As generative models achieve increasingly greater performance, the research frontier is shifting toward new challenges of multimodal and interactive world models. This talk presents recent advancements and insights across three interconnected themes. First, we introduce unified frameworks for multimodal understanding and generation, exploring methods to enhance their semantic-spatial reasoning abilities. Second, we demonstrate interactive video generation systems that incorporate action control mechanisms, enabling gaming-like experiences where users dynamically influence content evolution. Our solutions address critical challenges in memory and 3D consistency during prolonged interaction sessions. Finally, we propose autoregressive visual generation architectures that inherently support multimodal integration and interactivity. Through systematic architectural innovations, we overcome longstanding bottlenecks in output quality and computational efficiency, establishing a viable alternative to diffusion-based paradigms. Looking into the future, we aim to build multimodal and interactive visual generation models as world models.