Recursive Multi-Agent Systems
Xiyuan Yang, Jiaru Zou, Rui Pan, Ruizhong Qiu, Pan Lu, Shizhe Diao, Jindong Jiang, Hanghang Tong, Tong Zhang, Markus J. Buehler, Jingrui He, James Zou
2026-04-29
Summary
This paper introduces a new way to make groups of AI agents work together more effectively by letting them repeatedly refine their thinking through a recursive process, similar to how some language models work. It's about scaling up collaboration, not just individual agent abilities.
What's the problem?
Current AI systems, even those with multiple agents, often struggle with complex tasks that require deep reasoning and collaboration. Simply adding more agents doesn't always lead to better results, and existing methods can be slow and inefficient. The challenge is to find a way to make agent collaboration itself more powerful and scalable.
What's the solution?
The researchers developed a framework called RecursiveMAS. Imagine a loop where agents share their 'thoughts' (represented as data) and build upon each other's ideas repeatedly. This loop is made possible by a 'RecursiveLink' module that allows for easy communication and transfer of information between agents. They also created a special training method that optimizes the entire system at once, ensuring everyone learns to work together effectively. This system is designed to be faster and use fewer resources than traditional multi-agent systems.
Why it matters?
This work is important because it shows a promising path towards building more capable and efficient AI systems that can tackle complex problems. By scaling up collaboration through recursion, they've achieved significant improvements in accuracy, speed, and resource usage across a variety of tasks, including math, science, and coding. This could lead to AI that's better at problem-solving and more accessible due to reduced computational costs.
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 1.2times-2.4times end-to-end inference speedup, and 34.6%-75.6% token usage reduction. Code and Data are provided in https://recursivemas.github.io.