Reproduce “KV$^2$: A Self-Refining KV Cache” 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
Johannes Wesch, Danni Liu, Jan Niehues
KV² compresses long-context caches for later questions by selecting a small set of context tokens as reconstruction queries, then using attention to score entries. On RULER 16K with Qwen3-8B at 2% retention, it reports an average score of 62.19, versus 18.94 for KVzip; this is a subtask average, not a per-task guarantee, and distractor-heavy variants remain difficult.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.