arXiv:2603.19235v3 Announce Type: replace 
Abstract: While Multimodal Large Language Models demonstrate impressive semantic capabilities, they often suffer from spatial blindness, struggling with fine-grained geometric reasoning and physical dynamics. Existing solutions typically rely on explicit 3D modalities or complex geometric scaffolding, which are limited by data scarcity and generalization challenges. In this work, we propose a paradigm shift by leveraging the implicit spatial prior within large-scale video generation models. We posit that to synthesize temporally coherent videos, these models inherently learn robust 3D structural priors and physical laws. We introduce VEGA-3D (Video Extracted Generative Awareness), a plug-and-play framework that repurposes a pre-trained video diffusion model as a Latent World Simulator. By extracting spatiotemporal features from intermediate noise levels and integrating them with semantic representations via a token-level adaptive gated fusion mechanism, we enrich MLLMs with dense geometric cues without explicit 3D supervision. Extensive experiments across 3D scene understanding, spatial reasoning, and embodied manipulation benchmarks demonstrate that our method outperforms state-of-the-art baselines, validating that generative priors provide a scalable foundation for physical-world understanding. Code is publicly available at https://github.com/H-EmbodVis/VEGA-3D.

أدت التطورات الأخيرة في الذكاء الاصطناعي إلى تقديم VEGA-3D، وهو إطار يعيد استخدام نماذج انتشار الفيديو المدربة مسبقًا لتعزيز فهم المشاهد من خلال الاستفادة من الأولويات ثلاثية الأبعاد الضمنية. يتناول هذا التطور قيود نماذج اللغة متعددة الوسائط الحالية التي تكافح مع التفكير المكاني والديناميات الهندسية.

Los recientes avances en inteligencia artificial han llevado a la introducción de VEGA-3D, un marco que reutiliza modelos de difusión de video preentrenados para mejorar la comprensión de escenas aprovechando priors 3D implícitos. Este desarrollo aborda las limitaciones de los modelos de lenguaje multimodal (MLLM) existentes que luchan con el razonamiento espacial y las dinámicas geométricas.

Des avancées récentes en intelligence artificielle ont conduit à l'introduction de VEGA-3D, un cadre qui réutilise des modèles de diffusion vidéo pré-entraînés pour améliorer la compréhension des scènes en s'appuyant sur des priors 3D implicites. Ce développement répond aux limitations des modèles de langage multimodaux existants qui peinent avec le raisonnement spatial et les dynamiques géométriques.

Recent advancements in artificial intelligence have led to the introduction of VEGA-3D, a framework that repurposes pre-trained video diffusion models to enhance scene understanding by leveraging implicit 3D priors. This development addresses the limitations of existing multimodal large language models (MLLMs) that struggle with spatial reasoning and geometric dynamics.