WOVEN provides a training source and benchmark designed to improve the spatial, embodied, and temporal reasoning of MLLMs. It uses diverse video-pretrained generative rollouts to teach models visual transition reasoning.
HOW THIS AFFECTS YOU
●
builderThis framework can help you improve the physical and temporal reasoning of your multimodal models.
●
researcherYou can use this benchmark to conduct controlled comparisons of reasoning capabilities across different scenes and actions.