ReCBM: Uncertainty-Gated Relational Reasoning for Concept Bottleneck Models
A recent study presents ReCBM, a framework that employs uncertainty-gated relational reasoning for Concept Bottleneck Models (CBMs). These interpretable AI systems link predictions to concepts that are easily understood by humans, facilitating semantic analysis and intervention during testing. Although newer versions of CBMs have improved concept representations, uncertainty assessment, and dependency modeling, effective reasoning in the presence of unreliable concept states has not been thoroughly investigated. This gap can lead to misleading semantic information affecting explanations and subsequent predictions. ReCBM tackles this issue by integrating semantically defined relationships among concepts into the bottleneck and leveraging uncertainty to refine them. It captures co-occurrence, implication, and exclusion to clarify evidence exchange among concepts, with uncertainty influencing each concept's role. The paper can be found on arXiv with the identifier 2608.10004.
Key facts
- ReCBM is an uncertainty-gated relational reasoning framework for Concept Bottleneck Models.
- It introduces semantically defined concept relations into the bottleneck.
- It models co-occurrence, implication, and exclusion among concepts.
- Uncertainty modulates the contribution of each concept during reasoning.
- The framework addresses robust reasoning under unreliable concept states.
- The paper is available on arXiv with identifier 2608.10004.
- CBMs provide interpretable frameworks by grounding predictions in human-understandable concepts.
- Recent variants have improved CBMs through richer concept representations, uncertainty estimation, and dependency modeling.
Entities
Institutions
- arXiv