Preprint reports 31% lower sea-ice prediction error by preserving expert uncertainty

A machine-learning preprint reports a 31% reduction in mean absolute error on sea-ice concentration compared with training on hard labels. Its method preserves the ranges supplied by individual experts instead of collapsing their judgments into a single consensus. The model represents expert intervals as probability distributions and separates uncertainty within each label from disagreement between experts and uncertainty in the model itself. An additional training step aligns those components with their intended sources of uncertainty. The authors also report outperforming aggregation and interval-based baselines, with the demonstrated accuracy gain confined to the sea-ice task.