Scholar Feed

KV-cache compression

PagedAttention

Superseded baseline#37 of 234 most-superseded

Cited as a baseline — critiqued by newer work, not yet beaten on a benchmark here

5 papers critique it · 0 beat it on benchmarks

What papers say

Verbatim critique sentences, each from a paper that cites PagedAttention as a baseline.

System-level approaches like PagedAttention optimize KV storage to mitigate fragmentation, yet remain agnostic to the semantic importance of content under tight budgets.
— ArborKV: Structure-Aware KV Cache Management for Scaling Tree-based LLM Reasoning
“Orthogonally, system-level solutions like PagedAttention kwon2023efficientmemorymanagementlarge optimize memory management but do not reduce the fundamental size of the cache itself.”
— SWAN: Sparse Winnowed Attention for Reduced Inference Memory via Decompression-Free KV-Cache Compression
“Despite these advancements, no existing long-context benchmarks evaluate KV cache reuse scenarios.”
— SCBench: A KV Cache-Centric Analysis of Long-Context Methods
“applies paging techniques to reduce memory fragmentation but maintains full precision storage”
— GPU-Accelerated INT8 Quantization for KV Cache Compression in Large Language Models
“These systems treat all transferred KV at uniform precision; adds per-token precision differentiation”
— SpectrumKV: Per-Token Mixed-Precision KV Cache Transfer for Prefill-Decode Disaggregated LLM Serving

What to use instead

Recent methods in the same sub-problem, not yet superseded in the knowledge base — arXiv benchmark leaders, not vetted production recommendations.