Survey on the Evolution of Speculative Decoding

青稞AI · wechat · 2026-07-10

This article systematically reviews the evolution of Speculative Decoding, summarizing core insights, architectural innovations, and practical results from papers including Medusa, EAGLE-1/2/3, DFlash, DSpark, and JetSpec.

It highlights two main challenges speculative decoding aims to solve: the sequential and hard-to-parallelize nature of autoregressive decoding, and the difficulties in training and deploying draft models. The author also compares various methods regarding their trade-offs in prediction accuracy, position dependency, static/dynamic trees, and the balance between parallelism and quality.

Original post →

More from Research

Research channel →