Field Notes

Paper Notes: EAGLE-3

last updated 2026-06-15

• speculative-decoding

EAGLE-3 core improvements over EAGLE-2:

1. the model gets better with more data, not suffering from saturation nearly as strongly; this is because:
2. the model uses features from a lower, middle, and upper layer of the base LLM, as opposed to just the final layer; all of this leads to:
3. higher average acceptance length (AAL), and thus larger speedups.