Train Short, Test Long: Attention with Linear Biases Enables Input Length Extrapolation.

Ofir Press, Noah A. Smith, Mike Lewis

Browse the full ICLR paper archive.