Late fusion trains separate, independent models for each modality — each producing its own prediction, score, or decision — and combines those outputs only at the final decision stage, typically via averaging, weighted voting, a learned meta-classifier (stacking), or simple rule-based aggregation. This is the direct counterpart to early fusion (combination at the raw/feature input level, tested elsewhere in this set) and to intermediate/hybrid fusion (combination at one or more mid-network representation levels).

Late fusion's key practical advantage is modularity and robustness: because each modality's model operates independently until the final combination step, a missing or corrupted modality at inference time degrades performance gracefully rather than catastrophically — the surviving modalities' models can still contribute a prediction. It also allows each modality-specific model to be trained, validated, and even updated independently, which simplifies engineering in production systems. Its main disadvantage is that it cannot capture fine-grained, low-level cross-modal interactions, since by the time information reaches the fusion point, each modality has already been reduced to a high-level decision.

Options C and D describe feature-level and preprocessing-level combination respectively — both inconsistent with "late" in the fusion terminology, which specifically denotes the decision/output stage. Option B describes a training-schedule concept unrelated to fusion architecture.