Fixing overconfident AI predictions without losing local accuracy

Machine learning models often output confidently wrong predictions, a problem especially acute in high-stakes settings like medicine or finance. This paper tackles multiclass calibration—making a model's confidence match its actual accuracy—by dividing the latent representation space into quantized regions and learning region-specific correction maps that share structure. Unlike global methods that assume uniform miscalibration or local methods that lose information through dimensionality reduction, their Vector Quantization approach achieves better local calibration while preserving global performance on standard benchmarks.