Unlocking Efficiency in LLM Inference: A Comprehensive Survey of Speculative Decoding (ACL 2024)

Fields

id

sources/ev-speculative-decoding-survey

URL

https://arxiv.org/abs/2401.07851

Publisher

Heming Xia et al. (Peking University)

Kind

paper

Published

2024-01

Accessed

2026-10-04

notes

SURVEY PAPER - useful for understanding the landscape of speculative decoding methods and their claimed properties. The survey covers: (1) Formal definition of speculative decoding. (2) Drafter selection strategies. (3) Verification strategies. (4) 'Comparative analysis of leading methods under third-party testing environments' - this is a rare attempt at independent comparison. The survey is valuable because it is NOT authored by any single method's creators, providing a more neutral overview. However, the survey was published in January 2024 and does not cover more recent developments like EAGLE-2, MTP heads, or lossy verification methods. The survey's 'third-party testing' claim should be verified - it may still rely on numbers reported by original method authors.

Notes

SURVEY PAPER - useful for understanding the landscape of speculative decoding methods and their claimed properties. The survey covers: (1) Formal definition of speculative decoding. (2) Drafter selection strategies. (3) Verification strategies. (4) 'Comparative analysis of leading methods under third-party testing environments' - this is a rare attempt at independent comparison. The survey is valuable because it is NOT authored by any single method's creators, providing a more neutral overview. However, the survey was published in January 2024 and does not cover more recent developments like EAGLE-2, MTP heads, or lossy verification methods. The survey's 'third-party testing' claim should be verified - it may still rely on numbers reported by original method authors.

Cross-references

This record references no other records.