EDCN: Expert-Driven Dynamic Recruitment and Collaboration Network for MAS

Abstract

Large Language Model (LLM)-driven Multi-Agent Systems (MAS) have recently demonstrated potential in complex reasoning tasks. However, real-world scenarios often impose higher demands on “who should collaborate, how they should collaborate, and whose opinions should carry more weight.” Fixed roles and static orchestration can easily lead to capability mismatches and redundant overhead. To address this issue, this paper proposes a task-adaptive collaboration framework for multi-agent systems. Expert agents first decompose natural language tasks and dynamically recruit the most suitable executors. During execution, a directed collaboration network is constructed, where peer evaluation and historical performance are incorporated into a weighted voting scheme, allowing more stable and reliable agents to gain greater decision-making power. After task completion, evaluation results are fed back into subsequent recruitment, enabling continuous evolution. This paper conducts comparative experiments on a causal reasoning dataset. On the first 30 samples, the system achieved 22/30 correct answers, with an accuracy of 73.3%, outperforming single-agent GPT (63.3%) and majority-vote MAS without evaluation (66.7%). These results verify the advantages of the proposed framework in terms of both correctness and robustness.