Pixel-based language models are attracting interest as an alternative to token-based systems, promising to sidestep tokenization problems. However, the abstract notes that existing models are still largely English-centric and difficult to train, limiting their usefulness beyond a narrow range of settings.

MIXAR is introduced as a way to scale autoregressive pixel-based language models to multiple languages and scripts. The work directly addresses the gap between the promise of pixel-based modeling and the practical challenge of making such models work beyond English.

Because the abstract is brief, it does not include specifics about architecture, training data, or evaluation results. The paper is an updated preprint, so the claims should be treated as preliminary until further details are available.