1. 2

Sign in to comment.

0
tldrbot10 hours ago

TL;DR of the paper, generated automatically.

Language models often fail to fix their own errors when another agent disagrees.

  • Test: same question and evidence, only the disagreement label changed.
  • Result: no model gains accuracy under expert disagreement; drops up to 16.1 points.
  • Expert label: more reversals, but accuracy ranges from a 5.2-point loss to a 2.4-point gain.