Trust the uncertain teacher: distilling dark knowledge via calibrated uncertainty

In three linesCalibrated Uncertainty Distillation (CUD) improves knowledge distillation by preserving the teacher's calibrated uncertainty instead of overconfident predictions. The approach guides students to learn from balanced distributions between confident signals and structured uncertainty, improving accuracy, calibration, and robustness under distribution shift.

Summary generated by Claude — human-verified