SLED (Speculative decoding): cited as a baseline — critiqued by newer work, not yet beaten on a benchmark here. 2 paper(s) critique it, 0 beat it on benchmarks — #71 of 151 most-superseded. Sub-problem: cluster led by EdgeLLM. Newer alternatives in the same sub-problem include PipeSD, DiP-SD, PicoSpec.

Speculative decoding

SLED

SLED: A Speculative LLM Decoding Framework for Efficient Edge Serving

Superseded baseline#71 of 151 most-superseded · first seen Jun 11, 2025

Cited as a baseline — critiqued by newer work, not yet beaten on a benchmark here

2 papers critique it · 0 beat it on benchmarks

What papers say

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

existing edge SD systems typically adopt simple batching policies, e.g., static batching in SLED

— DiP-SD: Distributed Pipelined Speculative Decoding for Efficient LLM Inference at the Edge

“Similarly, SLED li2025sled focuses on multi-client throughput but fails to provide deep latency masking for individual users.”

— A Pipelined Collaborative Speculative Decoding Framework for Efficient Edge-Cloud LLM Inference

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Recent methods in the same sub-problem, not yet superseded in the knowledge base — arXiv benchmark leaders, not vetted production recommendations.

PipeSDPipeSD: An Efficient Cloud-Edge Collaborative Pipeline Inference Framework with Speculative Decoding
May 13, 2026
DiP-SDDiP-SD: Distributed Pipelined Speculative Decoding for Efficient LLM Inference at the Edge
Apr 22, 2026
PicoSpecA Pipelined Collaborative Speculative Decoding Framework for Efficient Edge-Cloud LLM Inference
Mar 19, 2026