Reproduce “StitchRAG: Constructing Compact and Evidence-Preserving Contexts for RAG with Diffusion-Based Stitching” 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

Minsang Kim, Junpyo Lim, Heeje Cho, Seungjun Baek

StitchRAG curates retrieved evidence by assigning each passage a retention budget, preserving answer-bearing sentences, and using text diffusion to fill gaps. In tested Top-10 to Top-30 retrieval, its average F1 across four datasets stayed near 51, while Naive RAG fell to 45.1 at Top-30. This shows stability in that sweep, not that more retrieval always harms RAG.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.