People solve unfamiliar problems by mapping them onto familiar ones. My work (e.g. GIFARC) builds datasets and training signals that surface these analogies, so a model’s first guess looks more like a human’s - fewer brittle shortcuts, more transferable structure.