Definition
Plain language
A way of building multi-agent AI systems where the agents are positions in a graph that get trained jointly.
As stated in the literature
A framework treating multi-agent LLM systems as neural-network-like graphs of role-free nodes trained jointly via REINFORCE on a final reward, with progressive growth used to scale topology without retraining from scratch.
Why it matters: It removes the manual labor of designing multi-agent topologies and lets the system scale by growing the graph rather than starting over.
For example, instead of hand-designing 'planner' and 'critic' roles, NeuroMAS lets the graph train its nodes into specialized roles from the reward.
Heard on the show
“… " The paper is called "NeuroMAS: Multi-Agent Systems as Neural Networks with Joint Reinforcement Learning," it went up on arXiv …”Episode 060 — When Splitting One Model Across Three Agents Doubles Its Accuracy