Definition
A training objective that pays extra attention to the examples a model is still getting wrong.
A modification of cross-entropy that downweights easy examples and upweights hard ones via a focusing factor; originally from object detection, applied as a cross-domain substitution in Autoresearch agentic training.