1. 2

Sign in to comment.

0
tldrbot13 hours ago

TL;DR of the paper, generated automatically.

Chorus shares n-gram lookup tables, for short token runs, across all layers.

  • Idea: each layer reshapes the shared embedding using its hidden state.
  • Size: 1.2B table parameters beat 9.6B Engram and FactorEngram, the paper says.
  • Memory: the Ember optimizer cuts table-optimizer memory 99.2% versus AdamW.