New CBM Improves Reasoning with Uncertain Concepts

A new research paper introduces ReCBM, an uncertainty-gated relational reasoning framework designed to enhance Concept Bottleneck Models (CBMs). CBMs are valuable for their interpretability, grounding predictions in human-understandable concepts, which allows for semantic inspection and intervention. However, their robustness can be compromised when concept states are unreliable, leading to misleading evidence propagation. ReCBM addresses this by integrating semantically defined concept relations directly into the bottleneck and using uncertainty to guide their refinement. It models various relationships, such as co-occurrence, implication, and exclusion, specifying how evidence is exchanged between concepts. Crucially, uncertainty modulates the contribution of each concept during this reasoning process, preventing unreliable semantic evidence from corrupting downstream predictions. Experiments across diverse datasets demonstrated that ReCBM significantly improved both concept and task recovery, even in scenarios with missing or flipped concepts. The framework also supported uncertainty-aware intervention and could extract compact, task-relevant concept subsets without degrading overall predictive performance, making CBMs more robust and reliable.