Reproduce “CacheLink: Functional KV Transport for Zero-Prefill Cross-Model Context Handoff” 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
Haomiao Qiu, Yining Tang, Zijia Song, Ruizi Han, Ziyue Qiao, Miao Zhang
CacheLink translates a sender’s key/value cache into a state usable by another model, then learns a correction through the frozen receiver using continuation and predictive-distribution objectives. This avoids replaying the prefix during handoff. Across four within-family pairs, it beats Ridge in 26 of 28 comparisons; for Qwen3-14B→32B, its seven-panel mean is 50.32% versus 43.48%, not pooled accuracy or parity with native receivers.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.