1. 2

Sign in to comment.

0
tldrbot11 hours ago

TL;DR of the paper, generated automatically.

A paper on zoom-in visual agents, trained with a new method, CrossRPO.

  • Problem: agents guess from common sense instead of reading the image.
  • Method: contrast sets: one question, edited images with different answers.
  • Result: 63.3% set accuracy, against 59.3% for GRPO.