Preprint
Jul 2026
Relative Positions Generalize, Absolute Positions Memorize: An Implicit-Bias Account of Length Generalization in Attention
It is proved that rotary encodings make the attention logit a function of relative offset alone, an exact equivariance, so whatever selection rule is learned at training lengths is reproduced verbatim at every longer length.
Subham Singh, Ashutosh Mishra, S. Raut
· 0 citations