MAnchors: Memorization-Based Acceleration of Anchors via Rule Reuse and Transformation
Published in International Conference on Machine Learning (ICML), 2026
Anchors is a widely used local model-agnostic explanation technique, but its sampling-heavy procedure can be too slow for time-sensitive applications. This work introduces a memorization-based acceleration strategy that stores reusable intermediate rules and adapts them to new inputs through horizontal and vertical transformations.
The framework reduces explanation generation time across tabular, text, image, and LLM settings while maintaining the fidelity and understandability of the original Anchors algorithm.
Recommended citation: Haonan Yu et al. (2026). "MAnchors: Memorization-Based Acceleration of Anchors via Rule Reuse and Transformation." International Conference on Machine Learning (ICML).
Download Paper