Image captioning plays a crucial role in numerous applications, including educational systems. However, ensuring caption quality remains a significant challenge, particularly for morphologically rich, low-resource languages such as Arabic. We investigate an evaluation of Arabic image captioning using state-of-the-art multimodal foundation models. We systematically assess the performance of leading models—Gemini, Gemma, LLaMA, and Fanar. Our evaluation framework employs a diverse set of metrics spanning rule-based, learnable, visually-grounded, and LLM-based approaches to capture semantic accuracy, linguistic fluency, and hallucination detection. Experiments are conducted on two benchmark datasets: Flickr8k-Arabic and JEEM. Our findings reveal significant performance variations across models and evaluation metrics, highlighting the need for Arabic-specific optimization in multimodal architectures.

Arabic social media platforms are increasingly using propaganda to deceive or influence people. This propaganda is often spread through multimodal content, such as memes. While substantial research has addressed the automatic detection of propaganda in English content, this paper presents the MODOS team’s participation in the Arabic Multimodal Propagandistic Memes Classification shared task. Our system deploys the Segment Anything Model (SAM) and CLIP for image representation and ARABIAN-GPT embeddings for text. Then, we employ LSTM encoders followed by a weighted fusion strategy to perform binary classification. Our system achieved competitive performance in distinguishing between propagandistic and non-propagandistic memes, scored 0.7290 macro F1, and ranked 6th among the participants.