RecursiveMAS
A multi-agent framework that scales agent collaboration by treating a heterogeneous team of LLM agents as one unified recursive computation, connected by lightweight modules that exchange and refine shared latent state.
RecursiveMAS casts a whole multi-agent system as a unified recursive computation rather than treating each large language model agent as an isolated module. Heterogeneous agents are connected by lightweight RecursiveLink modules that let them exchange, refine and evolve latent states across recursion rounds, and an Inner-Outer Loop training paradigm progressively co-optimizes the system: an inner loop gives each agent a model-level warm start, and an outer loop trains the RecursiveLink connections across agents at the system level.
Across nine benchmarks spanning mathematics, science, medicine, search and code generation, the paper reports that RecursiveMAS improves multi-agent coordination for sequential, mixture, distillation and deliberation collaboration patterns. The repository provides training implementations and data, inference and evaluation pipelines, released model checkpoints, and an interactive playground demo.