CARVE: Content-Aware Recurrent with Value Efficiency for Chunk-Parallel Linear Attention
arXiv:2606.27229v1 Announce Type: cross Abstract: Recurrent models must forget in order to remember, yet the state of the art decides what to erase without consulting what…