Measuring explanation quality without needing the ground truth

Evaluating explanation methods in AI is hard because we lack ground-truth data to compare against. This work proposes a quantifiable metric based on continuous input perturbation that measures both sufficiency (what information the model needs) and necessity (what it doesn't) of attributions. The authors also introduce an adapter module trained using this metric as a differentiable supervision signal, which can be attached to any black-box model to produce causal explanations without reducing model accuracy. Experiments show the approach aligns better with human intuition and outperforms existing XAI techniques on several benchmarks.