Unified multimodal models are designed to both answer questions about images and generate images using one shared set of weights. This design is efficient, but it introduces a risk that does not exist for single-purpose models: when the model is fine-tuned for one capability, the other capability may be damaged.
The new paper, posted on arXiv, focuses on this cross-modal forgetting and transfer during post-training. The abstract notes that fine-tuning one capability can harm another, and the work appears to investigate how these effects play out when a model is adapted after its initial training.
This is an important consideration for practitioners who want to update a model for a specific task without losing its general multimodal abilities. The paper underscores that post-training decisions in unified models require careful attention to the trade-offs between capabilities.