Scholar Feed

KV-cache compression

ReKV

Streaming Video Question-Answering with In-context Video KV-Cache Retrieval

Superseded baseline#23 of 234 most-superseded · first seen Mar 1, 2025

Superseded — cited as a baseline and beaten by newer methods

6 papers critique it · 2 beat it on benchmarks

What papers say

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

AGX+ReKV's coarse, frame-level selection offers modest latency gains but still requires selecting many tokens to maintain accuracy, limiting its effectiveness.
— V-Rex: Real-Time Streaming Video LLM Acceleration via Dynamic KV Cache Retrieval
“per-frame KV caches are insufficient for video modeling (lacking spatial details and temporal contexts) whilst consuming large amounts of storage that in turn reduces retrieval and QA performance”
— MuKV: Multi-Grained KV Cache Compression for Long Streaming Video Question-Answering
“KV cache offloading (e.g., ReKV) expands memory space yet incurs costly data transfer, repeated for each query.”
— InfiniPot-V: Memory-Constrained KV Cache Compression for Streaming Video Understanding
“Although retaining full Key-Value (KV) cache in the CPU and retrieving them based on queries facilitates the preservation and retrieval of fine-grained information, it leads to severe inefficiency.”
— LiveVLM: Efficient Online Video Understanding via Streaming-Oriented KV Cache and Retrieval
“However, it retains all generated KV caches, resulting in substantial memory consumption, and its retrieval strategy requires further optimization.”
— StreamKV: Streaming Video Question-Answering with Segment-based KV Cache Retrieval and Compression
“The offloading of KV cache could incur a lot of memory and is not scalable to ultra-long videos.”
— StreamMem: Query-Agnostic KV Cache Memory for Streaming Video Understanding

Beaten on benchmarks

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Recent methods in the same sub-problem, not yet superseded in the knowledge base — arXiv benchmark leaders, not vetted production recommendations.