Abstract
We introduce Dynamic Alpha Tuning (DAT), a dynamic model that adjusts the weighting coefficient between sparse and dense retrievers based on model confidence. DAT employs meta-learned schemes to adaptively skew contributions between retrieval methods. Our experimental results on HotPotQA and TriviaQA benchmarks show that this yields coverage and answer diversity advantages over static hybrid approaches. We also evaluate Bayesian methods, such as variational retrieval confidence and Monte Carlo dropout, as alternatives to estimate principled query-specific weighting.