How it works

Images, text, geometry, and motion all become tokens in one model. RL from task feedback teaches it when to lean on each representation type.