Spurious Correlation Robustness

This field studies how models can avoid relying on shortcuts—such as backgrounds, demographic attributes, or writing patterns—that correlate with labels in training data but may change at deployment. Common approaches include group-robust objectives and reweighting, counterfactual data generation, invariant or disentangled representations, and methods that identify shortcut features through explanations or model internals. Researchers evaluate these methods across vision, text, medical, and structured-data tasks, often using worst-group accuracy and out-of-distribution tests alongside average accuracy. A central challenge is reducing harmful reliance without discarding context that carries useful signal, especially when subgroup labels or representative validation data are scarce.