| license: mit | |
| library_name: transformers | |
| pipeline_tag: text-generation | |
| # Multi-Block Diffusion Language Models (MBD-LMs) | |
| Model weights of [MBD-LMs](https://arxiv.org/abs/2606.29215). | |
| - **Project Page:** [sjtu-deng-lab.github.io/mbd-lms](https://sjtu-deng-lab.github.io/mbd-lms) | |
| - **GitHub Repository (Training & Configs):** [SJTU-DENG-Lab/mbd-lms](https://github.com/SJTU-DENG-Lab/mbd-lms) | |
| - **Inference Engine (Diffulex):** [SJTU-DENG-Lab/Diffulex](https://github.com/SJTU-DENG-Lab/Diffulex) | |
| ## Introduction | |
| Block Diffusion Language Models (BD-LMs) improve diffusion-based text generation with KV caching and flexible-length generation. Multi-Block Diffusion Language Models (MBD-LMs) extend this from Single-Block Diffusion (SingleBD) to Multi-Block Diffusion (MultiBD), where a running-set of consecutive blocks is decoded concurrently for inter-block parallelism. | |
| MBD-LMs are obtained by post-training BD-LMs with **Multi-block Teacher Forcing (MultiTF)**, which integrates teacher forcing and diffusion forcing by training on bounded noise-groups conditioned on clean prefixes, matching practical MultiBD inference states. | |
| ## Citation | |
| If you find this work useful, please cite: | |
| ```bibtex | |
| @article{jin2026multiblock, | |
| title={Multi-Block Diffusion Language Models}, | |
| author={Jin, Yijie and Xu, Jiajun and Liu, Yuxuan and Xu, Chenkai and Tu, Yi and Li, Jiajun and Tu, Dandan and Yan, Xiaohui and Yu, Kai and Liu, Pengfei and Deng, Zhijie}, | |
| journal={arXiv preprint arXiv:2606.29215}, | |
| year={2026} | |
| } | |
| ``` |