Our work introduces Saber, a training-free sampling algorithm for diffusion language models that enhances code generation by adaptively accelerating inference and incorporating backtracking, thereby improving output quality and speed while narrowing the performance gap with autoregressive models.

conda create -n saber python=3.11
pip install -r requirements.txt

or

uv venv saber --python 3.11
source saber/bin/activate
uv pip install -r requirements.txt

Download the LLaDA-8B-Instruct model and place it in the models/ directory. You can download the model from Hugging Face.

We provide the HumanEval, MBPP, HumanEval-ET, and MBPP-ET datasets in the data/ directory. For LiveCodeBench, you will need to download it separately and place it in data/livecodebench/.

Firstly, ensure the model is placed correctly in the models/ directory (the configs use the path ./models/LLaDA-8B-Instruct by default). Secondly, verify the datasets are in the correct path. Finally, execute the following command to evaluate:

python eval.py --config ./configs/humaneval.yaml

If you want to test other methods, change the method in the yaml file. For the humaneval and MBPP datasets, our code will print pass@1 And steps. For the livecodebench dataset, our code will save the generated results, and you need to run the evaluation program yourself

@article{dong2025saber,
  title={Saber: An Efficient Sampling with Adaptive Acceleration and Backtracking Enhanced Remasking for Diffusion Language Model},
  author={Yihong Dong and Zhaoyu Ma and Xue Jiang and Zhiyuan Fan and Jiaru Qian and Yongmin Li and Jianha Xiao and Zhi Jin and Rongyu Cao and Binhua Li and Fei Huang and Yongbin Li and Ge Li},
  journal={arXiv preprint arXiv:2510.18165},
  year={2025}
}