The defining goal of multimodal machine learning is to learn a shared (joint) representation space that captures cross-modal relationships and correspondences --- allowing information from one modality to inform, constrain, or complete information from another. This shared representation is what enables tasks like cross-modal retrieval (finding images from a text query), cross-modal generation (text-to-image, image-to-text), and joint reasoning (visual question answering), all of which require the model to relate concepts across modality boundaries rather than process each in isolation.

How that shared representation is learned varies --- contrastive objectives (CLIP), joint embedding via co-attention (VisualBERT, LXMERT), or fusion layers that combine modality-specific features --- but the underlying principle is consistent across architectures: linkage happens through learned representations, not fixed rules or arbitrary connections.

Option C describes a specific, narrow ensembling strategy (tree-based combination of separate unimodal models) that is neither standard nor representative of how modern multimodal systems establish cross-modal relationships; it also conflates 'linking modalities' with 'combining model outputs,' which is closer to late fusion than to representation learning. Option D is simply the negation of the field's core premise. Option B introduces randomness where structure is explicitly what is being learned.