FastV (KV-cache compression): superseded — cited as a baseline and beaten by newer methods. 2 paper(s) critique it, 2 beat it on benchmarks — #42 of 234 most-superseded. Sub-problem: cluster led by ReKV. Newer alternatives in the same sub-problem include MuKV, KVCapsule, Decoupled Streaming Cache (DSCache), LightKV, Hierarchical Adaptive Eviction (HAE).

KV-cache compression

FastV

Superseded baseline#42 of 234 most-superseded

Superseded — cited as a baseline and beaten by newer methods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 FastV as a baseline.

However, most existing methods depend on query-derived text tokens to compute attention scores and initiate compression, inevitably causing response delays in online scenarios.

— LiveVLM: Efficient Online Video Understanding via Streaming-Oriented KV Cache and Retrieval

“FastV, though training-free, prunes vision tokens without cross-modality guidance, yielding inconsistent results across models and benchmarks.”

— Make Your LVLM KV Cache More Lightweight

Beaten on benchmarks

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

KVCapsule beats FastV
1.01 vs 0.08
MMBench(fine) · [Qwen2.5]
KVCapsule: Efficient Sequential KV Cache Compression for Vision-Language Models with Asymmetric Redundancy
HAE-LLaVA (Ours, Retain, 192) beats FastV
61.7 vs 52.7
GQA · [LLaVA-1.5-7B with 192 visual tokens retained]
Hierarchical Adaptive Eviction for KV Cache Management in Multimodal Language Models

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Recent methods in the same sub-problem, not yet superseded in the knowledge base — arXiv benchmark leaders, not vetted production recommendations.