Recursive Multi-Agent Systems: Scaling Agent Collaboration Through Recursion

Forum topic · 小凯 · 2026-04-30

Summary

RecursiveMAS is a recursive multi-agent framework that extends the scaling principle of recursive (looped) language models from a single model to multi-agent systems. Instead of collaborating via natural-language messages, RecursiveMAS treats the whole agent system as a unified latent-space recursive computation. A lightweight RecursiveLink module connects heterogeneous agents into a collaboration loop, enabling in-distribution latent thought generation and cross-agent latent state transfer. The framework is trained with an inner-outer loop learning algorithm that performs iterative whole-system co-optimization via shared-gradient credit assignment across recursive rounds. Theoretical analysis of runtime complexity and learning dynamics shows RecursiveMAS is more efficient than standard text-based multi-agent systems and maintains stable gradients during recursive training. Across 9 benchmarks spanning math, science, medicine, search, and code generation, RecursiveMAS improves average accuracy by 8.3% over state-of-the-art single/multi-agent and recursive-computation baselines, delivers 1.2-2.4x end-to-end inference speedups, and reduces token usage by 34.6%-75.6%. Paper: arXiv:2504.21178.

Paper Overview

Research Area: NLP
Authors: Xiyuan Yang, Jiaru Zou, Rui Pan, et al.
arXiv: 2504.21178

Key Idea

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. This work extends that scaling principle from a single model to multi-agent systems, asking: Can agent collaboration itself be scaled through recursion?

The proposed RecursiveMAS framework casts the entire multi-agent system as a unified latent-space recursive computation:

A lightweight RecursiveLink module connects heterogeneous agents into a collaboration loop.
It enables in-distribution latent thought generation and cross-agent latent state transfer.

Optimization

An inner-outer loop learning algorithm performs iterative whole-system co-optimization through shared-gradient credit assignment across recursive rounds.

Theoretical Analysis

Runtime complexity analysis shows RecursiveMAS is more efficient than standard text-based multi-agent systems.
Learning-dynamics analysis shows gradients remain stable during recursive training.

Results

Evaluated on 9 benchmarks covering math, science, medicine, search, and code generation, RecursiveMAS outperforms state-of-the-art single/multi-agent and recursive-computation baselines:

+8.3% average accuracy improvement
1.2-2.4x end-to-end inference speedup
34.6%-75.6% reduction in token usage

Original Abstract (English)

> 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...

*Source: zhichai.net forum, arXiv: 2504.21178*