AI2 researchers release new multimodal approach to boost AI capabilities using images and audio