“Applying text-only fine-tuning to a multimodal model can result in reasonable behavior when processing images, demonstrating cross-modal generalization.”