A. Hasnat

Arabic hateful memes pose unique challenges that current detection models are ill-equipped to handle, revealing critical gaps in our understanding of online hate.

Training VLMs on a unified, multilingual, multitask meme dataset reveals that robust meme understanding requires multimodal training and is highly sensitive to dataset-specific overfitting.