锦秋基金被投Manifold AI发布通用交互世界模型，让智能体具备实时未来预测能力｜Jinqiu Spotlight

Core Insights - The article discusses the advancements of Manifold AI in developing a universal interactive world model called WorldScape, which ranks first in the current mainstream world model evaluation, WorldScore [5][40]. Investment Background - Jinqiu Fund led the angel round investment in Manifold AI in 2025 and continued to invest in the angel+ round, focusing on breakthrough technologies and innovative business models in the field of general artificial intelligence [4]. World Model Overview - The vision of world models is to create an infinite and realistic "virtual laboratory" for agents, allowing them to explore, learn, and make decisions without the costs associated with real-world trial and error [7]. - WorldScape is designed to provide real-time interactive capabilities, integrating controllable video generation, 3D spatial consistency, and long-term world memory [9]. Key Features of WorldScape - Comprehensive Interaction Experience: WorldScape unifies action and world state modeling, supporting both spatial navigation and object manipulation [11]. - Stable 3D World Structure: The model incorporates explicit 3D geometric perception to maintain consistent spatial structures during interactions, addressing common issues like geometric drift [12]. - High Visual Quality in Real-Time Generation: Achieves near real-time interactive generation (6-16 FPS) on a single GPU without sacrificing visual quality [13]. - Memory Mechanism: WorldScape features a memory mechanism that allows for long-term consistency, distinguishing it from traditional video generation models [14]. Technical Innovations - Spatial Consistency Training: Utilizes a multi-task learning paradigm to integrate geometric constraints into the model's cognition [19]. - Joint Optimization: The training process employs complementary supervision from flow matching loss and 3D geometric signals to enforce strong constraints on scene structure [20]. - Efficient Long Sequence Consistency Modeling: Introduces a KV cache optimization strategy to manage memory efficiently during long video generation [24]. Interaction Control - WorldScape supports unified interaction perception, allowing for camera trajectory control and hand motion control, enabling both navigation and object manipulation within the same model [30][31]. Evaluation Metrics - WorldScape emphasizes balanced performance across multiple dimensions, including visual quality, interactivity, memory, and real-time capabilities [38]. - The model outperforms others in the WorldScore evaluation, which assesses the ability to maintain spatial structure, semantic content, and temporal evolution consistency [40]. Conclusion and Future Outlook - WorldScape addresses existing limitations in universality and real-time performance, aiming to become a foundational model for general embodied intelligence [42].