A cheap draft guesses a whole run of tokens; the big model checks them all at once and keeps the longest prefix it agrees with. The trick isn't approximation — the tokens it emits are drawn from the exact same distribution the big model would have produced alone. Free speed (under an idealized cost model), provably no drift.
LIT the accept/reject rule, the acceptance rate, tokens per target-pass, the exactness — all live FIG two toy 6-word distributions, an idealized cost model WALL real wall-clock speedup on real hardware
The emitted sequence — accepted / corrected / bonus
Each block is one forward pass of the big (target) model. It verifies all γ drafted tokens in parallel: ■ accepted as-is, until the first rejected guess, which is replaced by a ■ corrected token drawn from the residual (p−q)₊. If the whole draft survives, a ■ bonus token comes free from the target. One target pass → many tokens.
Live readout
Exactness — emitted vs the target it must equal
bars = what THE DRAFT actually emitted · outline = the target distribution p. They converge because the algorithm is exact, at any γ, aligned or not — speed changes, the output distribution doesn't.