The paper derives a characterization of minimax risk for KV cache compression and maps it to practical design principles and an algorithm tested on LongBench.
SnapKV: LLM knows what you are looking for before generation.Advances in Neural Information Processing Systems, 37:22947–22970
2 Pith papers cite this work. Polarity classification is still indexing.
2
Pith papers citing it
citation-role summary
baseline 1
citation-polarity summary
fields
cs.LG 2years
2026 2verdicts
UNVERDICTED 2roles
baseline 1polarities
baseline 1representative citing papers
RDKV derives per-token and per-channel weights from attention distortion, then uses reverse water-filling to assign bit-widths from full precision to zero after prefilling, recovering 97.81% accuracy with 2.48% cache retention on LongBench.
citing papers explorer
-
The risk of KV cache compression
The paper derives a characterization of minimax risk for KV cache compression and maps it to practical design principles and an algorithm tested on LongBench.
-
RDKV: Rate-Distortion Bit Allocation for Joint Eviction and Quantization of the KV Cache
RDKV derives per-token and per-channel weights from attention distortion, then uses reverse water-filling to assign bit-widths from full precision to zero after prefilling, recovering 97.81% accuracy with 2.48% cache retention on LongBench.