What’s in the Bottle? A Survey and Roadmap of Concept Bottleneck Models

Abstract:

Concept Bottleneck Models (CBMs) are interpretable learning architectures that factor predictions through intermediate, ideally human-understandable concepts, enabling explicit and inspectable reasoning. Although CBM research has gained substantial momentum in
recent years, this growth has also revealed numerous open challenges and a fragmented set of methodological choices. In this work, we systematically review the CBM literature, identify previously unidentified core components and challenges, and propose a unified taxonomy. Based on this taxonomy, we provide a detailed categorization of existing works. We hereby discuss current challenges for the CBM paradigm and outline important directions to extend it beyond its current scope. Overall, this survey aims to consolidate the CBM landscape, clarify open issues, and provide guidance for developing future models.

Submission Type: Long submission (more than 12 pages of main content)

Changes Since Last Submission:

## Changes Since Last Submission

We revised the manuscript in response to feedback from all three reviewers. All changes are highlighted in blue in the revised manuscript. The main changes are:

- Added a dedicated paragraph on causal representation learning (Sec. 6.2) and an explicit connection to epistemic uncertainty (Sec. 5.1) and knowledge representation.
- Added discussion of zero-shot generalization in the generalist/specialist section and a connection to System 1 & 2 thinking.
- Integrated suggested neurosymbolic AI references across Secs. 5.1, 5.2, 6.1, and 6.2.
- Expanded Sec. 3 to more explicitly compare our taxonomy to prior surveys
- Added Sec. 4.5 "Summary of Observed Trends" synthesizing key patterns from the taxonomy.
- Split the evaluation section into "Performance Evaluation and Benchmarking" and "Evaluating Interpretability," with a new discussion on the absence of a shared interpretability definition.
- Added clarification that the paper focuses on conceptual synthesis rather than empirical benchmarking, and that standardized evaluation protocols for CBMs are still lacking.

Assigned Action Editor: ~Adin_Ramirez_Rivera1

Submission Number: 7700