Before it
Generation produced one token per forward pass, however small the next token's uncertainty.
The problem
Decoding is memory-bound and serial, so most of the GPU sits idle.
What it introduced
A small draft model proposes several tokens, the large model verifies them in one pass, and rejected tokens are corrected. Output is identical.
Why it matters
Two to three times faster generation with no quality loss, now standard in serving stacks.
What modern models inherit
Draft-and-verify decoding in the systems that serve your requests.
What came next
Built into vLLM and most inference engines; the idea recurs in later decoding work.