Evaluating Robustness of Concept Bottleneck Models
October 1, 2026
This study introduces a generator-based framework to evaluate Concept Bottleneck Models (CBMs) under continuous geometric and discrete semantic perturbations. The research reconciles conflicting findings regarding CBM robustness by distinguishing between different perturbation regimes.
HOW THIS AFFECTS YOU
●
researcherYou can use this framework to perform more rigorous, controlled robustness testing on interpretable models.