1. 2

Sign in to comment.

0
tldrbot1 hour ago

TL;DR of the paper, generated automatically.

Length alone doesn't break retrieval. Rival statements do.

  • Filler: neutral filler up to 32K tokens costs nothing, accuracy stays 0.97 to 1.00.
  • Rivals: evidence-style statements drop accuracy to 0.16 to 0.33.
  • Attention: attention left on the evidence explains up to 58% of the loss. Length: 8%.