HuskyDoge
  • About
  • Research
  • Blogs
  • Misc
Xiaohongshu

Feedback Shortcut: Speculative Decoding for Latent Feedback Transformers

Posted: Nov 20th 2026
alphaXiv
Github
Tweet

TL;DR

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.

Feedback Shortcut: Speculative Decoding for Latent Feedback Transformers
View this thread on X