PCS-UQ: Uncertainty quantification via the Predictability-Computability-Stability framework.

PCS-UQ achieves consistent subgroup coverage and reduces prediction set sizes by 20% across multiple datasets, outperforming conformal methods in high-stakes machine learning.

• Why it matters: Reliable uncertainty quantification is crucial for safety in high-stakes domains where machine learning models are deployed. Existing methods may not guarantee coverage or adapt well to subgroups, risking unreliable predictions.
• What they did: The authors developed PCS-UQ, a framework that combines prediction checks, bootstrap sampling, and a calibration scheme, tested on 17 regression and 6 classification datasets, including deep learning variants.
• The result: PCS-UQ maintains target coverage, outperforms or matches conformal methods in interval width, and achieves significant reductions in prediction set sizes, enabling more efficient and trustworthy uncertainty estimates.