NeuroMAS: Multi-Agent Systems as Neural Networks with Joint Reinforcement Learning

A new method called NeuroMAS reimagines multi-agent language systems as trainable neural-network-like architectures. Instead of hand-designed workflows with predefined roles and communication protocols, NeuroMAS treats LLM agents as nodes and textual signals as edges. The topology determines information flow, while reinforcement learning trains nodes to communicate, specialize, and coordinate. This shifts design from workflow engineering to architecture design, where depth, width, connectivity, and growth become scalable. A theoretical perspective shows modular textual computation is more parameter-efficient for hierarchically decomposable tasks.