Split conformal prediction is widely used for distribution-free uncertainty quantification, but most theoretical analyses of interval length assume clean data. A new preprint on arXiv (2610.15458) highlights this gap, asking how contamination at the training and calibration stages alters interval length relative to the corresponding clean population.

The abstract frames this as an important but understudied question for robust conformal inference. While the available text does not detail specific results, the paper's motivation is clear: real-world datasets are often noisy, and understanding how contamination affects interval length is essential for reliable uncertainty estimates in practice.