Approximating parameter posteriors in likelihood-free settings is a practical challenge common to many scientific disciplines. While recent advances in both computer simulation and generative modeling have paved the way for tractable inference in high-fidelity environments, they often require prohibitively large sample sizes in practice. Sequential posterior estimation methods attempt to mitigate this by iteratively producing proposal distributions that refine the inverse model, but they lack explicit selection mechanisms for reducing information overlap in proposed simulations. In this work, we introduce a mutual information-based acquisition scheme for identifying informative simulation parameters, operating on disagreement in the parameter space across a weighted posterior ensemble of atomic proposals. Our approach crucially leverages only an inverse model, making it compatible with existing direct posterior estimation procedures. We demonstrate the potential of this method on several common simulation-based inference (SBI) benchmark tasks, and observe performance advantages over non-ensemble counterparts in low-data regimes.