Multi-hop retrieval-augmented generation (RAG) requires finding documents whose evidence is useful in combination. Standard retrieval methods often rank by query relevance or graph connectivity, but the authors of a new arXiv paper argue that neither directly measures the additional value a document brings to an already-selected set.

The paper, titled "Retrieving Complementary Evidence for Multi-Hop RAG with Semantic Gain Graphs" (arXiv:2610.15285), proposes a graph-based approach that scores documents by their semantic gain—how much new, useful information they add when combined with other evidence. This shifts the focus from individual relevance to complementary utility.

The work addresses a core challenge in multi-hop reasoning, where answers depend on synthesizing facts across multiple sources. As an arXiv preprint, the findings have not yet undergone peer review, but they point toward a more explicit way to model evidence combination in retrieval systems.