Positional Encoding

A method that injects order information into Transformer models so sequence positions become visible.

Transformer architectures do not inherently encode sequential order, so positional encoding is used to inject order information into the model. This can be implemented through sinusoidal patterns or learned positional representations. Without it, the ordering of words or elements is lost and semantic structure weakens. Positional encoding is therefore one of the core building blocks of attention-based architectures.