Multimodal models often combine inputs such as images and text into a single latent representation. This fusion tends to boost accuracy on downstream tasks, but the resulting representation is opaque, which is a problem in settings where decisions must be explained.

The paper, released on arXiv, proposes aligning these multimodal embeddings with domain-specific language. By tying the latent space to terminology and concepts from the target domain, the authors aim to make predictions more interpretable without sacrificing the benefits of fusion.

The abstract notes that the lack of interpretability has limited deployment in domains that require explainable model decisions. The proposed alignment is presented as a way to address that limitation, though the abstract does not provide experimental details. Readers interested in the method will need to consult the full paper.