LKV learns task-optimized global budgets and intrinsic KV token importance without attention matrices, delivering near-lossless performance at 15% cache retention on LongBench.
GQA : Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints , December 2023
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
-
LKV: End-to-End Learning of Head-wise Budgets and Token Selection for LLM KV Cache Eviction
LKV learns task-optimized global budgets and intrinsic KV token importance without attention matrices, delivering near-lossless performance at 15% cache retention on LongBench.