Sparse retrieval relies on exact term matches, so vocabulary mismatch between a query and relevant documents can hurt performance. Large language models offer a workaround by generating alternative views of the query, which are then searched independently to broaden the set of matching terms.

But once those views are searched, the resulting ranked lists are not independent in a useful sense. They may be complementary, with each view surfacing different relevant documents, or redundant, repeatedly returning the same results. A naive fusion of scores would treat both cases the same way.

XViewRank, introduced in a new arXiv paper, proposes overlap-conditioned score fusion. Rather than combining scores blindly, the method accounts for how much the ranked lists overlap, aiming to exploit complementarity while discounting redundancy. The abstract does not detail the full algorithm, but the core idea is to make fusion sensitive to the relationship between the lists.