Scholar Feed

KV-cache compression

Loki

Loki: Low-rank Keys for Efficient Sparse Attention

Superseded baseline#45 of 234 most-superseded · first seen Jun 4, 2024

Superseded — cited as a baseline and beaten by newer methods

2 papers critique it · 2 beat it on benchmarks

What papers say

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

Approaches like SparQ Attention ribar2024sparqattentionbandwidthefficientllm, AQUA Attention s2025aquaattentionquerymagnitudes, and Loki singhania2024lokilowrankkeysefficient prioritize computational savings over memory reduction.
— SWAN: Sparse Winnowed Attention for Reduced Inference Memory via Decompression-Free KV-Cache Compression
“Another representative work is Loki~singhania2024loki, which performs the PCA transformation of KV cache and then selects % tokens based on attention scores computed in low-dimensional space for sparse attention. However, this method still suffers from the increasing KV cache size.”
— SALS: Sparse Attention in Latent Space for KV cache Compression

Beaten on benchmarks

Head-to-head results where a newer method reports beating Loki. Values are copied from the source paper's tables — verify against the cited paper.

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Recent methods in the same sub-problem, not yet superseded in the knowledge base — arXiv benchmark leaders, not vetted production recommendations.