SQUINT Teaches Vision Models to Read Hidden Text Humans See Easily

Chinese researchers' SQUINT framework lifted hidden-text recognition from 26.50 to 88.61 on SquintBench using knowledge distillation, exposing a blind spot in production moderation.