Efficient and Interpretable Multi-Agent LLM Routing via Ant Colony Optimization

Jiaquan Zhang, Xudong Wang, Chaoning Zhang, Chenghao Li, Qigan Sun, Sung-Ho Bae, Peng Wang, Ning Xie, Jie Zou, Yang Yang, Heng Tao Shen

Keywords: Large Language Models，Multi-Agent

Abstract: The instruction-following and semantic understanding capabilities of large language models (LLMs) serve as the core competence of Multi-Agent Systems (MAS), enabling collective strategy coordination. However, task routing in MAS remains a critical performance bottleneck, especially in dynamic and resource-constrained environments. Existing LLM-based routing approaches often suffer from limited transparency, static allocation strategies, and insufficient system state awareness. To address these challenges, we propose AMRO: Ant colony inspired Multi-agent Routing Optimization. AMRO models agent interactions as a function-based directed graph, utilizing a pheromone-driven node update mechanism and an adaptive pheromone decay strategy to achieve real-time perception and response to environmental changes, thereby continuously optimizing routing assignments. This approach significantly enhances routing efficiency and overall system performance, while the pheromone-guided path selection offers strong interpretability for the routing process. We conduct extensive experiments on five public benchmark datasets. The results show that AMRO achieves an average improvement of 1.97\% in pass@1 accuracy over the baseline and demonstrates superior efficiency and robustness under high concurrency. These findings indicate that AMRO provides an efficient and interpretable solution to the routing problem in LLM-based MAS.

Primary Area: foundation or frontier models, including LLMs

Submission Number: 10824