Reciprocal Rank Fusion (RRF) is a standard method for combining sparse and dense retrievers in hybrid search, but it typically applies a fixed, uniform weight to each retriever regardless of the query. A new paper on arXiv argues that this is a critical limitation, because different queries may benefit from different balances between lexical and semantic signals.

The authors propose a query-adaptive version of weighted RRF (WRRF) and examine how optimal fusion weights behave. Their central claim is that the optimal weight for a given query is not a single point but an interval of values that yield comparable effectiveness. This interval perspective, they suggest, is more robust and practical than trying to pinpoint one best weight.

The study also incorporates latency awareness, meaning the choice of weights is evaluated not only for retrieval quality but also for computational cost. The abstract does not provide specific experimental numbers, so the full paper would be needed to assess the magnitude of the gains. As a single source, this is a standalone contribution; no other sources are available for comparison.