Reproduce “KVPRIOR: On the Surprising Effectiveness of Marginalizing KVCaches Over Prefixes” 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
Masataro Asai, Bo Wu, Dan Gutfreund
KVPrior estimates reusable document key-value caches by prefilling each target after sampled prefix documents, resetting rotary-position keys, and averaging caches offline. It is training-free and adds no inference-time recomputation, though it requires extra offline prefills. With 4 sampled prefixes, it substantially improves over direct cache reuse and approaches full-prefill accuracy on Biography, but gains are more modest on multi-hop tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.