Explanation-based Counterfactual Retraining(XCR): A Calibration Method for Black-box Models

Explanation methods can mislead models: The paper identifies a critical problem where popular explanation techniques (used to make AI decisions interpretable) can actually push AI models to behave in ways that don't align with real-world data or social values, a phenomenon called "social misalignment."

A new retraining approach improves model reliability: XCR works by using AI explanations as a form of feedback to retrain models, helping them perform better on out-of-distribution data (situations different from their training data). The method maintained performance while using only 12.5% of the most important features, suggesting models can be made more efficient and focused.

Robustness improvements without extra validation data: Unlike existing methods, XCR improved model robustness on corrupted or unusual data without requiring additional validation data for calibration, making it more practical to implement in real-world settings where such data may be scarce.

Addresses a fundamental tension in explainable AI: The paper highlights that post-hoc explanation methods (explanations added after a model is built) have inherent limitations in accurately representing how black-box models actually work, suggesting the field needs to reconsider how we validate and use AI explanations for improving model trustworthiness.