HiCEMs add hierarchy to interpretable models

This is a meaningful upgrade for concept-based explainability because it tackles the two things that usually keep interpretable models niche: brittle flat concept vocabularies and expensive annotation pipelines.

• –HiCEMs extends classic Concept Embedding Models with explicit concept hierarchies, so explanations can move between coarse and fine-grained concepts instead of stopping at one label level
• –The Concept Splitting method is the real hook: it discovers human-interpretable sub-concepts from embedding space without requiring new annotations for every granularity
• –The paper claims gains beyond interpretability theater, showing that interventions at different concept levels can improve downstream task accuracy
• –The inclusion of a user study and the new PseudoKitchens 3D kitchen-render dataset makes this feel more grounded than a purely theoretical interpretability paper