shell pulsar
Recursive or looped language models have recently emerged as a new scaling axis by iteratively refining the same model computation over latent states to deepen reasoning. We extend this scaling principle from a single model to multi-agent systems, and ask: can agent collaboration itself be scaled through recursion?
We introduce RecursiveMAS, a recursive multi-agent framework that casts the entire system as a unified latent-space recursive computation. RecursiveMAS connects heterogeneous agents as a collaboration loop through a lightweight RecursiveLink module, enabling in-distribution latent thoughts generation and cross-agent latent state transfer. To optimize the framework, we develop an inner–outer loop learning algorithm for iterative whole-system co-optimization through shared gradient-based credit assignment across recursion rounds.
Theoretical analyses of runtime complexity and learning dynamics establish that RecursiveMAS is more efficient than text-based MAS and maintains stable gradients during recursive training. Empirically, we instantiate RecursiveMAS under 4 representative collaboration patterns and evaluate across 9 benchmarks spanning mathematics, science, medicine, search, and code. Compared with advanced single/multi-agent and recursive computation baselines, RecursiveMAS consistently delivers an average accuracy improvement of +8.3%, together with 1.2×–2.4× speedup, and 34.6%–75.6% token reduction. https://recursivemas.github.io/
