Why it matters

EAGLE improves on Medusa. Understanding shapes selection.

Advertisement

The architecture

Draft: small model on target's feature embeddings.

Higher accept rate.

EAGLE structureTarget featureshidden statesEAGLE drafton featuresHigher accept ratevs token-onlyUses target model features not just tokens; better accept rate
EAGLE flow.
Advertisement

How it works end to end

Feature-level draft: uses target's rich features.

Better predictions than token-level draft.

EAGLE-2: further improvements.