Is SpecTr superseded?

SpecTr (Speculative decoding): superseded — cited as a baseline and beaten by newer methods. 3 paper(s) critique it, 2 beat it on benchmarks — #18 of 151 most-superseded. Sub-problem: cluster led by SpecInfer. Newer alternatives in the same sub-problem include SpecKV, component-aware self-speculative decoding, FASER, ConfLayers, Goose.

Method Drift›Speculative decoding

Superseded baseline#18 of 151 most-superseded

SpecTr

SpecTr: Spectral Transformer for Hyperspectral Pathology Image Segmentation

Speculative decoding · first seen Mar 5, 2021

superseded — cited as a baseline and beaten by newer methods

3 papers critique it · 2 beat it on benchmarks

What papers say

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

“However, a common limitation of these methods is their reliance on fixed patterns of tree construction, which can lead to suboptimal performance across diverse query distributions, resulting in a relatively low acceptance rate as tree size grows.”
— DySpec: Faster Speculative Decoding with Dynamic Token Tree Structure
“However, sun2403block proved that the position-by-position verification procedure does not yield the optimal expected number of accepted tokens (see Lemma 1 in sun2403block).”
— SpecTr-GBV: Multi-Draft Block Verification Accelerating Speculative Decoding
“However due to complexity reasons, the authors instead propose a modified sequential rejection sampling scheme which has much lower complexity.”
— Multi-Draft Speculative Sampling: Canonical Decomposition and Theoretical Limits

Beaten on benchmarks

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

GBV beats SpecTr · Block Efficiency [Temp 1.0, K=2]
4.430 vs 3.716
Greedy Multi-Path Block Verification for Faster Decoding in Speculative Sampling
GBV beats SpecTr · Throughput [Temp 1.0, K=2]
7.508 vs 6.355
Greedy Multi-Path Block Verification for Faster Decoding in Speculative Sampling
GBV beats SpecTr · Walltime [Temp 1.0, K=2]
133.188 vs 157.365
Greedy Multi-Path Block Verification for Faster Decoding in Speculative Sampling
GBV beats SpecTr · Block Efficiency [Temp 1.0, K=3]
4.271 vs 3.540
Greedy Multi-Path Block Verification for Faster Decoding in Speculative Sampling
GBV beats SpecTr · Throughput [Temp 1.0, K=3]
7.338 vs 6.211
Greedy Multi-Path Block Verification for Faster Decoding in Speculative Sampling
GBV beats SpecTr · Walltime [Temp 1.0, K=3]
136.272 vs 161.002
Greedy Multi-Path Block Verification for Faster Decoding in Speculative Sampling
Hybrid beats SpecTr · Block Efficiency [K=2 drafts, top-k]
2.21 vs 1.97
Multi-Draft Speculative Sampling: Canonical Decomposition and Theoretical Limits
Hybrid beats SpecTr · Token Rate [K=2 drafts, top-k]
14.78 vs 13.20
Multi-Draft Speculative Sampling: Canonical Decomposition and Theoretical Limits
Hybrid beats SpecTr · Block Efficiency [K=2 drafts, top-p]
2.09 vs 1.92
Multi-Draft Speculative Sampling: Canonical Decomposition and Theoretical Limits
Hybrid beats SpecTr · Token Rate [K=2 drafts, top-p]
14.50 vs 13.22
Multi-Draft Speculative Sampling: Canonical Decomposition and Theoretical Limits
Hybrid beats SpecTr · Block Efficiency [K=3 drafts, top-k]
2.24 vs 2.01
Multi-Draft Speculative Sampling: Canonical Decomposition and Theoretical Limits
Hybrid beats SpecTr · Token Rate [K=3 drafts, top-k]
14.67 vs 13.26
Multi-Draft Speculative Sampling: Canonical Decomposition and Theoretical Limits

Newer alternatives

Recent methods in the same sub-problem, not yet superseded in the knowledge base.