Large language models generate text one word at a time, and each word costs a full pass through the model, so output length drives latency. speeds this up by having a small helper model (a drafter) guess several upcoming words, which the full model then checks in a single pass. Drafters that guess word by word get slow as guesses lengthen; drafters that guess all positions at once ignore how words depend on each other.
ChainFlow makes the drafter start from the model's own recent internal state rather than noise, then nudges that estimate onto the model's predicted future state in one or two cheap steps. A compression step keeps this cost flat as the model grows, and small correction heads restore word-to-word dependence.
On Qwen3.5 models it runs 39 percent faster than one strong baseline, EAGLE-3, and 11 percent faster than another, DFlash, at equal drafter size. The idea matters because the drafter cost stops scaling with model size.