From Interpretability Methods to Interpretable Models

Over a decade of explainable AI research in computer vision has assembled a mature arsenal of attribution maps, feature visualizations, and circuit analyses, yet the discipline has spent most of its energy benchmarking the tools rather than the networks themselves. This position paper proposes shifting the field's focus from interpretability methods to interpretable models. Drawing parallels to systems neuroscience, the authors advocate using existing diagnostic suites to systematically compare representations across architectures, while establishing rigorous empirical tests to measure whether independent evaluators can truly understand model behavior.