LinQuant: A Quantization Framework for Low-Precision Linear Recurrences in Mamba-2 and Beyond
SonglinYang4 · x · 2026-11-06
Researchers introduce LinQuant, a framework for low-precision recurrent prefill targeting linear recurrences such as Mamba-2, GDN, and KDA.
Key insight: while attention quantization has been studied for years, the linear recurrences replacing attention lacked an equivalent treatment. The authors show their computation can be rewritten to reduce quantization error while preserving the full-precision function exactly.
More from Research
- Kling AI team and partners unveil JointLight, a SIGGRAPH Asia 2026 video relighting method — jiqizhixin · 2026-11-06
- Cornell's Trainspotting: 12k objects, 312k found photos to train feed-forward 3D reconstruction — kwangmoo_yi · 2026-11-06
- How a 118-base stretch of ultraconserved DNA became HAR1, the fastest-evolving region of the human genome — iskander · 2026-11-06
- ICLR reviews are out today, setting off the annual score discourse — khademinori · 2026-11-06
- Meta researcher: do as much SFT as possible before touching RL, or you're wasting money — DimitrisPapail · 2026-11-06
- SWE-bench author responds to benchmaxx concerns: SWE needs human-in-the-loop benchmarks — OfirPress · 2026-11-06