Class-conditional conformal prediction works by calibrating a separate score threshold for each class, which requires labelled calibration examples of every class. As the abstract notes, in long-tailed deployments these examples are exactly what is missing for rare classes. The new paper, posted on arXiv, addresses this gap with a method called predicted class correction.
The proposed approach sets conformal thresholds for classes that have no calibration data, rather than skipping them or falling back to a global threshold. The abstract does not provide implementation details, but the framing suggests a correction applied to the predicted class distribution itself.
Because only one source was available, there are no differing findings to compare. The significance is clear: it targets a practical failure mode of conformal prediction in real-world, imbalanced datasets.