A new arXiv study finds that GraphRAG can over-cite sources while delivering low citation precision, and that the effect changes depending on the corpus.
The researchers varied the embedder, corpus, and judging model across thousands of runs. They report that GraphRAG emitted 11 to 15 citation IDs per answer while citation precision ranged from 0.12 to 0.23, even as retrieval recall remained much higher. That means the system often found relevant material but attached too many weak citations to its answers.
The consequence was not identical everywhere. In typed-edge DO-178C requirements, faithfulness reportedly fell from 74 percent to 40 percent across hops. On Wikipedia-style chains, the same pipeline improved because over-cited paragraphs still provided topical support.
The practical lesson is that retrieval architecture alone does not guarantee trustworthy citations. Teams using GraphRAG need to evaluate precision, corpus structure, and judge reliability rather than assuming that more graph context automatically means better evidence.