mlsanityData to 14 Oct 2026

Leakage and Interpretability in Concept-Based Models

arXiv 2504.14094 ↗First posted 18 Apr 2025cs.LG cs.AI stat.MLcited by 13

Not on the map. Why?

Abstract (arXiv)

Concept-based Models aim to improve interpretability by predicting high-level intermediate concepts, representing a promising approach for deployment in high-risk scenarios. However, they are known to suffer from information leakage, whereby models exploit unintended information encoded within the learned concepts. We introduce an information-theoretic framework to rigorously characterise and quantify leakage, and define two complementary measures: the concepts-task leakage (CTL) and interconcept leakage (ICL) scores. We show that these measures are strongly predictive of model behaviour under interventions and outperform existing alternatives. Using this framework, we identify the primary causes of leakage and, as a case study, analyse how it manifests in Concept Embedding Models, revealing interconcept and alignment leakage in addition to the concepts-task leakage present by design. Finally, we present a set of practical guidelines for designing concept-based models to reduce leakage and ensure interpretability.

Key facts. Cited by 13 papers in the map (data through 14 Oct 2026). Builds on 44 of its 69 references that are in the map. First posted on arXiv 18 Apr 2025.

Builds on

44 of 69 references.

References · 44

  1. 2021Learning Transferable Visual Models From Natural Language Supervisioncited by 33,744
  2. 2016Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localizationcited by 6,733
  3. 2017A Unified Approach to Interpreting Model Predictionscited by 6,555
  4. 2016"Why Should I Trust You?": Explaining the Predictions of Any Classifiercited by 5,070
  5. 2015Learning Deep Features for Discriminative Localizationcited by 2,436
  6. 2020Concept Bottleneck Modelscited by 1,032
  7. 2018This Looks Like That: Deep Learning for Interpretable Image Recognitioncited by 737
  8. 2022Language in a Bottle: Language Model Guided Concept Bottlenecks for Interpretable Image Classificationcited by 312
  9. 2022Post-hoc Concept Bottleneck Modelscited by 293
  10. 2023Label-Free Concept Bottleneck Modelscited by 285
  11. 2019On Completeness-aware Concept-Based Explanations in Deep Neural Networkscited by 239
  12. 2021Promises and Pitfalls of Black-Box Concept Learning Modelscited by 131
32 more
  1. 2021Do Concept Bottleneck Models Learn as Intended?cited by 116
  2. 2023Probabilistic Concept Bottleneck Modelscited by 101
  3. 2022B-cos Networks: Alignment is All We Need for Interpretabilitycited by 88
  4. 2022Concept Bottleneck Model with Additional Unsupervised Conceptscited by 79
  5. 2022Interactive Concept Bottleneck Modelscited by 79
  6. 2022GlanceNets: Interpretabile, Leak-proof Concept-based Modelscited by 73
  7. 2024VLG-CBM: Training Concept Bottleneck Models with Vision-Language Guidancecited by 65
  8. 2023A Closer Look at the Intervention Procedure of Concept Bottleneck Modelscited by 64
  9. 2022Concept-level Debugging of Part-Prototype Networkscited by 52
  10. 2024Stochastic Concept Bottleneck Modelscited by 51
  11. 2022Overlooked factors in concept-based explanations: Dataset choice, concept learnability, and human capabilitycited by 46
  12. 2023Learning to Receive Help: Intervention-Aware Concept Embedding Modelscited by 41
  13. 2023Beyond Normal: On the Evaluation of Mutual Information Estimatorscited by 37
  14. 2023Auxiliary Losses for Learning Generalizable Concept-based Modelscited by 35
  15. 2024Energy-Based Concept Bottleneck Models: Unifying Prediction, Concept Intervention, and Probabilistic Interpretationscited by 33
  16. 2023Towards Robust Metrics for Concept Representation Evaluationcited by 29
  17. 2022Understanding and Enhancing Robustness of Concept-based Modelscited by 22
  18. 2024On the Concept Trustworthiness in Concept Bottleneck Modelscited by 22
  19. 2024Semi-supervised Concept Bottleneck Modelscited by 21
  20. 2024Approximating mutual information of high-dimensional variables using learned representationscited by 21
  21. 2021Is Disentanglement all you need? Comparing Concept-based & Disentanglement Approachescited by 17
  22. 2024Concept Bottleneck Language Models For protein designcited by 15
  23. 2025Measuring Leakage in Concept-Based Methods: An Information Theoretic Approachcited by 9
  24. 2024Tree-Based Leakage Inspection and Control in Concept Bottleneck Modelscited by 8
  25. 2017Geometric k-nearest neighbor estimation of entropy and mutual informationcited by 7
  26. 2024Counterfactual Concept Bottleneck Modelscited by 7
  27. 2024Evidential Concept Embedding Models: Towards Reliable Concept Explanations for Skin Disease Diagnosiscited by 7
  28. 2022Towards learning to explain with concept bottleneck models: mitigating information leakagecited by 6
  29. 2024EQ-CBM: A Probabilistic Concept Bottleneck with Energy-based Models and Quantized Vectorscited by 6
  30. 2025Birds look like cars: Adversarial analysis of intrinsically interpretable deep learningcited by 3
  31. 2025V-CEM: Bridging Performance and Intervenability in Concept-based Modelscited by 3
  32. 2024HuLP: Human-in-the-Loop for Prognosiscited by 1

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2025: 5; 2026: 8.

13 papers.

Sort:

By area:

  1. 2025There Was Never a Bottleneck in Concept Bottleneck Modelscited by 8
  2. 2025Towards Reasonable Concept Bottleneck Modelscited by 3
  3. 2026Rethinking Concept Bottleneck Models: From Pitfalls to Solutionscited by 2
  4. 2026Matryoshka Concept Bottleneck Modelscited by 2
  5. 2025Parameter-Efficient Construction of the Rashomon Slice for Concept Bottleneck Modelscited by 1
  6. 2026Hoeffding Concept Bottleneck Models with Applications to Overhead Imagescited by 1
  7. 2025Improving ARDS Diagnosis Through Context-Aware Concept Bottleneck Modelscited by 0
  8. 2025Uncertainty-Aware Concept Bottleneck Models with Enhanced Interpretabilitycited by 0
  9. 2026Towards Faithful Multimodal Concept Bottleneck Modelscited by 0
  10. 2026Metonymy in vision models undermines attention-based interpretabilitycited by 0
  11. 2026Measuring What Matters: Synthetic Benchmarks for Concept Bottleneck Modelscited by 0
  12. 2026In Defense of Information Leakage in Concept-based Modelscited by 0
1 more
  1. 2026Concept Flow Models: Anchoring Concept-Based Reasoning with Hierarchical Bottleneckscited by 0