A mechanism that dynamically selects which text tokens activate at which geometric scales, concentrating cross-modal computation where alignment is most needed.