One Distribution, Two Channels: Coverage-Calibrated KV Retention and Compensation
Long-context decoding is increasingly constrained by key--value (KV) cache memory and bandwidth. Existing fixed-budget compression methods typically separate retention from compensation, while a retention ranking specifies neither discarded attention mass nor the direction of induced output error. We start from an exac...
Shu-Xin Liu, Qing Liu, Yi Du et al.
· 0 citations