ReaDy Uses Recurrent States as Dynamic Context for Speculative Decoding in Hybrid Linear Attention LLMs
Bohan Sun · hf · 2026-10-19
A paper from StartLux, UCAS, and the University of Hong Kong introduces ReaDy, claimed to be the first speculative decoding method designed for hybrid linear attention (HLA) LLMs. Instead of relying on target-model hidden features whose emphasis may misalign with draft positions, ReaDy retrieves position-specific context from the compact recurrent states that HLA models maintain, using them as dynamic context for drafting.
This improves both generation speed and draft acceptance rates, easing the sequential bottleneck of autoregressive decoding. The paper was posted to arXiv on Oct 14.
More from Infra
- "Slowest pretrain I've ever supervised": researcher's one-line lament — cephaloform · 2026-10-19
- Sipeed's MaixStation challenges DGX Spark: 128GB LPDDR5X at 40W, one-third the size — AIFlow_ML · 2026-10-19
- TRELLIS.2 image-to-3D running on AMD RX 9060 XT with ROCm: fixes, workflows, MIT swaps — ova-mojica · 2026-10-19
- Celesto Claims CI Runner With 96% Lower Per-Minute Cost Than GitHub Actions — aniketmaurya · 2026-10-19
- Huawei's Atlas 960E superpod links 4,096 NPUs with near-packaged optics — bookwormengr · 2026-10-19
- Robot brain inference cost fell 84% in two months, from $2.36 to $0.38 per task — victor_explore · 2026-10-19