Our work devoted to rigorous comparison of independent vs joint start/end probability modelling assumptions in span prediction for extractive QA got accepted into MRQA@EMNLP’21. The pre-print & code is here.