Advances in World Models and Multimodal AI
ABot-World-0 achieves infinite interactive world rollouts on a single RTX 5090 at 16 FPS. Generative World Renderer pushed to 31.54 FPS. Video-Oasis paper finds 55% of video understanding benchmarks solvable without visual input. Vidu S1 launched real-time interactive video generation. OpenCoF learns reasoning through video generation. NVIDIA's Cosmos 3, Qianxun's Spirit v1.6, Flash-WAM. TESSERA v2. Odyssey raises $310M. UniAR. Real-time controllable 3D human motion generation at SIGGRAPH 2026. GenCeption shows video generation models as general-purpose vision learners. ReChannel. RoboTTT scales robot policy context to 8K timesteps. VideoChat3, UniVR. Survey frames video generative models as game engines. Xiaomi-Robotics-1. Nvidia's SIGGRAPH papers. VideoRAE. AV-Flamingo. Meta's SAM 3 and DINOv3 deployed at DOE labs. New: Game-native world model with explicit state (WildWorld dataset). Video world model trained on 15 hours of robot video generalizes zero-shot to unseen embodiments. Yann LeCun's SIGReg for JEPA world models (anti-collapse mechanism). ByteDance's FlowMimic enables mask-free video editing from image edits. Apple-PI benchmark tests physical law reasoning in video models via a three-stage protocol.