Explanation Stability

Explanation stability studies whether a model gives consistent feature rankings, token attributions, or saliency maps when inputs are perturbed, models are retrained, or conditions change. Researchers measure consistency with metrics such as rank correlation, overlap, and attribution similarity, and test methods including SHAP, LIME, CAM, and gradient-based explanations across vision, language, graphs, and structured data. Other approaches train models to produce steadier explanations or align explanations across transformed inputs, while benchmarks examine shifts such as test-time adaptation and temporal change. A central tension is that stability is distinct from predictive accuracy and explanation faithfulness, and can depend on the particular model–method pair rather than the model alone.