ReCIRC: Rectified Conformal Risk Control
ReCIRC adjusts conformal risk control around each input’s estimated local risk while keeping CRC’s finite-sample marginal guarantee.
The paper targets cases where one global CRC threshold can be too cautious for easy inputs and too weak for hard ones. ReCIRC reframes the threshold as a shared conditional-risk budget, then runs CRC on that transformed family. The authors say the guarantee holds even if the local risk estimates are poor, while accurate estimates can move it toward conditional risk control. In tests across synthetic and real segmentation, multilabel, multiclass, and regression settings, ReCIRC had the lowest average worst-group risk and mean positive group excess while keeping marginal risk near target. ArXiv · AI/CL/LG's note
The paper targets cases where one global CRC threshold can be too cautious for easy inputs and too weak for hard ones. ReCIRC reframes the threshold as a shared conditional-risk budget, then runs CRC on that transformed family. The authors say the guarantee holds even if the local risk estimates are poor, while accurate estimates can move it toward conditional risk control. In tests across synthetic and real segmentation, multilabel, multiclass, and regression settings, ReCIRC had the lowest average worst-group risk and mean positive group excess while keeping marginal risk near target. ArXiv · AI/CL/LG's note
score 4