TY  - RPRT
TI  - Folding Attention Biases into Keys Makes Compacted KV Caches Run on Standard Kernels
AU  - Utkarsh Kanwat
PY  - 2026
UR  - https://arxiv.org/abs/2610.36520
ID  - 2610.36520
ER  -