X-MAS: Towards Building Multi-Agent Systems with Heterogeneous LLMs

Source: arXiv Authors: Rui Ye, Xiangrui Liu, Qimin Wu, Xianghe Pang, Zhenfei Yin, Lei Bai, Siheng Chen Date: 2025-05-22 Primary category: cs.AI All categories: cs.AI, cs.CL, cs.MA

Abstract

X-MAS asks a coordination question that many multi-agent papers dodge: why should every role run the same model? The paper benchmarks 27 LLMs across domains and multi-agent functions, then shows that assigning different models to different roles can improve performance substantially without redesigning the surrounding structure. It is a strong argument that coordination should include model ecology and capability routing, not just conversation structure.