Tag
GAVEL introduces a new task for verifying, explaining, and localizing errors in image-text pairs, along with a dataset and benchmark. A supervised baseline shows improvements over strong closed-source models.