Match each advanced speculative-decoding variant to its defining mechanism

Autoregressive decoding is slow because it is sequential: the model must finish token T before it can start token T+1. Each step reads the whole model from memory to produce just one token, so the GPU sits underused. **Speculative decoding** breaks that bottleneck. The idea is to guess several future tokens cheaply, then have the big target model verify all the guesses in a single parallel forward pass. Verifying many tokens costs roughly the same as decoding one, because decode is memory-bound. Correct guesses are kept; wrong ones are discarded. The four variants differ only in WHO produces the guesses. Vanilla speculative decoding uses a separate small draft model that generates tokens autoregressively. Medusa skips the draft model and trains extra prediction heads onto the target itself, so one pass emits several candidate tokens. EAGLE goes deeper: instead of predicting tokens, it predicts the target's future hidden states, which gives more accurate guesses and a higher acceptance rate. Lookahead decoding needs no training and no draft model at all. It runs a fixed-point iteration over n-grams seen during generation. The shared payoff is identical: spend extra parallel compute each round to shrink the slow sequential steps.

All four techniques attack the same bottleneck. Autoregressive decode is sequential and memory-bound, so each step wastes compute. They guess multiple tokens, then verify in parallel. The match-ups turn on WHO drafts. **Vanilla speculative decoding** uses a separate small draft model that proposes tokens autoregressively. The large target then verifies the whole block in one pass. Medusa removes the draft model. It trains extra decoding heads on top of the target. Each head predicts a future position, so one forward pass emits several candidate tokens in parallel. EAGLE drafts at the feature level. It predicts the target's next hidden states, not just tokens. Richer signal means higher acceptance, so more drafted tokens survive verification per round. Lookahead decoding needs no draft model and no training. It runs Jacobi fixed-point iteration over an n-gram pool collected during generation. Verified n-grams advance the sequence. The common payoff axis ties them together. Each trades extra parallel compute per round for fewer sequential decode steps. None changes the output distribution when verification is exact.