Title: Mixture of Concept Bottleneck Experts

Abstract: Concept Bottleneck Models (CBMs) enhance interpretability by anchoring predictions in concepts that are understandable to humans. Nevertheless, conventional CBMs often restrict their task predictors to a single, pre-determined functional expression, which hampers both predictive performance and the ability to meet varied user requirements. To address this, we introduce Mixture of Concept Bottleneck Experts (M-CBE), a framework that expands the design space of CBMs across two axes: the quantity of expressions—termed experts—that the task predictor utilizes to map concepts to outcomes, and the specific functional structure of each expression. This approach reveals an under-investigated area of model design. We explore this space through the development of two new architectures: Linear M-CBE, which acquires a finite collection of linear expressions, and Symbolic M-CBE, which employs symbolic regression to identify expert functions from data, constrained by user-defined operator sets. Our empirical results show that adjusting both the number of expressions and their functional forms yields a resilient framework for balancing the trade-off between accuracy and interpretability.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC