NeuroMAS: Multi-Agent Systems as Neural Networks with Joint Reinforcement Learning
For researchers building multi-agent LLM systems, NeuroMAS shifts design from hand-crafted workflows to scalable architecture design, offering a new paradigm for improving capability through organizational scaling.
NeuroMAS treats multi-agent language systems as trainable neural-network-like architectures with LLM agents as nodes, enabling role-free, structure-aware coordination via reinforcement learning. It significantly outperforms both inference-time and trained multi-agent baselines, and shows that progressive growth from smaller trained systems enables scaling to larger architectures.
Multi-agent language systems are often built as hand-designed workflows, where agents are assigned semantic roles and communication protocols are specified in advance. We propose NeuroMAS, a method that first treats a multi-agent language system as a trainable and scalable neural-network-like architecture with LLM agents as nodes and intermediate textual signals as edges. In NeuroMAS, agent nodes are role-free but structure-aware: the topology only determines how information can flow in general, while reinforcement learning training determines how nodes communicate, specialize, and coordinate. This formulation shifts multi-agent design from workflow engineering toward architecture design, where depth, width, connectivity, and growth protocol become scalable sources of capability. Further, we provide a theoretical perspective showing why such modular textual computation is more parameter-efficient when tasks admit hierarchical decompositions. Experiments show that NeuroMAS improves significantly over both inference-time and trained multi-agent baselines. We further find that organizational scaling is path-dependent: larger systems can be challenging to train from scratch, but become feasible when grown progressively from smaller trained systems. These results suggest that learned neural multi-agent systems are a promising scaling axis for LLMs.