Reusing pre-computed vision tokens during language generation to avoid re-encoding images, reducing computation at inference time.