A decoding strategy where a fast drafter model generates candidate tokens and a verifier checks them, reducing total inference time compared to standard generation.