Standard CP
One ball around the mean
Coverage-
Mean area-
ORCA · Project page
NeurIPS 2026
A generative model gives many plausible outcomes, and they often concentrate in several separate modes. Standard CP draws one region around a single prediction, so it must also cover the empty space between modes. PCP keeps the modes apart with a ball around every sample, but every ball gets the same radius. ORCA reads the local geometry of the samples: it ranks them from crowded to isolated and learns a separate radius for each rank. The result is a set of disconnected pieces with adaptive radii that stay close to the likely outcomes, and a final calibration step keeps the guaranteed coverage, such as 90%.
Draw samples from the generative model and rank them by their mean distance to a few nearest neighbours: densest first, stray samples last.
On an exploration split, learn one radius per rank: the smallest balls that still cover the target share of outcomes. Isolated samples can get no ball at all.
On a separate calibration split, scale every radius by one common factor, chosen so the target share of outcomes is covered. This step alone guarantees the coverage.
Standard CP, PCP and ORCA build prediction sets from the same generated samples. Move a slider and everything is recomputed on the spot, from sampling to the final coverage check.
One ball around the mean
Equal balls around every sample
A trained generator is never perfect, so a few of its samples land where Y never goes: the dots that land away from the true distribution. Those samples sit alone. ORCA ranks them last and gives them little or no radius, while PCP draws the same ball around every sample.
@inproceedings{zheng2026orca,
title = {Generative Conformal Prediction with Optimized Coverage Allocation},
author = {Zheng, Minxing and Zhu, Shixiang},
booktitle = {Advances in Neural Information Processing Systems},
year = {2026}
}