Is simple top-K KV cache pruning superseded?

Q: Is simple top-K KV cache pruning superseded?

simple top-K KV cache pruning (KV-cache compression): cited as a baseline — critiqued by newer work, not yet beaten on a benchmark here. 1 paper(s) critique it, 0 beat it on benchmarks — #209 of 234 most-superseded. Sub-problem: cluster led by LURE. Newer alternatives in the same sub-problem include PruneHal.

KV-cache compression

simple top-K KV cache pruning

Superseded baseline#209 of 234 most-superseded

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

1 papers critique it · 0 beat it on benchmarks

What papers say

Verbatim critique sentences, each from a paper that cites simple top-K KV cache pruning as a baseline.

this strategy struggles to determine the optimal time for pruning, often leading to excessive pruning. This, in turn, results in the loss of crucial visual information, negatively impacting the overall performance of the model.

— PruneHal: Reducing Hallucinations in Multi-modal Large Language Models through Adaptive KV Cache Pruning

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Recent methods in the same sub-problem, not yet superseded in the knowledge base — arXiv benchmark leaders, not vetted production recommendations.

PruneHalPruneHal: Reducing Hallucinations in Multi-modal Large Language Models through Adaptive KV Cache Pruning
Oct 22, 2025