Reproduce “Cache Locally, Recompute Selectively for Streaming Video Understanding” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Junhyeok Lee, Sangyoon Lee, Jaeho Lee
FrameBlend prepares video frames with a short attention window, then selects query-relevant cached tokens, recomputes their key–value representations in temporal order, and replaces them before answering. On RVS-Ego, it raises LLaVA-OneVision-7B with ReKV accuracy from 59.3% to 62.4%, while TTFT rises from 0.17 to 0.93 seconds, so the accuracy gain does not imply faster first-token response.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.