Introduction

We introduce ReFusion, a novel masked diffusion model featuring two core innovations:

1. It unifies a causal attention mechanism with global, any-order slot generation, enabling full KV cache reuse without sacrificing flexibility.
2. It simplifies the learning objective from an intractable token-combination space to a manageable slot-permutation space, significantly boosting learning efficiency.

Empirically, ReFusion not only outperforms prior MDMs with a 34% performance gain and an over 18× speedup on average, but also bridges the performance gap to strong ARMs while maintaining a 2.33× average speedup.

Usage

Environment Setup

git clone https://github.com/ML-GSAI/ReFusion.git
cd ReFusion
conda env create -f refusion_full_env.yml
conda activate refusion_py10

Training

1. Data Preparation

We provide a sample dataset in data/train_data.json to illustrate the required format. The full training dataset is available on Hugging Face at GSAI-ML/ReFusion.

Please ensure your training data is organized as a JSON list of objects, where each object contains a query and a response.

Data Format Example:

[
    {
        "query": "...",
        "response": "..."
    },
    {
        "query": "...",
        "response": "..."
    }
]

2. Running the Training Script

Single-Node Training:

To train on a single machine, simply run:

bash train.sh

Multi-Node Training:

For distributed training across multiple nodes (e.g., 2 nodes), specify the node count (-n), current rank (-r), and the master node IP address (-m):

# Example: Running on the master node (Rank 0)
bash train.sh -n 2 -r 0 -m 192.168.1.1

# Example: Running on the worker node (Rank 1)
bash train.sh -n 2 -r 1 -m 192.168.1.1

Inference

python generate.py

Evaluation

bash eval.sh

Experimental Results

Key Results:

• Superior Performance: ReFusion achieves a 34% performance gain over prior MDMs.
• High Efficiency: It provides an over 18× speedup compared to MDMs and a 2.33× speedup compared to strong ARMs.
• Gap Bridging: ReFusion effectively bridges the performance gap to strong ARMs while maintaining significantly faster inference speeds.

Citation

If you find our work helpful, please consider citing our paper.

@misc{li2025refusiondiffusionlargelanguage,
      title={ReFusion: A Diffusion Large Language Model with Parallel Autoregressive Decoding}, 
      author={Jia-Nan Li and Jian Guan and Wei Wu and Chongxuan Li},
      year={2025},
      eprint={2512.13586},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2512.13586}, 
}