base_model:
- Qwen/Qwen3.5-4B
license: mit
library_name: transformers
pipeline_tag: text-generation
Deliberation-Toolcaller-Qwen3.5-4B
🌐 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 Tool-Caller Agent performs tool-oriented reasoning and execution, while collaborating with the Reflector Agent through RecursiveLink modules for iterative refinement.
Model Details
| Item | Description |
|---|---|
| Model | Deliberation-Toolcaller-Qwen3.5-4B |
| Collaboration Style | Deliberation-Style |
| Agent Role | Tool-Caller Agent |
| Base Model | Qwen3.5-4B |
⚠️ Note: This checkpoint is a role-specific agent in RecursiveMAS, rather than a standalone model intended for plain-text generation.
For detailed usage instructions, please refer to our GitHub repository.
Usage
This model is intended to be used as part of the RecursiveMAS framework. You can load the deliberation system as follows:
from system_loader import load_mas_system
mas = load_mas_system(
style="deliberation",
device="cuda",
trust_remote_code=True,
)
# Access the specific agents
reflector = mas.agents["reflector"].model
toolcaller = mas.agents["toolcaller"].model
Alternatively, you can run inference using the provided script from the repository:
python run.py --style deliberation --batch_size 16 --temperature 0.6 --top_p 0.95 --dataset math500 --seed 42 --trust_remote_code 1 --device cuda
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},
}