Instead of treating annotator disagreement as noise, you can use it as a signal to train models that reproduce individual annotators' reasoning styles—useful for building explanation systems grounded in real human decision-making patterns.
This paper shows that large language models can learn to mimic how individual annotators explain their decisions, not just their labels. The researchers developed a method called CAPO that trains models on one annotator's explanations while contrasting them with other annotators' valid but different explanations for the same input, revealing stable patterns in how people reason about tasks.