subtype-robustness

This paper presents the first systematic study of calibration under unseen subtype shift, showing that models become overconfident on novel subtypes within known coarse categories, and argues that subtype robustness should be evaluated with calibration metrics rather than accuracy alone.