Rethinking Concept Bottleneck Models: From Pitfalls to Solutions

Merve Tapli; Quentin Bouniot; Wolfgang Stammer; Zeynep Akata; Emre Akbas

cs.CVMar 5, 2026

Rethinking Concept Bottleneck Models: From Pitfalls to Solutions

Authors: Merve Tapli, Quentin Bouniot, Wolfgang Stammer, Zeynep Akata, Emre Akbas

Organizations: Middle East Technical University (METU) · Helmholtz Munich · Munich Center for Machine Learning · Munich Data Science Institute ·

Abstract

Concept Bottleneck Models (CBMs) ground predictions in human-understandable concepts but face fundamental limitations: the absence of a metric to pre-evaluate concept relevance, the "linearity problem" causing recent CBMs to bypass the concept bottleneck entirely, an accuracy gap compared to opaque models, and finally the lack of systematic study on the impact of different visual backbones and VLMs. We introduce CBM-Suite, a methodological framework to systematically addresses these challenges. First, we propose an entropy-based metric to quantify the intrinsic suitability of a concept set for a given dataset. Second, we resolve the linearity problem by inserting a non-linear layer between concept activations and the classifier, which ensures that model accuracy faithfully reflects concept relevance. Third, we narrow the accuracy gap by leveraging a distillation loss guided by a linear teacher probe. Finally, we provide comprehensive analyses on how different vision encoders, vision-language models, and concept sets interact to influence accuracy and interpretability in CBMs. Extensive evaluations show that CBM-Suite yields more accurate models and provides insights for improving concept-based interpretability.

Figures & tables

Figure 1: Common issues among CBM approaches. (i) Models can achieve high accuracy even with irrelevant or random concepts due to concept leakage. (ii) Overlooking the importance of non-linearity reduces CBMs to mere linear probes. (iii) Introducing a concept bottleneck layer often leads to an accuracy gap compared to opaque models. (iv) Although CBMs are modular by design, evaluations do not consider different encoder choices.

Panel of Figure 2 — Figure 2: Linearity problem undermines interpretability. Linear CBMs attain high accuracy (85+% on ImageNet100 dataset) with random strings ( left ) and Roman Law terminology ( right ).

Panel of Figure 3 — Figure 3: Histogram of concept activations computed with SAIL for relevant and irrelevant concept sets on ImageNet100 and Places365, illustrating the difference in activation distributions between meaningful and random concepts.

Figure 4: Overview of the CBM-Suite, designed to address the key pitfalls of conventional Concept Bottleneck Models. First, the most relevant concept set for the dataset is selected using the Goodness of Concepts metric. Then, a non-linear concept encoder is trained to construct a trustworthy CBM. Next, the final classifier module is trained under teacher supervision. Finally, CBM-Suite provides a comprehensive evaluation suite across different vision encoders and vision-language models.

Task Agnostic $\downarrow$
Concept Set	IN100	CUB200	Places365
Relevant	3.660	4.305	3.077
Irrelevant	4.504	4.491	4.505
Random words	4.565	4.553	4.566
Task Specific $\downarrow$
Concept Set	IN100	CUB200	Places365

Table 1: Goodness of concepts : suitability of different concept sets for different datasets. Relevant concepts are GPT-generated in ImageNet100 (IN100) and Places365 and GT attributes in CUB200. The irrelevant concept set consists of Roman Law terminology. Random strings denote non-word concepts. Bold / underline indicate lowest/highest entropy (best/worst).

Panel of Figure 5 — Figure 5: Effect of non-linear concept encoders. We report classification accuracy on ImageNet100 and Places365 using both relevant and irrelevant concept sets. Linear CBM and Non-linear CBM differ only in the concept encoder layer, i.e . , the latter includes a non-linear function. Error bars represent the standard deviation over ten independent runs with random initializations.

Panel of Figure 6 — Figure 6: Effect of knowledge distillation on CBM performance. Classification accuracy on ImageNet100 and Places365 datasets. Distilled CBM significantly outperforms Vanilla CBM by incorporating distillation loss alongside the elastic loss, approaching Oracle performance. Error bars represent the standard deviation over ten independent runs with random initializations.

