README.md

---
license: mit
pipeline_tag: text-generation
---

# Mixture-Outerlinks

[🌐 Project Page](https://recursivemas.github.io) | [💻 Code](https://github.com/RecursiveMAS/RecursiveMAS) | [📄 Paper](https://arxiv.org/abs/2604.25917)

We introduce RecursiveMAS, a multi-agent framework that scales agent collaboration through latent-space recursion.
RecursiveMAS treats a multi-agent system as a unified recursive computation, where heterogeneous agents iteratively exchange, refine, and evolve their latent states across recursion rounds. In the Mixture-Style setting, domain-specialized agents collaborate with the Summarizer Agent through Outer RecursiveLink modules for integrating specialized reasoning into the final response.

## Usage

This repository provides the **Outer RecursiveLink modules** for the **Mixture-Style** configuration of **RecursiveMAS**. You can load the full multi-agent system using the following snippet (requires the [GitHub repository](https://github.com/RecursiveMAS/RecursiveMAS) code):

```python
from system_loader import load_mas_system

# Load the full Mixture-Style MAS pipeline
mas = load_mas_system(
    style="mixture",
    device="cuda",
    trust_remote_code=True,
)

# Access individual agents and their RecursiveLink-connected components
math_agent = mas.agents["Math"].model
code_agent = mas.agents["Code"].model
science_agent = mas.agents["Science"].model
summarizer = mas.agents["Summarizer"].model
```

## Model Details

| Item | Description |
|---|---|
| Model | Mixture-Outerlinks |
| Collaboration Style | Mixture-Style |
| Component Role | Outer RecursiveLink Modules |
| Math-Summarizer-Outerlink.pt | Math Specialist Agent → Summarizer Agent |
| Summarizer-Math-Outerlink.pt | Summarizer Agent → Math Specialist Agent |
| Code-Summarizer-Outerlink.pt | Code Specialist Agent → Summarizer Agent |
| Summarizer-Code-Outerlink.pt | Summarizer Agent → Code Specialist Agent |
| Science-Summarizer-Outerlink.pt | Science Specialist Agent → Summarizer Agent |
| Summarizer-Science-Outerlink.pt | Summarizer Agent → Science Specialist Agent |

⚠️ **Note:** This checkpoint contains **Outer RecursiveLink modules** in [**RecursiveMAS**](https://arxiv.org/abs/2604.25917), rather than a standalone model intended for plain-text generation.  
For detailed usage instructions, please refer to our [GitHub repository](https://github.com/RecursiveMAS/RecursiveMAS).

## Model Collections for RecursiveMAS

| Style | Model Collection |
|---|---|
| Sequential-Style | [🤗 HuggingFace](https://huggingface.co/collections/RecursiveMAS/sequential-style-recursivemas) |
| Mixture-Style | [🤗 HuggingFace](https://huggingface.co/collections/RecursiveMAS/mixture-style-recursivemas) |
| Distillation-Style | [🤗 HuggingFace](https://huggingface.co/collections/RecursiveMAS/distillation-style-recursivemas) |
| Deliberation-Style | [🤗 HuggingFace](https://huggingface.co/collections/RecursiveMAS/deliberation-style-recursivemas) |

## Experiment Results

<p align="center">
  <img src="https://raw.githubusercontent.com/RecursiveMAS/RecursiveMAS/main/assets/hero_fig.png" width="95%" alt="RecursiveMAS Experiment Results">
</p>

## Citation

```bibtex
@misc{recursivemas,
      title={Recursive Multi-Agent Systems}, 
      author={Xiyuan Yang and Jiaru Zou and Rui Pan and Ruizhong Qiu and Pan Lu and Shizhe Diao and Jindong Jiang and Hanghang Tong and Tong Zhang and Markus J. Buehler and Jingrui He and James Zou},
      year={2026},
      eprint={2604.25917},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2604.25917}, 
}
```