Is Speculative Sampling superseded?

Speculative Sampling (Speculative decoding): superseded — cited as a baseline and beaten by newer methods. 2 paper(s) critique it, 10 beat it on benchmarks — #9 of 151 most-superseded. Sub-problem: cluster led by Lookahead. Newer alternatives in the same sub-problem include Collaborative Speculative Decoding (CoSpec), ToolSpec, Quasar, FLy (Training-Free Loosely Speculative Decoding), Pivot-Aware Speculative Decoding.

Speculative decoding

Speculative Sampling

Speculative Sampling for Parametric Temporal Point Processes

Superseded baseline#9 of 151 most-superseded · first seen Oct 22, 2025

Superseded — cited as a baseline and beaten by newer methods

2 papers critique it · 10 beat it on benchmarks

What papers say

Verbatim critique sentences, each from a paper that cites Speculative Sampling as a baseline.

vanilla speculative decoding often suffers from high drafting latency, which can account for over 75\% and 60\% of the total inference time when using Qwen3-1.7B to accelerate Qwen3-14B and Qwen3-32B with draft length 5.

— DART: Diffusion-Inspired Speculative Decoding for Fast LLM Inference

“The draft model with limited capacity struggles to precisely approximate the large-scale target model.”

— EAGLE-3: Scaling up Inference Acceleration of Large Language Models via Training-Time Test

Beaten on benchmarks

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

Spiffy beats Speculative Sampling
3.07 vs 1
Speedup · [LLaDA-Base-8B HumanEval]
Spiffy: Multiplying Diffusion LLM Acceleration via Lossless Speculative Decoding
ECHO beats Speculative Sampling
5.25 vs 1.76
Avg. Speedup · [Vicuna-13B]
ECHO: Elastic Speculative Decoding with Sparse Gating for High-Concurrency Scenarios
DART beats Speculative Sampling
2.77 vs 0.98
Speedup · [Qwen14B Temperature=0]
DART: Diffusion-Inspired Speculative Decoding for Fast LLM Inference
LogitSpec beats Speculative Sampling
3.68 vs 1.87
Speedup · [Llama 2 7B]
LogitSpec: Accelerating Retrieval-based Speculative Decoding via Next Next Token Speculation
FLy beats Speculative Sampling
5.34 vs 2.82
Speedup · [L31 405B, Temperature=0]
Training-Free Loosely Speculative Decoding: Accepting Semantically Correct Drafts Beyond Exact Match
DiffuSpec beats Speculative Sampling
3.08 vs 1.67
Speedup · [training-free block, Qwen2.5-32B target]
DiffuSpec: Unlocking Diffusion Language Models for Speculative Decoding
DuoDec beats Speculative Sampling
2.61 vs 1.80
speedup (φ) · [Llama2-7b]
DuoDecoding: Hardware-aware Heterogeneous Speculative Decoding with Dynamic Multi-Sequence Drafting
+ Reflect Verify beats Speculative Sampling
1.36 vs 1.18
Speed · [Llama3.2-1B-Instruct & Llama3.1-8B-Instruct]
Think Before You Accept: Semantic Reflective Verification for Faster Speculative Decoding
CoSpec beats Speculative Sampling
86.97 vs 83.98
Mean · [Qwen3-32B, T=1]
Beyond the Target: From Imitation to Collaboration in Speculative Decoding

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.