Reproduce “Context as Infrastructure: Knowledge Compilation for Retrieval in Agentic Systems with Harnesses” 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
Zhichang Yu, Baining Cao, Sbaaoui Idriss, Qi Wang, Gangqiang Xu, Ling Qin, Xiaoling Jin, An Liu and 10 more
The authors test whether compiling documents into claim–evidence pairs can reduce what an agent reads while answering questions. A BM25-and-grep harness reaches 91% accuracy on FRAMES without compilation. On 200 FRAMES questions, projected evidence cuts tokens 91% when accepted, but raises them 1.9% when rejected; overall tokens fall 38%, with no reliable accuracy change in this English, Wikipedia-based benchmark.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.