Scholar Feed

Speculative decoding

PLD

PLD+: Accelerating LLM inference by leveraging Language Model Artifacts

Superseded baseline#6 of 151 most-superseded · first seen Dec 2, 2024

Superseded — cited as a baseline and beaten by newer methods

7 papers critique it · 13 beat it on benchmarks

What papers say

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

existing model-free approaches, such as prompt-lookup decoding (PLD)~saxena2023prompt, achieve low overhead and rapid token generation, but typically lack adaptivity.
— SuffixDecoding: Extreme Speculative Decoding for Emerging AI Applications
“there are no matched tokens in more than 30% of decoding steps in PLD”
— LogitSpec: Accelerating Retrieval-based Speculative Decoding via Next Next Token Speculation
“These training-free approaches are highly effective for tasks with high repetition (e.g., code editing) but struggle with open-ended generation where context reuse is minimal.”
— Quasar: Quantized Self-Speculative Acceleration for Rapid Inference via Memory-Efficient Verification
“pattern-repeating scenarios such as code generation, but can only propose a single continuation at a time. Moreover, because pattern matches are sparse and fail to capture the full diversity of target model outputs, PLD is constrained to specific domains and cannot generalize broadly.”
— RACER: Retrieval-Augmented Contextual Rapid Speculative Decoding
“Prompt- and retrieval-based approaches (PLD, Lookahead, CLLMs) improve draft quality but degrade with scarce context”
— Bridging Draft Policy Misalignment: Group Tree Optimization for Speculative Decoding
“PLD~pld-saxena-2023 focuses on current text while REST~rest-he-2024 uses a text corpus.”
— SAM Decoding: Speculative Decoding via Suffix Automaton
“However, it cannot predict new tokens or their combinations.”
— RASD: Retrieval-Augmented Speculative Decoding

Beaten on benchmarks

Head-to-head results where a newer method reports beating PLD. 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Recent methods in the same sub-problem, not yet superseded in the knowledge base — arXiv benchmark leaders, not vetted production recommendations.