Accurate KV Cache Eviction via Anchor Direction Projection for Efficient LLM Inference
anchor directioncache accuracyinference latencykey-value cache evictionkv cache budgetlongbench benchmarkmemory optimizationperformance degradationprojection techniquesprojection-based scoringspatial relationshipsstate-of-the-art methodstoken importancetoken selectionvalue vectors
Key-Value (KV) cache eviction