Bidirectional Attention

Allowing tokens to attend to both past and future context, used in encoder models like BERT.

This concept is essential for understanding large language models and forms a key part of modern AI systems.