Deciding the Path: Leveraging Multi-Agent Systems for Solving Complex Tasks

Iman Abbasnejad, Xuefeng Liu, Atanu Roy; Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR) Workshops, 2025, pp. 4216-4225

We present a multi-agent framework that enhances the capabilities of LLMs through intelligent task distribution and resource optimization for solving complex problems. The framework uses a dynamic routing mechanism that automatically delegates queries to specialized agents, complemented by an efficient tool selection system that reduces computational complexity. This autonomous architecture eliminates the need for human intervention while maintaining high performance across diverse tasks. Through comprehensive empirical evaluation on three challenging benchmarks, our multi-agent framework outperforms existing state-of-the-art LLMs and even specialized systems. Our method achieves 90.29% accuracy on Math 401 (surpassing MathViz-E's 89.53%), 91.3% pass@1 on MBPP (exceeding QualityFlow's 88.53%), and obtains state-of-the-art valid efficiency and execution scores of 56.28% and 54.39% respectively on BIRD SQL (outperforming DeepSeek-V3's 51.28% and 51.02%). These findings validate the effectiveness of our collaborative multi-agent approach in enhancing LLM capabilities, demonstrating significant improvements in accuracy, reasoning depth, and computational efficiency across varied task domains. Our research establishes a foundation for more adaptable and sophisticated AI systems capable of addressing complex real-world problems.