What's this about?
WM-VLM adds a small branch that generates intermediate visual states, then uses those states to solve spatial tasks rather than reasoning only in text.
What exactly is a mental rotation task?
The model must imagine an object turned in two or three dimensions and determine how its orientation or parts will look afterward.
Why does this matter?
Large gains and corruption tests suggest the images contribute causally rather than serving as decoration. The tasks are synthetic and narrower than open-world visual reasoning.