PosS improves speculative decoding by training a position-specialized draft model that generates higher-quality drafts, leading to improved efficiency and acceptance rates during decoding.
The full implementation, training details, and evaluation scripts are available at:
If the model is not automatically downloaded by your framework, you may manually download the following files from this repository:
1@misc{huang2025posspositionspecialistgenerates,
2 title = {POSS: Position Specialist Generates Better Draft for Speculative Decoding},
3 author = {Langlin Huang and Chengsong Huang and Jixuan Leng and Di Huang and Jiaxin Huang},
4 year = {2025},
5 eprint = {2506.03566},
6 archivePrefix= {arXiv},
7 primaryClass = {cs.CL},
8 url = {https://arxiv.org/abs/2506.03566}
9}
10
