Reproducibility study on how to find Spurious Correlations, Shortcut Learning, Clever Hans or Group-Distributional non-robustness and how to fix them

Fragmented research community is addressing the same problem: Researchers across different fields (medical AI, autonomous driving, etc.) use different terminology—"spurious correlations," "shortcut learning," "Clever Hans effect," and "group-distributional non-robustness"—but are all trying to solve the same core issue: preventing AI models from relying on misleading patterns instead of causally relevant features.

XAI-based methods outperform traditional approaches: Explainable AI (XAI) techniques for fixing these problems generally work better than conventional methods, with Counterfactual Knowledge Distillation (CFKD) being the most consistently effective at helping models generalize better to new situations.

A critical practical barrier: the need for labeled group data: Most correction methods require manual annotation of which data subgroups exist and how they differ—a task that is often infeasible in real-world applications. Automated tools designed to identify these groups (like Spectral Relevance Analysis) currently struggle with complex scenarios, limiting the practical deployment of these solutions.

Limited data and imbalanced datasets make the problem harder: The study tested methods under realistic constraints including scarce training data and severe imbalances between subgroups, revealing that many promising techniques become less effective when these real-world challenges are present.