to dynamically adjust their personality during conversations. This system successfully reproduces interpersonal complementarity, such as shifting from initial hostile to more communal behavior in virtual patients, demonstrating enhanced realism and adaptability in human-AI interactions.

Key insights

• Dynamic personality adaptation in LLMs can be effectively achieved through a state machine architecture grounded in established psychological models like the Interpersonal Circumplex.
• Lightweight, fine-tuned models can achieve comparable performance to larger LLMs for personality scoring, significantly reducing computational costs and enabling broader deployment for real-time applications.
• The framework successfully reproduces interpersonal complementarity, a fundamental aspect of human social interaction, in human-AI dialogues, indicating improved psychological realism.

Summary by alphaXiv.