【#Tech24H】The Beijing Institute of Artificial Intelligence (BAAI) recently unveiled a new multimodal representation world model, the Physis · RoboBrain Orca, which breaks through the existing paradigm of large models that merely predict single-modal outputs such as text, images, or actions. It is the first to achieve forward and backward evolution inference over the overall state of the world, opening up a brand-new technical path for the development of foundational general artificial intelligence models. The core innovation of this model lies in constructing a unified latent representation space of the world, effectively building a virtual “mental world” for AI. After receiving any multimodal input, whether video, images, text instructions, or event descriptions, the model integrates diverse signals such as vision, language, and task intentions into a coherent whole. It autonomously learns object motion laws, scene evolution logic, causal relationships between actions, and temporal event sequences. It can not only infer the future evolution trajectory of a current scene but also simulate multiple possible developments of the world under different conditions, completing a paradigm shift from “single-output prediction” to “global state inference”. [ By Zhang Liyan | Tang Ruohan ]