An arXiv preprint reports that steering vectors derived from one modality can transfer zero-shot to another in several unified multimodal models. The authors report bidirectional transfers between text-to-image and image-to-text generation across concepts including emotion and object size. They argue that steering provides a causal test beyond correlational alignment such as representational similarity, and that highly transferable vectors tend to point in similar directions for the same concept despite a gap between text and image representations. The preprint includes links to code, data and a demo; the supplied report does not identify the evaluated models.