Uncertainty Quantification

Uncertainty quantification studies what is unknown in a prediction or inference, where that uncertainty comes from, and how it should affect decisions. Common approaches include Bayesian methods, ensembles and Gaussian processes, which model predictive distributions; conformal prediction, which provides distribution-free coverage guarantees; and calibration methods that adjust confidence or intervals against observed errors. Researchers apply these tools to settings from medical imaging and physical simulations to language models, evaluating calibration, coverage, error detection and performance under distribution shift. A central challenge is that well-calibrated or narrow estimates do not always identify actual failures, so methods must be tested in the conditions and for the decisions where they will be used, while balancing reliability against computational cost.