We compare against state-of-the-art models for both
explicit and implicit knowledge storage on three popular Open-QA benchmarks, and find that we outperform all previous
methods by a significant margin (4-16% absolute accuracy), while also providing qualitative benefits such as
interpretability and modularity.