Pixel-Based Models Show Robustness to Typographical Noise

Pixel-based language models process rendered text directly, bypassing subword vocabularies to handle multilingual and noisy inputs. While they require longer training and more data to adapt, they maintain higher accuracy than text-based models under typographical distortions.