Backbone	LP	VLM
		CLIP [ 21 ]	SAIL [ 35 ]	FLAIR [ 30 ]	SigLIP [ 34 ]
ResNet [ 8 ]	91.30	89.16	85.56	87.00	90.12
DINOv2 [ 3 ]	92.72	92.22	89.28	89.68	91.74
Franca [ 27 ]	94.10	93.66	88.40	90.08	93.68
Perc. Enc. [ 1 ]	96.56	95.84	93.56	93.84	96.14

Table 2: Backbone choice matters more than VLM selection. Performance accuracies (%) of different VLM and vision backbones on ImageNet100 dataset. Perc.Enc. consistently outperforms other backbones across all methods with performance generally varying more across backbones (rows) than VLMs (columns). SigLIP performs best among VLMs in 3 of 4 cases, with the Perc.Enc. + SigLIP combination yielding best overall performance.

Models	CUB200 [ 28 ]	CIFAR100 [ 12 ]	Places365 [ 36 ]	ImageNet [ 6 ]
LaBo	81.00%	86.82%	-	83.97%
LFCBM	74.31%	65.13%	43.68%	71.95%
CBM-Suite	86.73%	92.50%	54.64%	81.56%

Table 3: Quantitative comparison with state-of-the-art CBMs. Results for LaBo and LFCBM are taken from their respective papers. CBM-Suite uses the best-performing vision backbone–VLM pair for each dataset.

Panel of Figure 7 — Figure 7: Top concepts are well-matched with their respective classes. Illustration of the Top-5 concepts, ranked by their contribution to the final prediction for randomly selected classes, along with representative images from ImageNet, CIFAR100, CUB200, and Places365, demonstrating that the selected backbone–VLM pairs capture meaningful semantic structure across domains.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Concept Encoder		Classifier
	$\eta$	Scheduler	$\alpha$	$\beta$	Temperature	$\eta$	Scheduler
ImageNet100	0.001	✓	1	1	2	0.0001	-
CUB200	0.0001	-	1	0.5	4	0.001	✓
Places365	0.0001	-	1	2	2	0.001	✓
CIFAR100	0.001	✓	1	1	4	0.0001	-

Appendix

Table 4: Hyperparameter settings for each dataset, including learning rates (

\eta

) and scheduler usage for the concept encoder and classifier, along with the

\alpha

and

\beta

loss weights and distillation temperature Identical settings are used across all vision backbones and VLMs.

Roman Law Concepts		Random Words
Jurisprudential	historians	gsankm	sjklrveogjc
administrative	impede	ehxrpq	pdbhhidbnvjig
bureaucratization	individual	swsbkzguqjil	cygmao
centuries	iudicum	odbhhqr	foqmegkdcp
civil	knowledge	rwkkmbezusn	fdqumo
delegation	latitude	qkfns	hkeeyhw

Appendix

Table 5: Sample of 30 concepts from the two concept sets used in our experiments. Roman Law concepts extracted from the content of the relevant Wikipedia page, and a set of randomly generated strings.

Panel of Figure 8 — Figure 8: Histogram of concept activations computed with different VLMs for the relevant and irrelevant concept sets on ImageNet100, highlighting differences in activation distributions across models when evaluated on meaningful versus random concepts.

Task Agnostic $\downarrow$
Concept Set	Top-10	Top-100	Top-250
Relevant	1.751	3.660	4.521
Irrelevant	2.254	4.504	5.376
Random	2.284	4.565	5.465

Appendix

Table 6: Robustness of the goodness of concepts metric to varying cutoff values on ImageNet100.

ImageNet100
Concept Set	CLIP	SAIL	FLAIR	SigLIP
Relevant	4.479	3.660	4.545	4.494
Irrelevant	4.480	4.504	4.511	4.534
Random words	4.555	4.565	4.565	4.522

Appendix

Table 7: Task-agnostic goodness of concepts: suitability of different concept sets for different datasets. Relevant concepts are GPT-generated in ImageNet100 and Places365 and ground-truth attributes in CUB200. The irrelevant concept set consists of Roman Law terminology. Random strings denote non-word concepts. Lower entropy indicates better alignment.

