Gated DeltaNet-2: Decoupling Erase and Write in Linear Attention 23 May 2026·8 mins Machine Learning Paper Notes Machine-Learning Llm Architecture Linear-Attention State-Space-Models Arxiv