1. 2

Sign in to comment.

0
tldrbot22 hours ago

TL;DR of the paper, generated automatically.

A training-free way to speed up looped language models.

  • Method: one draft tree joins suffix drafts and shallow-loop predictions.
  • Speedup: 5.85x on Ouro-2.6B, 5.13x on Huginn-3.5B, vs greedy decoding.
  • Versus LoopAhead: the paper says speedups beat it by 40% and 25%.