Let Confidence Change, Not the Prediction: Prediction-Preserving Repair for Post-hoc Calibration
CORD repairs calibrated probability outputs so the original top-1 class stays the prediction.
The paper says post-hoc multiclass calibration can change which class wins, even when the goal is only to adjust confidence. Its proposed adapter, CORD, uses the original and calibrated outputs to rebuild the probability vector while preserving the original argmax. The authors report zero Top-1 Prediction Change Rate by construction, with lower mean ECE, NLL, and Brier than direct calibrated outputs across CIFAR-10/100 and ImageNet-1K. The method does not alter the fitted calibrator, train another supervised map, or require tuned hyperparameters. ArXiv · AI/CL/LG's note
The paper says post-hoc multiclass calibration can change which class wins, even when the goal is only to adjust confidence. Its proposed adapter, CORD, uses the original and calibrated outputs to rebuild the probability vector while preserving the original argmax. The authors report zero Top-1 Prediction Change Rate by construction, with lower mean ECE, NLL, and Brier than direct calibrated outputs across CIFAR-10/100 and ImageNet-1K. The method does not alter the fitted calibrator, train another supervised map, or require tuned hyperparameters. ArXiv · AI/CL/LG's note
score 4