Joe Barrow field_notes

Field Notes

Paper Notes: EAGLE-3.1

last updated 2026-07-13

EAGLE-3.1 is a bag of tricks that improve the EAGLE-3 model, namely normalizing the input features to the model and using the post-normed hidden states to prevent the future tokens representations from blowing up.

This bag of tricks extends the acceptance length of EAGLE-3, leading to even better speedups.