Generative Conformal Prediction with Optimized Coverage Allocation

ORCA ranks samples from a conditional generative model using a local density proxy, optimizes rank-specific radii on an exploration split, and calibrates the result on a separate split. This yields adaptive prediction regions with finite-sample marginal coverage.

The project page and interactive demo explains the geometry. Cite the current arXiv preprint using its 2024 title; we will update the citation when the revised version and proceedings are available.

Recommended citation: Minxing Zheng and Shixiang Zhu. (2024). "Generative Conformal Prediction with Vectorized Non-Conformity Scores." arXiv:2410.13735.
Download Paper