LLaDA 1.5: Variance-Reduced Preference Optimization for Large Language Diffusion Models

Introduction

We introduce LLaDA 1.5, a competitive large diffusion language model, trained by variance-reduced preference optimization (VRPO).

Compared with LLaDA-8B-Instruct, LLaDA 1.5 achieves better performance on a wide range of tasks, including Math, Code, and Alignment tasks.

Inference

The LLaDA 1.5 model is available on Huggingface. Please employ the transformers to load.

from transformers import AutoModel, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained('GSAI-ML/LLaDA-1.5', trust_remote_code=True)
model = AutoModel.from_pretrained('GSAI-ML/LLaDA-1.5', trust_remote_code=True, torch_dtype=torch.bfloat16)

The model is based on LLaDA-8B-Instruct, you can use the code for LLaDA-8B-Instruct to inference.

Evaluation

We have open-sourced the evaluation code to reproduce the benchmark results of the LLaDA 1.5 model reported in the paper as closely as possible. Please refer to the documentation here.

Contact

If you have any questions, please feel free to contact fengqizhu@ruc.edu.cn.

Citation

Please consider cite:

@article{zhu2025llada,
  title={LLaDA 1.5: Variance-Reduced Preference Optimization for Large Language Diffusion Models},
  author={Zhu, Fengqi and Wang, Rongzhen and Nie, Shen and Zhang, Xiaolu and Wu, Chunwei and Hu, Jun and Zhou, Jun and Chen, Jianfei and Lin, Yankai and Wen, Ji-Rong and others},
  journal={arXiv preprint arXiv:2505.19223},
  year={2025}
}

LLaDA Group Wechat QR Code