Recursive Multi-Agent Systems

TL;DR

RecursiveMAS introduces a recursive multi-agent framework that enhances collaboration, reasoning, and efficiency across various tasks by iteratively refining system-wide latent states and optimizing through shared gradient-based learning.

Contribution

It extends recursive language model principles to multi-agent systems, enabling scalable agent collaboration via a unified recursive computation and an innovative co-optimization algorithm.

Findings

Achieves an average accuracy improvement of 8.3% over baselines.

Provides 1.2×–2.4× inference speedup.

Reduces token usage by 34.6%–75.6%.

Abstract

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 such scaling principle from a single model to multi-agent systems, and ask: Can agent collaboration itself be scaled through recursion? To this end, 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 the lightweight RecursiveLink module, enabling in-distribution latent thoughts generation and cross-agent latent state transfer. To optimize our 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 standard text-based MAS and maintains stable gradients during recursive training. Empirically, we instantiate RecursiveMAS under 4 representative agent collaboration patterns and evaluate across 9 benchmarks spanning mathematics, science, medicine, search, and code generation. In comparison with advanced single/multi-agent and recursive computation baselines, RecursiveMAS consistently delivers an average accuracy improvement of 8.3%, together with 1.2$\times$-2.4$\times$ end-to-end inference speedup, and 34.6%-75.6% token usage reduction. Code and Data are provided in https://recursivemas.github.io.