QEvict keeps the key-value cache in three tiers, preserves uncertain windows in compressed form, and restores them when attention returns, improving retention under fixed memory budgets.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
QEvict: Recoverable Quantized KV Eviction for Attention-Drift-Robust Long-Context Decoding
QEvict keeps the key-value cache in three tiers, preserves uncertain windows in compressed form, and restores them when attention returns, improving retention under fixed memory budgets.