license: mit
pipeline_tag: text-generation
Deliberation-Outerlinks
🌐 Project Page | 💻 Code | 📄 Paper
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 Deliberation-Style setting, the Reflector Agent and Tool-Caller Agent collaborate through Outer RecursiveLink modules for iterative reasoning, tool-oriented execution, and refinement.
Model Details
| Item | Description |
|---|---|
| Model | Deliberation-Outerlinks |
| Collaboration Style | Deliberation-Style |
| Component Role | Outer RecursiveLink Modules |
| Reflector-Toolcaller-Outerlink.pt | Reflector Agent → Tool-Caller Agent |
| Toolcaller-Reflector-Outerlink.pt | Tool-Caller Agent → Reflector Agent |
⚠️ Note: This checkpoint contains Outer RecursiveLink modules for RecursiveMAS, rather than a standalone model intended for plain-text generation.
Sample Usage
To load the full Deliberation-Style system using these modules, you can use the high-level API provided in the official repository:
from system_loader import load_mas_system
# Load the deliberation style multi-agent system
mas = load_mas_system(
style="deliberation",
device="cuda",
trust_remote_code=True,
)
reflector = mas.agents["reflector"].model
toolcaller = mas.agents["toolcaller"].model
For detailed execution commands on downstream tasks, please refer to the GitHub repository.
Model Collections for RecursiveMAS
| Style | Model Collection |
|---|---|
| Sequential-Style | 🤗 HuggingFace |
| Mixture-Style | 🤗 HuggingFace |
| Distillation-Style | 🤗 HuggingFace |
| Deliberation-Style | 🤗 HuggingFace |
Experiment Results
Citation
@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},
}