Tag
This paper studies swap-agnostic learning of proper losses, showing that prediction-level comparisons can be controlled jointly via second-order multicalibration, achieving tight rates for finite hypothesis classes and families of losses.