Reproduce “Why Retrieval-Augmented Generation Fails: A Graph Perspective” 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

Kai Guo, Xinnan Dai, Zhibo Zhang, Nuohan Lin, Shenglai Zeng, Jie Ren, Haoyu Han, Jiliang Tang

Researchers from Michigan State University and MIT investigated why Retrieval-Augmented Generation (RAG) models generate incorrect answers even when provided with relevant context, by developing a graph-based framework to analyze internal evidence processing. They found that correct RAG predictions correspond to deeper, more structured internal reasoning paths, while errors result from shallower, fragmented information flow. The study introduced an attribution-graph based error detection method that achieved an 11.53% average accuracy improvement over logit-based baselines, and a targeted intervention that increased accuracy by 9% on a challenging mixed-context task.

Key insights

  • Correct RAG answers consistently arise from deeper, more structured, and more evenly distributed internal computational circuits, in contrast to shallower, fragmented, and concentrated information flow in incorrect predictions.
  • Successful reasoning follows a 'Question-Constrained Evidence Grounding' (QCEG) pattern, characterized by strong early question comprehension and consistent question constraint during answer formation, while failures exhibit 'Surface-Aligned Evidence Grounding' (SAEG), where models prematurely rely on superficially aligned external context.
  • Correct predictions utilize more mid-layer processing, indicating a deeper integration of information, whereas incorrect ones rely more on early layers, suggesting surface-level pattern matching without thorough integration.

Summary by alphaXiv.

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