Muir: Linear Attention with Sensitivity-Allocated Gating and Matched-Filter Readout
Muir redesigns linear attention by allocating gate rank according to sensitivity and using matched-filter readout, improving memory retrieval, language modeling, and long-context efficiency.