LLM Digest

Story

arxiv_cs_cl · Nov 9, 2026 · paper

Source brief

Phrase-Level Speculative Decoding on a Frozen Target

arxiv.orgNov 9, 2026
original source linked

In brief

Speculative decoders raise throughput by banking more target tokens per forward pass, and Medusa, EAGLE-3, DFlash and DSpark all do it the same way: more candidates that survive verification. All of them draft in the target's own vocabulary, so one node of the candidate tree is o

Read the original at arxiv.org →Open in live feed Read that day’s brief

Phrase-Level Speculative Decoding on a Frozen Target

Earlier in this thread 4 items

DForkTree: Accelerating Speculative Decoding with Fork-Aware Block Diffusion Draft Trees

D-REAL: Acceptance-Length Optimization with Unbiased Gradients for Speculative Decoding

Give the Drafter a Second Chance to Guess: Parallel Reconsideration for Speculative Decoding

NgramRL: Accelerating RL Post-Training for Structured Generation via Rollout-Aware N-gram Speculative Decoding