Uncertainty-aware Language Guidance for Concept Bottleneck Models

Researchers Yangyi Li and Mengdi Huai have introduced a novel method called 'Uncertainty-aware Language Guidance' to solve a critical problem in building interpretable AI systems known as Concept Bottleneck Models (CBMs). CBMs work by first identifying human-understandable concepts (like 'has stripes' or 'is metallic') before making a final prediction, making their reasoning transparent. However, labeling these concepts traditionally requires expensive expert annotation. While recent approaches use Large Language Models (LLMs) to generate concept labels automatically, they fail to account for the LLM's inherent uncertainty, leading to unreliable models prone to hallucination errors. The new method directly tackles this by providing a mathematically rigorous, distribution-free way to quantify how uncertain an LLM is about each concept it labels.

This technical advancement allows the CBM training process to weigh concept labels differently based on their calculated reliability. A concept the LLM is confident about gets more influence than one it's uncertain of, creating a more robust and accurate final model. The paper includes theoretical analysis and shows improved performance on real-world datasets. This work is a significant step toward practical, trustworthy, and interpretable AI, as it enables the automated creation of reliable CBMs without manual labeling, potentially accelerating their adoption in high-stakes fields like healthcare and finance where understanding an AI's decision is as important as its accuracy.