Water-MAS: A multi-agent LLM framework with instruction-data decoupling for smart water management.

Area of Science:

• Environmental Science
• Artificial Intelligence
• Data Science

Background:

• Sustainable water management requires advanced processing of complex water data.
• Single-agent systems lack the accuracy and adaptability for multi-step water management tasks.
• Limitations in big data analysis hinder effective decision-making in water resource management.

Purpose of the Study:

• To introduce Water-MAS, a novel LLM-based multi-agent framework for water management.
• To enhance the accuracy, adaptability, and big data analysis capabilities for water-related workflows.
• To enable automated tasks in environmental monitoring and data analysis.

Main Methods:

• Developed a multi-agent framework (Water-MAS) with planning, execution, and checking agents.
• Implemented an instruction-data decoupling mechanism to optimize computational overhead.
• Conducted experiments on tool invocation tasks and analyzed performance across different foundation models.

Main Results:

• Multi-agent collaboration under instruction-data decoupling achieved 89% higher recall compared to single-agent systems.
• The decoupling mechanism improved recall by 143% within the multi-agent framework.
• Smaller foundation models demonstrated comparable performance to larger counterparts, offering flexibility in system design.

Conclusions:

• Water-MAS provides a reusable framework for deploying LLM-based multi-agent systems in water management.
• The framework significantly enhances task automation and data-driven insights for environmental monitoring.
• Findings guide the selection of foundation models to balance performance and computational resources.