arXivAli Hatamizadeh, Yejin Choi, Jan KautzThu, May 21, 2026, 10:44 AM PDT
score 14.8
New linear attention model better balances memory updates
Original: Gated DeltaNet-2: Decoupling Erase and Write in Linear Attention
Source: arxiv.org ↗
Writing ELI5 summary…