[3J5-GS-5] Agents:

[3J5-GS-5-01] Multi-agent system for dynamic Expert knowledge linkage.

〇Atsushi Yamamoto¹, Taito Naruki², Akihiko Katagiri¹, Yudai KoikKoike¹, Takumi Iida², Ryuta Shimogauchi², Kota Shimomura², Eri Onami², Koki Inoue², Osamu Ito¹ (1. Honda Honda Motor Co., Ltd., 2. Elith Inc.)

Keywords:Agent, LLM, RAG

This research proposes a multi-agent system that dynamically integrates and updates knowledge, focusing on multi-layered communication challenges in complex product development sites such as the automotive industry. The system achieves cross-disciplinary knowledge integration by having multiple specialized agents cooperate with each other, which has been a limitation of a single agent in the past. Each agent consults different databases and dynamically selects and updates the most appropriate references according to the context of the conversation, leading to new perspectives and deeper insights. The proposed system employs four types of schemes: Decentralized, Centralized, Layered, and Shared-Pool, and their properties are compared. Experiments on the arXiv dataset showed that multi-agents are more accurate and stable than single-agents, and in particular, the use of expert agents in the specialized domain enables accurate inference of relevant knowledge. The results suggest that the Decentralized model can accelerate the exchange of diverse expert knowledge through dense interaction and provide more accurate results.