UniMoE-World: A Unified Mixture-of-Experts Architecture for Scalable Multi-Control Video Generation World Modeling

With the rapid progress of interactive video generation, video generation world models have gradually emerged as one of the mainstream paradigms in world model research and are increasingly regarded as a promising path toward efficient intelligent agents. However, existing video generation world models are typically developed under different forms of control supervision and mainly focus either on interactive world modeling or embodied world modeling, leaving the compatibility of heterogeneous control signals largely unexplored. In this work, we introduce UniMoE-World, the first training framework that enables unified learning of world models under heterogeneous supervisory controls by incorporating a Mixture-of-Experts (MoE) design into Diffusion Transformers (DiT). We further propose UniMoE-World Tuning, a continually extensible heterogeneous training strategy for world models, which supports diverse control signals, including robotic arms, hand joints, and camera poses, within a single world model and allows the model to be progressively expanded to new control settings. By enabling joint learning from more diverse data sources, this training strategy alleviates the scaling bottleneck of current world models. Our experiments show that world models trained with MoE over heterogeneous supervision consistently outperform those trained with any single control modality alone, demonstrating clear mutual gains across different control types. UniMoE-World achieves state-of-the-art performance on the WorldArena benchmark and shows particularly strong advantages in both locomotion and hand-motion capabilities over existing methods.