Selecting Pre-trained Model Architectures for Specific Tasks

Self-supervised pre-training can be applied to different underlying model structures to create systems optimized for specific kinds of tasks. Match each model architecture with the description of the task it is most fundamentally suited for.

A team is building a foundation model intended primarily for abstractive summarization tasks, which require processing a source document and generating a new, coherent summary. They choose a full encoder-decoder architecture for self-supervised pre-training. What is the most critical reason this architecture is better suited for this task than an encoder-only or decoder-only model?