A team called SZED has described its submission to ArGuard 2026 Track A, the Arabic hateful meme detection task, in a paper posted on arXiv. The task is split into two parts: A1, a binary judgment of whether a meme is hateful, and A2, a fine-grained multi-label tagging problem.

For A1, the proposed system fuses MARBERTv2, a language model pretrained for Arabic, with CLIP ViT-L/14 visual features. The title of the paper also references multi-task label-attention and cross-modal fusion, suggesting the two subtasks share a common modeling framework rather than being solved completely separately.