"In\^{t}elegi Rom\^ane\c{s}te?'' A Recipe for Romanian Vision-Language Models

A systematic effort to build a Romanian-specific vision-language model shows that language-adapted architectures can outperform larger, general-purpose counterparts, according to a study submitted on 29 May 2026 [1]. The work addresses a persistent performance gap for low-resource languages in multimodal AI. Vision-language models (VLMs) have largely followed the trajectory of text-only large language models, achieving strong results on English benchmarks while degrading sharply on low-resource languages where large-scale image-text corpora and culturally grounded evaluations are absent [1]. The researchers present a full pipeline for Romanian, from data construction to architectural choices, translating established English VLM training and evaluation corpora into Romanian [1]. Machine translation was applied to textual annotations and in-image text, preserving visual grounding while adapting the textual content [1]. Using this data, the team trained and ablated a series of VLMs to isolate the contributions of vision backbones of varying scale and pretraining, language backbones ranging from multilingual to Romanian-adapted LLMs, and OCR-style image-text data [1]. The Romanian-adapted VLMs consistently outperformed their same-sized counterparts and, across all evaluated benchmarks, surpassed models from the next larger size category [1]. The study also introduces HoraVQA, a culturally native evaluation set grounded in Romanian everyday scenes [1]. This benchmark provides a targeted measure of how well a model understands visual contexts specific to Romanian culture, rather than relying solely on translated English evaluations. While the Romanian VLM work focuses on adapting multimodal understanding to a specific language, parallel research continues to advance the underlying architectures of unified multimodal models. One recent paper proposes Representation Forcing, a technique that eliminates the need for an external generative latent space by making representation prediction a native capability of the model, closing a quality gap that appears when removing a frozen VAE for image generation [2]. Another study introduces Lumos-Nexus, a training-efficient unified video generation framework that uses a two-stage design to align a lightweight generator during training and then progressively hand off generation to a high-capacity pretrained generator during inference, achieving gains in visual realism and temporal coherence [3]. These architectural innovations, while not specific to low-resource languages, represent the broader technical landscape in which language-specific adaptations like the Romanian VLM must operate.