Rotary Position Encoding (RoPE)

Encodes token positions via vector rotations in complex space, enabling natural extrapolation to sequences longer than those seen during training.