Tencent's DSpine speculative decoding paper earns rare high praise from toughest reviewer
teortaxesTex · x · 2026-11-09
Tencent's DSpine speculative decoding paper is drawing unusually strong praise in the AI community. TeortaxesTex notes that Arthur Chakhvadze, known for "Eastern European standards" (i.e., brutally strict reviews), called it cool after finally working through it, and specifically liked its mechinterp-inspired design.
Technical highlights from the paper figures:
- A unified transfer space built from the target LM head's output embeddings: ℓ2-normalize rows, apply PCA whitening, keep r principal directions, and normalize again into a fixed embedding table C — a single pathway carries either a prediction or a decided token.
- Adjacent causal injection: after each layer, position t receives a message only from its immediate predecessor; receiver and predecessor features jointly gate (sigmoid) how much of the message is written into the residual stream. Messages use pre-update features, so injection runs in parallel across positions; the anchor stays unchanged.
- After the final layer, transition-cached decoding selects each token conditioned on its chosen predecessor.
Worth a close read for anyone interested in speculative decoding.
More from Research
- TACL caps first authors at one submission every three months amid LLM-written paper flood — EhudReiter · 2026-11-09
- Exact Relative-Position Attention Kernels Deliver Up to 21x Backward Speedup at 8K Context — INSAIT-Institute · 2026-11-09
- HKUST-Led Benchmark Shows LLM Agents Drift Beyond Authorization Even Without Attackers — Chuxue Cao · 2026-11-09
- S-Agent 2.0 gives coding agents spatial investigation tools, lifting ReVSI by 20.5 points — jiqizhixin · 2026-11-09
- Papers with Code weekly trending: Inkling 2 Small and Drex top the charts — NielsRogge · 2026-11-09
- Opus 5.5 agents dig through 12,000 pages of Gauss and Riemann notebooks, find lost physics — roydanroy · 2026-11-09