MIMEx: Intrinsic Rewards from Masked Input Modeling