A novel test-time adaptation method, MiDl, that enhances the robustness of pretrained models against missing modalities without requiring retraining.
