DiffuSpec: Unlocking Diffusion Language Models for Speculative Decoding

Method and contribution

Uses diffusion drafting, Causal-consistency Path Search over token lattices, and adaptive draft lengths.

Connects candidate-lattice path selection to autoregressive verification of parallel diffusion proposals.

Read in context

Structured Decoding: Higher-Order CRFs, Diffusion Language Models, and Candidate Lattices →

Citation

Guanghao Li, Zhihui Fu, Min Fang, Qibin Zhao, Ming Tang, Chun Yuan, Jun Wang. DiffuSpec: Unlocking Diffusion Language Models for Speculative Decoding. Findings of ACL 2026.

@inproceedings{li2026diffuspec,
  title = {{DiffuSpec: Unlocking Diffusion Language Models for Speculative Decoding}},
  author = {Li, Guanghao and Fu, Zhihui and Fang, Min and Zhao, Qibin and Tang, Ming and Yuan, Chun and Wang, Jun},
  year = {2026},
  booktitle = {Findings of ACL 2026},
  url = {https://aclanthology.org/2026.findings-acl.1048/},
  doi = {10.18653/v1/2026.findings-acl.1048}
}