Title: "Do You Understand Romanian?": A Framework for Developing Romanian Vision-Language Models

Abstract: While Vision-Language Models (VLMs) have largely mirrored the development path of text-only Large Language Models (LLMs), achieving high performance on English benchmarks, their efficacy drops significantly in low-resource languages. This decline is largely due to the absence of large-scale image-text datasets and culturally relevant evaluation metrics for these languages. This paper presents a comprehensive investigation into the creation of a Romanian-specific VLM, addressing the entire workflow from data preparation to architectural design. We adapt existing English VLM training and evaluation datasets by translating them into Romanian, utilizing machine translation for both textual annotations and in-image text. This process maintains visual grounding while localizing the linguistic content. Leveraging this localized data, we train and conduct ablation studies on various VLM configurations to determine the impact of three key factors: (i) vision backbones differing in scale and pretraining regimes, (ii) language backbones ranging from multilingual models to those specifically adapted for Romanian, and (iii) OCR-style image-text data. Additionally, we introduce HoraVQA, a novel evaluation benchmark rooted in everyday Romanian cultural contexts. Our results demonstrate that Romanian-adapted VLMs not only outperform same-sized baseline models but also exceed the performance of models from the next larger size category across all assessed benchmarks.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC