Attention lets a model weight relevant tokens when building representations.

Attention helps models connect distant words and focus on important context, which is central to transformer architectures.