Dynamic Knowledge Integration in Multi-Agent Systems for Content Inference

Abstract

Advancements in cutting-edge science and technology have resulted from the integration of multiple interdisciplinary domains beyond traditional academic boundaries. Achieving effective cross-domain knowledge-sharing and consensus-building is crucial. However, single-agent Large Language Models (LLMs) solutions often struggle to integrate the diverse and highly specialized knowledge required in these contexts. This study proposes a multi-agent system with dynamic knowledge integration, where multiple specialized LLM-based agents cooperatively infer content by referencing different domain-specific databases. Each agent selectively and dynamically updates references based on conversational context to achieve deeper insight and more robust solutions. We propose four system architectures---Decentralized, Centralized, Layered, and Shared Pool---for agent coordination. We then evaluate these approaches on a title-to-abstract inference task using a subset of the arXiv dataset, demonstrating that multi-agent systems significantly outperform single-agent models in both accuracy and stability. Notably, expert agents, restricted to domain-specific data, produce more precise and consistent outputs, and the Decentralized architecture fosters increased domain interaction. These findings suggest that the collaboration of specialized multi-agent systems can more effectively facilitate the consensus-building process in the advancement of complex interdisciplinary scientific domains.