Calibration Framework for Probabilistic Label Ranking

A new study formalizes calibration for probabilistic label ranking, a task where models predict distributions over orderings of a label set. The authors define a hierarchy of calibration notions covering full rankings, sub-rankings, and top-k rankings, proving that full-rank calibration implies the others but not vice versa, and that sub-ranking and top-k calibration are incomparable. Empirical tests show popular label ranking models are often poorly calibrated, with significant discrepancies between sub-ranking and top-k metrics. The framework is applied to RLHF reward models, revealing calibration issues in preference learning.