D³MAS: Decompose, Deduce, and Distribute for Enhanced Knowledge Sharing in Multi-Agent Systems

Heng Zhang, Yuling Shi, Xiaodong Gu, Haochen You, Zijian Zhang, Lubin Gan, Yilei Yuan, Jin Huang

Keywords: Multi-agent Systems, Multi-agent Reasoning, Graph-based Learning, Knowledge Sharing, Collaborative Inference

Abstract: Multi-agent systems powered by large language models (LLMs) exhibit strong capabilities in collaborative problem-solving. However, these systems often face significant knowledge redundancy, where agents duplicate efforts in retrieval and reasoning processes. This leads us to ask: \textit{Can current knowledge sharing mechanisms effectively reduce redundancy in multi-agent reasoning?} Empirical analysis reveals that agents still experience an average knowledge duplication rate of 47.3\%, highlighting inefficiencies in current approaches. To tackle this, we propose \textbf{D³MAS}(\textbf{D}ecompose, \textbf{D}educe and \textbf{D}istribute), a hierarchical coordination framework designed to reduce redundancy and enhance reasoning quality. D³MAS organizes the collaborative process into three key layers: task decomposition for segmenting complex queries, collaborative reasoning for cooperative inference, and distributed memory for sharing complementary knowledge. Using these layers, agents coordinate through structured message passing in a unified heterogeneous graph, eliminating the need for explicit protocols. Experiments demonstrate that D³MAS consistently boosts reasoning accuracy by 8.7\% to 15.6\% on four challenging datasets and reduces knowledge redundancy by 46\% on average, leading to substantial improvements in computational efficiency.

Area: Generative and Agentic AI (GAAI)

Generative A I: I acknowledge that I have read and will follow this policy.

Submission Number: 1830