How it works

It reframes recurrent state space models as a form of structured attention. This duality lets it use efficient matrix-multiply hardware, removing the quadratic cost.