Conceptual
Browse Instruct Curate

Channel Polarization for Recency Bias and Over-Smoothing in State Space Models

A technique for structured state space models that reserves two channels of each state-transition matrix and fixes them to one and to zero, simultaneously countering the intrinsic recency bias that blocks long-range recall and the over-smoothing that makes deep SSM token representations indistinguishable, thereby improving long-range associative recall and depth scalability.