RESEARCH LIBRARY

Attention

注意力计算、混合机制与位置表示。

共 89 篇论文 · 按研读更新时间排序
2611.08376v1 · r1

Muir

Muir: Linear Attention with Sensitivity-Allocated Gating and Matched-Filter Readout

Muir reallocates low-rank gate capacity by state sensitivity, adds matched-filter readout to reduce interference and write-time erasure, and uses a boundary-pivoted TPU chunk kernel; experiments show strong language modeling and multi-hop retrieval with static/dynamic readout trade-offs.