Conformal prediction provides distribution-free coverage guarantees, but the size of its prediction regions depends heavily on the chosen conformity score. In multivariate regression, dependencies between output dimensions make this choice especially important. CoCoNet addresses this by learning a convex conformity score from held-out data, optimizing for prediction-region efficiency while enforcing convexity with respect to the residual.

The authors propose two variants: CoCoNet_Spline, which uses piecewise-linear scores, and CoCoNet_ICNN, based on Input Convex Neural Networks. Both are combined with split-conformal calibration to retain finite-sample coverage guarantees.

In experiments across 8 models and 30 multi-dimensional regression tasks with output dimensions from 2 to 16, CoCoNet_ICNN achieved the best average rank of 1.45 and ranked first on 22 of 30 tasks. CoCoNet_Spline-l1 followed with an average rank of 2.75, demonstrating the flexibility of the approach across different response types.