While video generation models like Sora and Veo are still racing to "make a single video look more cinematic," the next battlefield in the video generation field has quietly shifted — interactive world models. AnchorWorld, released by Kuaishou Kling team in collaboration with Tsinghua on arXiv, is a hardcore response from an industry-leading player to a "customizable, interactive, self-evolving" world simulation framework.
The core entry point of the paper is very clear: use 3D human motion as the first modality of interaction. The first-person view naturally suffers from view occlusion and body truncation, and the authors introduce an auxiliary supervision signal "decoupled from the agent's first-person perception," letting the model observe the agent's full body position relative to the environment from an external view, thus making the spatial anchoring of "person-world interaction" more solid.
More noteworthy is the self-evolution mechanism of "Anchor View + text-driven": several anchor views are defined under a unified world coordinate system, paired with text descriptions to constrain the dynamic evolution of local scenes. Simple, but effective — experiments show it strictly follows preset dynamics on spatiotemporal geometric consistency, and significantly leads on multiple SOTA benchmarks.
If previous world models (Project Genie, SANA-WM, etc.) answered "can it generate a video you can walk into," AnchorWorld answers "after walking in, can you rewrite this world like a game." This may well be the real bridge to embodied intelligence and AGI.