We used to do vision-language grounding in a unidirectional way: language tells the model what to look for, and the model finds it in the image. But real grounding should go both ways — given an image and text, a model should discover what concepts are being referred to, where they appear, and how the two sides correspond, without being told the relevant text spans in advance. ConCor-1 is our attempt to revisit this standard paradigm. We formulate grounding as bidirectional concept correspondence, where the model jointly discovers referential concepts in language and vision and explicitly recovers their correspondence. I’m excited about this direction because it moves grounding from answering “where is this?” toward understanding the full correspondence structure between what we say and what we see. ConCor-1 is just the first step — we’re already working on extending this bidirectional correspondence view to video! Huge kudos to @ziqigaothu , whom I’ve had the pleasure of mentoring and who will be joining UW as a PhD student this fall!

MolmoMotion is not just another cool work built on top of Molmo2’s video grounding capabilities. It shows grounding in 3D can enable motion forecasting (for both data and capabilities), which serves as an objective and representation for controllable video generation and robotics!