arxivcs.CLcs.AIcs.LGcs.NE2026-06-25
CARVE: Content-Aware Recurrent with Value Efficiency for Chunk-Parallel Linear Attention
Recurrent delta-rule models keep a fixed-size state matrix S (d_v x d_k) that compresses all past context. The state of the art (GDN-2) gates this update with element-wise matrix erase/write masks. This is powerful but has two defects. First, both gates are computed from the inco…