HuskyDoge
  • About
  • Research
  • Blogs
  • Misc
Xiaohongshu

#Speculative Decoding

Research

Feedback Shortcut: Speculative Decoding for Latent Feedback Transformers
Benhao Huang, Xuezhe Ma, Hector Liu
Nov 20th 2026
Institute of Foundation Models
#Latent Feedback
#Speculative Decoding
#Efficient Inference

Latent feedback breaks speculative decoding. Feedback Shortcut fixes it with MTP stand-ins and pipelined correction: 2.4–2.8× faster decoding, within 10⁻⁴ bits per byte of exact decoding.

alphaXiv
Github
Tweet
Feedback Shortcut: Speculative Decoding for Latent Feedback Transformers