ImageNet100
Concept Set	CLIP	SAIL	FLAIR	SigLIP
Relevant	4.517	3.738	4.578	4.522
Irrelevant	4.529	4.545	4.572	4.559
Random words	4.576	4.579	4.589	4.556

Appendix

Table 8: Task-specific goodness of concepts: suitability of different concept sets for different datasets. Relevant concepts are GPT-generated in ImageNet100 and Places365 and ground-truth attributes in CUB200. The irrelevant concept set consists of Roman Law terminology. Random strings denote non-word concepts. Lower entropy indicates better alignment.

Panel of Figure 9 — Figure 9: Concept set refinement. Entropy-guided removal (blue) outperforms random removal (red, averaged over 10 trials) and preferentially eliminates random concepts, demonstrating the effectiveness of the goodness-of-concepts metric. Solid lines correspond to the entropy (left y-axis), while dashed lines correspond to the number of removed random words (right y-axis).

Configuration	Accuracy (%)
Linear	87.96
Linear + Distillation	90.37
Non-linear	82.27
Non-linear + Distillation	85.56

Appendix

Table 9: Ablation study disentangling the effects of linearity and knowledge distillation.

Panel of Figure 10 — Figure 10: Illustration of the Top-5 concepts, ranked by their contribution to the final prediction for randomly selected classes, along with representative images from ImageNet100, CIFAR100, CUB200, and Places365, highlighting model interpretability across diverse domains.

Concept Bottleneck Models

Concept Bottleneck Models route predictions through human-understandable concepts that people can inspect, correct, or intervene on.

See all 82 papers

Caption Bottleneck Models
Jul 1, 2026Seref Baris Cagliyan, Umut Ozdemir, Merve Tapli +1Vision-Language Models Concept Bottleneck Models
Dept. of Computer Eng., Middle East Technical University (METU) · Robotics & AI Center (ROMER), METU
Learning Concept Bottleneck Models from Mechanistic Explanations
Mar 7, 2026Antonio De Santis, Schrasing Tong, Marco Brambilla +1Concept Bottleneck Models Explainability Evaluation
Politecnico di Milano · MIT CSAIL
Concepts' Information Bottleneck Models
Feb 16, 2026Karim Galliamov, Syed M Ahsan Kazmi, Adil Khan +1Concept Bottleneck Models Information Bottleneck
University of Amsterdam · University of West of England · University of Hull +1

Jul 1, 2026J/K move · Enter open · S save

Caption Bottleneck Models

Seref Baris Cagliyan, Umut Ozdemir, Merve Tapli, Emre Akbas

Dept. of Computer Eng., Middle East Technical University (METU) · Robotics & AI Center (ROMER), METU

Vision-Language Models Concept Bottleneck Models Interpretable ML

Open paper AskPDF arXiv

Concept Bottleneck Models (CBMs) provide interpretability by routing predictions through a layer of human-understandable concepts. However, defining an optimal concept set for a specific dataset remains an open challenge. Existing approaches rely on expensive expert annotations or LLM-generated lists based solely on class names. Even "open-vocabulary" variants typically depend on static concept sets, which restrict discovery and introduce label bias. Furthermore, traditional CBMs often suffer from information leakage, where unmodeled visual features bypass the bottleneck and compromise the integrity of the explanations. To overcome these limitations, we propose Caption Bottleneck Models (CaBM), a framework that circumvents the need for predefined concept sets by replacing rigid concept layers with free-form natural language. By representing images via LMM-generated captions and training a classifier strictly on this text, CaBM ensures a leakage-free architecture by construction. Additionally, by analyzing the text classifier post-training, CaBM autonomously discovers high-quality, dataset-specific concepts. Our results across fine- and coarse-grained benchmarks demonstrate that CaBM achieves competitive accuracy while preserving interpretability without the constraints of external dictionaries or manual labeling.

Abstract

Concept Bottleneck Models

Explore similar work

Caption Bottleneck Models

Learning Concept Bottleneck Models from Mechanistic Explanations

Concepts' Information Bottleneck Models