An appraisal-based chain-of-emotion architecture for affective language model game agents

Area of Science:

• Artificial Intelligence
• Cognitive Science
• Human-Computer Interaction

Background:

• Developing believable and interactive digital artificial agents is a growing research area.
• Simulating human emotions in artificial agents faces significant theoretical and technical challenges.
• Large language models (LLMs) offer potential solutions by leveraging common patterns in situational appraisal.

Purpose of the Study:

• To test the capability of LLMs in solving emotional intelligence tasks.
• To evaluate the effectiveness of LLMs in simulating emotions in digital agents.
• To introduce and assess a novel Chain-of-Emotion architecture for emotion simulation in video games.

Main Methods:

• Three empirical experiments were conducted to assess LLM performance.
• A new Chain-of-Emotion architecture was developed, grounded in psychological appraisal theory.
• LLM architectures were compared using user experience and content analysis metrics.

Main Results:

• LLMs demonstrated capabilities in emotional intelligence tasks and emotion simulation.
• The Chain-of-Emotion architecture outperformed control LLM architectures.
• Significant improvements were observed in user experience and content analysis.

Conclusions:

• LLMs show promise for creating more natural and interactive digital artificial agents.
• The Chain-of-Emotion architecture provides a viable method for affective agent development.
• This research offers foundational evidence for constructing and testing affective agents based on cognitive processes within LLMs.