DAMI Optimizes Gated Linear Attention via Dual-Axis Memory Initialization

DAMI improves long-context retrieval in Gated Linear Attention by jointly configuring gate parameters and input-sensitivity profiles during initialization. By assigning geometric timescales across heads, DAMI enables better memory retention and varied retrieval distances in GDN-family models.