DFlare : Scaling Up Draft Capacity for Block Diffusion Speculative Decoding
Jiebin Zhang, Zhenghan Yu, Song Liu, Eugene J. Yu, Zheng Li, Dawei Zhu, [0.5ex] Jiangshan Duo, Weimin Xiong, Yifan Song, Guanghua Yu, Jianchen Zhu, Sujian Li, [1.5ex] School of Computer Science, Peking University · 2026
Read DFlare : Scaling Up Draft Capacity for Block Diffusion Speculative Decoding by Jiebin Zhang, Zhenghan Yu, Song Liu, Eugene J. Yu, Zheng Li, Dawei Zhu, [0.5ex] Jiangshan Duo, Weimin Xiong, Yifan Song, Guanghua Yu, Jianchen Zhu, Sujian Li, [1.5ex] School of Computer Science, Peking University (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-02091.txt.
Description
Zhang and co-authors scale up draft capacity for block diffusion decoding. Speculative decoding speeds up language model inference. A draft model predicts a whole block of tokens at once. The target model verifies them in parallel. A capable draft model is key to the speedup.
The prior best method, DFlash, shares one fused representation. All draft layers draw from a few target layers. That narrow bottleneck limits per-layer expressiveness. Draft capacity cannot scale well under that design. Their DFlare method flares the bottleneck out. Each draft layer attends to its own learnable mix. Broad target layers feed that mix at low cost. Richer target knowledge enters every draft layer. Distinct conditioning per layer is the result.
Catalog record
| CREATOR | Jiebin Zhang, Zhenghan Yu, Song Liu, Eugene J. Yu, Zheng Li, Dawei Zhu, [0.5ex] Jiangshan Duo, Weimin Xiong, Yifan Song, Guanghua Yu, Jianchen Zhu, Sujian Li, [1.5ex] School of Computer Science, Peking University |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 40 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2606.02091