KVpop supervises fixed-budget KV eviction with an efficiently computed future-attention target and delayed mLSTM scoring, retaining ~97–100% of full-attention math performance at 75–88% compression.
T oken S elect: Efficient Long-Context Inference and Length Extrapolation for LLM s via Dynamic Token-Level KV Cache Selection
1 Pith paper cite this work, alongside 3 external citations. Polarity classification is still indexing.
1
Pith paper citing it
3
external citations · OpenAlex
fields
cs.LG 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
KVpop -- Key-Value Cache Compression with Predictive Online Pruning
KVpop supervises fixed-budget KV eviction with an efficiently computed future-attention target and delayed mLSTM scoring, retaining ~97–100% of full-attention math performance at 75–88% compression.