ReCBM: Uncertainty-Gated Relational Reasoning for Concept Bottleneck Models

2026-08-07

AI Summary

This work addresses the vulnerability of concept bottleneck models to unreliable concept states, which can propagate erroneous semantic evidence and degrade both interpretability and predictive performance. To mitigate this issue, the authors propose a novel approach that integrates semantic relationships—such as co-occurrence, entailment, and mutual exclusion—with an uncertainty-aware gating mechanism. This framework dynamically modulates the contribution of individual concepts during inference, enabling robust evidence propagation among concepts. Notably, it introduces uncertainty-guided relational reasoning into concept bottleneck models for the first time, facilitating stable inference and intervention even under missing or corrupted concepts, while automatically identifying a task-relevant subset of core concepts. Experiments demonstrate that the method significantly improves concept recovery and prediction accuracy across multiple datasets, exhibiting strong robustness under concept perturbations without compromising downstream task performance.

Abstract

Concept Bottleneck Models (CBMs) provide an interpretable framework by grounding predictions in human-understandable concepts, enabling semantic inspection and test-time intervention. Recent variants have improved CBMs through richer concept representations, uncertainty estimation, and dependency modeling. However, robust reasoning under unreliable concept states remains underexplored. Without such reasoning, misleading semantic evidence can propagate through the bottleneck, compromising both explanations and downstream predictions. To address this issue, we propose ReCBM, an uncertainty-gated relational reasoning framework for CBMs. ReCBM introduces semantically defined concept relations into the bottleneck and uses uncertainty to guide their refinement. By modeling co-occurrence, implication, and exclusion, ReCBM specifies how evidence is exchanged across concepts, while uncertainty modulates the contribution of each concept during this process. Experiments across diverse datasets showed that ReCBM improved concept and task recovery under missing and flipped concepts, supported uncertainty-aware intervention, and extracted compact task-relevant concept subsets without degrading downstream performance.

Problem

• Concept Bottleneck Models
• uncertainty
• relational reasoning
• semantic evidence
• concept reliability

Innovation

• Concept Bottleneck Models
• Uncertainty Estimation
• Relational Reasoning
• Interpretable AI
• Semantic Intervention