Scholar Feed

Speculative decoding

FR-Spec

FR-Spec: Accelerating Large-Vocabulary Language Models via Frequency-Ranked Speculative Sampling

Superseded baseline#12 of 151 most-superseded · first seen Feb 20, 2025

Superseded — cited as a baseline and beaten by newer methods

4 papers critique it · 5 beat it on benchmarks

What papers say

Verbatim critique sentences, each from a paper that cites FR-Spec as a baseline.

static methods rely on the assumption of static word frequencies and fail to capture long-tail tokens that become locally probable in specialized domains or during topic shifts
— EvoSpec: Evolving Speculative Decoding via Real-Time Vocabulary and Parameter AdaptationTarget
“While effective in reducing latency, these static subsets may suppress rare or domain-specific tokens, lowering acceptance in speculative decoding.”
— DynaSpec: Context-aware Dynamic Speculative Sampling for Large-Vocabulary Language Models
“are context-insensitive and struggle with long-tail tokens, leading to lower acceptance rates in diverse scenarios.”
— MicroSpec: Accelerating Speculative Decoding with Lightweight In-Context Vocabularies
“All tokens outside this vocabulary are assigned zero probability and can never be proposed by the drafter, which typically reduces acceptance quality.”
— SlimSpec: Low-Rank Draft LM-Head for Accelerated Speculative Decoding

Beaten on benchmarks

Head-to-head results where a newer method reports beating FR-Spec. Values are copied from the source paper's tables — verify against the cited paper.

What to use instead

Recent methods in the same sub-problem, not yet superseded in the knowledge base — arXiv benchmark leaders, not vetted production recommendations.