Modality alignment is the process of establishing correspondence between semantically related elements across different data types — for example, matching a spoken word to its corresponding lip movement in video, or a caption phrase to the image region it describes. It is distinct from fusion (combining modalities into a joint representation) and from data integration (option B, which describes ingestion rather than alignment). Alignment can be explicit, as in dynamic time warping for audio-text synchronization, or implicit, learned end-to-end through attention mechanisms such as cross-attention in transformer architectures. CLIP's contrastive objective is itself a form of learned alignment: it pulls matching image-text pairs together in embedding space while pushing non-matching pairs apart, producing an aligned shared representation without explicit temporal correspondence. Alignment quality directly affects downstream fusion: poorly aligned modalities introduce noise that fusion layers cannot fully compensate for, which is why alignment is typically treated as a prerequisite step, not an afterthought.

Option A describes model reuse for custom tasks (closer to transfer learning), while C describes handling missing modality data, a separate robustness concern. Neither captures the correspondence-building nature of alignment. On the NCA-GENM exam, expect alignment questions to be paired with fusion and co-embedding concepts.