Learning the Pareto Frontier of Predictive Models under Distribution Shift

Researchers have published a new framework for analyzing the performance of pretrained models when they are applied to different data distributions. By identifying the Pareto frontier, the study provides a method to evaluate the trade-offs between model accuracy and reliability across varying task environments.