SJTU's DLite Enables History-Free Speculative Decoding with up to 70% Speedup

SJTU, together with Tsinghua and Infinigence-AI, introduced DLite, a history-free speculative decoding method that tackles acceptance-length degradation of draft models under agentic workloads, achieving up to 70% speedup over state-of-the-art approaches.

2026-10-19 ~ 2026-10-20 · 2 related posts

• DLite: History-Free Speculative Decoding Fixes 80% Acceptance Drop in Agentic AI — SJTU · 2026-10-19
• DLite: History-Free Speculative Decoding Beats DSpark by Up to 70% for Agentic AI — pmttyji · 2026-10-20