AI and the World's Hard Problems · Theoretical computer science · arXiv · 2026-10-23
Fragility-aware Classification for Risk Characterization and Robust Learning
Accuracy and AUC do not capture how large confident misjudgments are, which matters in safety-critical settings. The authors propose the Fragility Index, a risk-averse metric for the tail risk of such errors, along with a trainable framework and a regularizer for deep networks. On real medical diagnosis data, FI revealed error risk that accuracy and AUC missed, and FI-based models kept competitive accuracy with lower FI and lower average operational cost than standard empirical risk minimization in a simulated human–AI workflow.