EWMV: An algorithm to improve the efficiency of conformal methods

ICLR 2026 Conference Submission21488 Authors

Keywords: conformal prediction, aggregation, efficiency, validity, uncertainty quantification

TL;DR: An algorithm to improve the efficiency of conformal methods by leveraging calibration data and multiple available predictors.

Abstract: Conformal prediction is a framework that augments machine learning models to return a prediction set in lieu of a single prediction. Although, these sets commonly contain the correct answer with guaranteed probability, their size can be ineffectively large and thus lead to costly erroneous decisions. To mitigate this, we propose EWMV, an algorithm that leverages the available calibration data to aggregate multiple accessible predictors into a single, smaller conformal predictor. Empirical evidence across a variety of tasks and conformal methods suggests EWMV often produces smaller and more efficient prediction sets than any of the individual predictors being aggregated. Accordingly, these findings encourage a new paradigm to improve the efficiency of conformal methods with two readily available resources: calibration data and a plethora of pre-trained predictors.

Primary Area: probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)

Submission Number: 21488