Scholar Feed

KV-cache compression

CaM

Superseded baseline#9 of 234 most-superseded

Superseded — cited as a baseline and beaten by newer methods

6 papers critique it · 6 beat it on benchmarks

What papers say

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

these single-modal optimizations exhibit limited efficacy in MLLMs due to cross-modal distribution shifts and attention pattern divergence, failing to preserve modality-specific information fidelity.
— FlowMM: Cross-Modal Information Flow Guided KV Cache Merging for Efficient Multimodal Context Inference
“In contrast, CaM zhang2024cam adaptively merges evicted value states into others but does not merge the corresponding keys.”
— KeepKV: Eliminating Output Perturbation in KV Cache Compression for Efficient LLMs Inference
“most merging methods rely on local heuristics, they often funnel many evicted tokens into a small set of span-boundary tokens. These boundary tokens therefore become the main information carriers, overloading their representations and making them prone to over-merging: excessive aggregation can blur or even erase their original semantics, thereby degrading overall performance.”
— GRKV: Global Regression for Training-Free KV Cache Compression in Long-Context LLMs
“CaM merges the values of evicted tokens only probabilistically, with a non-negligible probability of discarding them and hence losing information.”
— WeightedKV: Attention Scores Weighted Key-Value Cache Merging for Large Language Models
“a fundamental limitation of these approaches is their uniform treatment of keys and values during merging despite their distinct distributional characteristics.”
— Homogeneous Keys, Heterogeneous Values: Exploiting Local KV Cache Asymmetry for Long-Context LLMs
“Notably, both Quest and CaM report results only on LLaMA2, without comparisons to other eviction methods, limiting their relevance to current frontier models.”
— CAOTE: KV Cache Selection for LLMs via Attention Output Error-Based Token Eviction

Beaten on benchmarks

Head-to-head results where a newer method reports beating CaM. 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Recent methods in the same sub-problem, not yet superseded in the knowledge base — arXiv benchmark leaders, not vetted production recommendations.