arXiv:2603.17541v1 Announce Type: new 
Abstract: Multimodal large language models (MLLMs) are typically trained in multiple stages, with video-based supervised fine-tuning (Video-SFT) serving as a key step for improving visual understanding. Yet its effect on the fine-grained evolution of visual capabilities, particularly the balance between spatial and temporal understanding, remains poorly understood. In this paper, we systematically study how Video-SFT reshapes visual capabilities in MLLMs. Across architectures, parameter scales, and frame sampling settings, we observe a consistent pattern: Video-SFT reliably improves video performance, but often yields limited gains or even degradation on static image benchmarks. We further show that this trade-off is closely tied to temporal budget: increasing the number of sampled frames generally improves video performance, but does not reliably improve static image performance. Motivated by this finding, we study an instruction-aware Hybrid-Frame strategy that adaptively allocates frame counts and partially mitigates the image-video trade-off. Our results indicate that Video-SFT is not a free lunch for MLLMs, and preserving spatial understanding remains a central challenge in joint image-video training.

دراسة حديثة حول نماذج اللغة متعددة الوسائط (MLLMs) تسلط الضوء على دور التعديل الخاضع للإشراف القائم على الفيديو (Video-SFT) في تحسين الفهم البصري. تكشف الأبحاث أنه بينما يحسن Video-SFT أداء الفيديو بشكل كبير، فإنه غالبًا ما يؤدي إلى مكاسب محدودة أو حتى تدهور في المعايير الخاصة بالصور الثابتة، مما يشير إلى وجود توازن معقد بين القدرات الزمنية والمكانية.

Un estudio reciente sobre Modelos de Lenguaje Multimodal (MLLMs) destaca el papel de la afinación supervisada basada en video (Video-SFT) en la mejora de la comprensión visual. La investigación revela que, aunque Video-SFT mejora significativamente el rendimiento en video, a menudo conduce a ganancias limitadas o incluso a una degradación en los benchmarks de imágenes estáticas, lo que indica un complejo compromiso entre las capacidades temporales y espaciales.

Une étude récente sur les modèles de langage multimodaux (MLLMs) met en lumière le rôle de l'affinage supervisé basé sur la vidéo (Video-SFT) dans l'amélioration de la compréhension visuelle. La recherche révèle que, bien que Video-SFT améliore considérablement les performances vidéo, il entraîne souvent des effets limités ou même une réduction de l'efficacité sur les benchmarks d'images statiques, indiquant un compromis complexe entre les capacités temporelles et spatiales.

A recent study on Multimodal Large Language Models (MLLMs) highlights the role of video-based supervised fine-tuning (Video-SFT) in enhancing visual understanding. The research reveals that while Video-SFT significantly improves video performance, it often leads to limited or even reduced effectiveness on static image benchmarks, indicating a complex trade-off between temporal and spatial capabilities.