Label-Free Attribution for Interpretability

ICLR 2026 Conference Submission2654 Authors

Keywords: Interpretability, Attribution

Abstract: The importance of attribution algorithms in the AI field lies in enhancing model transparency, diagnosing and improving models, ensuring fairness, and increasing user understanding. Gradient-based attribution methods have become the most critical because of their high computational efficiency, continuity, wide applicability, and flexibility. However, current gradient-based attribution algorithms require the introduction of additional class information to interpret model decisions, which can lead to issues of information ignorance and extra information. Information ignorance can obscure important features relevant to the current model decision, while extra information introduces the incorrect identification of irrelevant features as significant. To address these issues, we propose the Label-Agnostic Attribution for Interpretability (LAAI) algorithm, which analyzes model decisions without the need for specified class information. Additionally, to more rigorously assess the potential of current attribution algorithms, we introduce a variety of new evaluation metrics, combined with the traditional Insertion \& Deletion Scores, to comprehensively assess the performance of our algorithm. To continuously advance research in the field of explainable AI (XAI), our algorithm is open-sourced at https://anonymous.4open.science/r/LFAI-336C

Primary Area: interpretability and explainable AI

Submission Number: 2654