arXiv study: Diverse RAG documents lift answer accuracy 17-47%

Retrieval-augmented generation (RAG) typically ranks documents by individual relevance, but the generator ultimately answers from the whole retrieved set. In a new arXiv preprint (2608.13956), researchers Jonathan J Ross, Bevan Koopman, Anton van der Vegt, and Guido Zuccon ask a deceptively simple question: does redundancy or diversity in retrieved documents actually improve LLM answers? Prior findings were mixed—some claimed repeated information reinforces correctness, others said LLM paraphrasing helps—but many studies failed to control for whether documents contained the exact answer or whether the model already knew it from training.

To isolate the effect, the team built FictionalQA, a synthetic dataset with fictional questions that LLMs cannot answer from parametric knowledge. They then fed generators three controlled retrieval scenarios: exact duplicate copies of a document, LLM-paraphrased versions of the same document, and documents from different genres (news, blogs, etc.) all containing relevant information in varied forms. Their results are striking: duplicate redundancy and paraphrasing produced no significant correctness gains. In contrast, diverse document sets improved answer correctness by 17% to 47%—and further testing confirmed this boost comes from genre diversity alone, not from simply having more answer-bearing content. The authors argue future retrieval methods should be designed to maximize set-level diversity for the generator, not just per-document relevance.