Building trust in AI requires making "black-box" models transparent, but today's most popular post-hoc explanation methods share a critical flaw: they rely on random sampling that leads to unstable, fluctuating explanations. To solve this, we developed Expected Active Gain for Local Explanations (EAGLE). EAGLE reimagines how we unveil the model's reasoning by leveraging active learning. Rather than randomly probing the model, EAGLE intelligently queries the exact data points that maximize information gain, directly shrinking the surrogate model's epistemic uncertainty, without losing focus on the specific prediction at hand. Backed by rigorous mathematical guarantees for sample complexity, our framework drastically outperforms state-of-the-art baselines by delivering highly stable explanations, faster convergence, and tighter confidence intervals across both image and tabular datasets. Ultimately, EAGLE offers a robust, principled path forward for truly trustworthy Explainable AI.

Read the full work here: arxiv.org/pdf/2603.14894