Reproduce “Request Order Matters: Cache-History Sensitivity in Selective KV-Cache Reuse for Rolling Agents” 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
Tiffany Gu, Annie Guan, Manshu Huang, Nitin Rao, Siddhant Shah, Margaret Capetz
The authors test document-aligned recomputation for reusing cached key–value states in rolling agent prompts. With the same 5% token budget, recomputing contiguous spans reduces mean pairwise exact-answer disagreement across five request orders from 69.0% to 26.1%, compared with token top-k. Chronological fidelity shows no significant difference, and reversed-order fidelity reaches only 54.3%.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.