Diffusion Soft Labeling for Model Calibration

Overconfidence is a pervasive problem for modern classification models based on neural networks. One key source is the hard sample labels used in cross-entropy loss, which are one-hot probability vectors that encourage high confidence. Existing model calibration approaches have tried introducing softer supervision signals, which however, rely on heuristic rules and may not accurately reflect the label uncertainty. In this work, we propose Diffusion Soft Labeling (DiffSL), which assigns soft labels according to a data-driven principle. Specifically, DiffSL constructs initial soft labels from diffusion classifier predictions with top-k class clipping, then corrects wrong labels using ground-truth labels. Extensive experiments show that DiffSL significantly improves calibration performance over existing methods while maintaining competitive classification accuracy.