@inproceedings {pub-5576,
	title = {Meta-Defined Virtual Humans and Moderated Multiparty Dialogues: A Novel GPTs Agent Architecture},
	author = {Frank Joublin AND Antonello Ceravola AND Johane Takeuchi},
	year = {2024},
	month = {April},
	abstract = {This research delves into the burgeoning field of human-machine interface for multiparty open dialogues, focusing on the development of human simulators using ChatGPT-4. Our work introduces a novel framework where virtual humans are meta-defined, with their personalities, preferences, skills, and personal vitae generated by Large Language Models (LLMs) based on concise specifications. The dialogues{\textquoteright} topics are similarly meta-defined, transformed into dialog synopsis by an LLM. Each virtual human{\textquoteright}s dialogue contributions are synthesized by LLMs, ensuring consistency with their personality traits. A virtual moderator, also powered by an LLM, oversees the dialogue, guiding it in accordance with the defined synopsis.

Our analysis employs a blend of quantitative Key Performance Indicators (KPIs) and qualitative assessments conducted by another LLM instance. The results demonstrate that our approach can successfully generate realistic human dialogues in a multiparty context. This capability has significant implications for testing artificial systems, creating dialogue benchmarks, and evaluating human-machine interactions in dialogue settings. Our findings suggest that the use of meta-defined virtual humans and moderated dialogues can enhance the realism and applicability of simulated human-machine interactions.
},
	publisher = {HFSE},
	booktitle = {HFES2024},
	city = {L{\"u}beck}
}