Genpei Zhang

Rare word combinations expose a critical flaw in scene text recognition systems that scaling up models alone cannot fix.

Vision-language models are often blind to visual evidence, with up to 97% of samples showing a striking insensitivity that undermines their effectiveness.