Weaver speculative decoding hits 4.08x speedup; authors cry foul as Tencent's DSpine skips their citation
teortaxesTex · x · 2026-11-09
The thread mixes a citation grievance with a real technical contribution: the authors complain that follow-up speculative-decoding papers (Tencent's DSpine, DSpark, DFlash2) cite everyone but them, despite Weaver predating several of them — suspecting the phone-friendly format makes the paper look less serious.
Paper: "Trees from Marginals: Autoregressive drafting with factorized priors"
- Factorized drafters predict future-token marginals in parallel and are cheap, but their independence assumption craters acceptance rates as the speculative budget grows
- Weaver is a lightweight autoregressive adapter that builds proposal trees from the top-K marginals, restoring conditional dependencies without a full-vocabulary projection
- Includes a rollback-free tree-verification algorithm for Gated Delta Net layers plus optimized CUDA kernels in SGLang
- Achieves 4.08x speedup over autoregressive decoding and beats a highly optimized DFlash baseline by 16.3%
More from Research
- Owain Evans proposes a practitioner guide to subliminal learning and phantom transfer — OwainEvans_UK · 2026-11-09
- Owain Evans proposes a research direction: data poisoning risks during RSI — OwainEvans_UK · 2026-11-09
- Owain Evans flags open research question: can reward hacking emerge without ever being rewarded — OwainEvans_UK · 2026-11-09
- Report author admits transparency failure on AI R&D acceleration section amid METR critiques — ShakeelHashim · 2026-11-09
- Philosopher argues the brain-computer metaphor is AI's fatal scientific error — AnnaCiaunica · 2026-11-09
- Doesn't Contradict the Bitter Lesson: Soft Inductive Biases Plus Scaling Win — irinarish · 2026-11-09