Recursive Multi-Agent Systems: Scaling Agent Collaboration Through Recursion

A recent study presents RecursiveMAS, a framework that utilizes recursive scaling in multi-agent systems. This method builds on the idea of recursive language models, which progressively enhance computations across latent states, facilitating collaborative loops among diverse agents. At its core, RecursiveMAS incorporates a streamlined RecursiveLink module for generating in-distribution latent thoughts and transferring states between agents. An inner-outer loop learning algorithm fine-tunes the system by employing shared gradient-based credit assignment throughout the recursion rounds. Theoretical evaluations address runtime complexity and learning dynamics. The full paper can be accessed on arXiv with the ID 2604.25917.