Survey and roadmap · TMLR 2026

What’s in the bottle?

A CBM survey and research guide to Concept Bottleneck Models (CBMs): their architectural choices, literature, and the open problems that determine whether concept-based reasoning is interpretable and reliable.

Explore the literature See the framework

Purpose of the surveyCBM papers differ in their input representation, concept source, prediction task, and training procedure. The taxonomy makes these choices explicit.

Perform the prediction through an intermediate concept representation.

CBM pipeline showing input modality, input encoder, embeddings, concept source with semantics and grounding, activations, predictor, and output — Paper figure: the pipeline separates the four modules used in the taxonomy.

inspect / intervene

Concept Bottleneck Models.

A CBM maps an input x to human-understandable concepts c and uses those concepts to produce the output y. This explicit interface supports inspection of the predicted concepts, interventions that replace concept values and steer the output, and debugging when the concept representation or prediction is wrong.

CBM figure showing concept explanations, a bed concept changed from 0.8 to 0.0, the output changing to living room, and model debugging — Inspection, intervention, steering, and debugging at the concept interface.

Taxonomy

Architectural choices in four modules.

The survey separates input, concept, output, and training modules. Within the concept module, it distinguishes semantics—which concepts are used and what they mean—from grounding—how those concepts are detected and instantiated in the input.

Taxonomy of CBM input, concept, output, and training modules — Supplied taxonomy figure · the literature matrix below uses these categories

Categorization

A structured view of the CBM literature.

Each row assigns a named primary work to the taxonomy dimensions used in the survey. Search and compare modalities, concept sources, representations, predictors, tasks, and training strategies.

		Input module		Concept module			Output module		Training
		Modality	Encoder / embedding	Semantics	Grounding	Activations	Predictor	Task	Strategy
Concept Bottleneck Models ↗Koh et al.	2020	Image	CNNStructured	GT labels	GT labels	Continuous	Linear	Discriminative	Joint
Promises and Pitfalls of Black-Box Concept Learning Models ↗Mahinpei et al. · ICML XAI workshop	2021	Image	CNNStructured	GT labels	GT labels	Continuous	Linear	Discriminative	Joint
Addressing Leakage in Concept Bottleneck Models ↗Havasi et al.	2022	Image	CNNStructured	GT labels	GT labels	Continuous	Linear	Discriminative	Joint
Concept Embedding Models ↗Espinosa Zarlenga et al.	2022	Image	CNNStructured	GT labels	GT labels	Continuous	MLP	Discriminative	Joint
Label-free Concept Bottleneck Models ↗Oikarinen et al.	2023	Image	CLIPUnstructured	VLM-based	VLM-based	Continuous	Linear	Discriminative	Joint
Post-hoc Concept Bottleneck Models ↗Yüksekgönül et al.	2023	Image	CNNStructured	Mixed	VLM-based	Continuous	Linear	Discriminative	Sequential
Intervenable and editable CBMs ↗A growing line of work	2023+	Mixed	TransformerUnstructured	LLM-based	VLM-based	Cont. & disc.	MLP	Disc. & gen.	Joint
Discovering concepts from data ↗Prototype-, clustering-, and VLM-guided methods	2023+	Image	VLMUnstructured	VLM-based	Mixed	Continuous	Linear	Discriminative	Joint
Beyond complete concept sets ↗Incomplete, probabilistic, and structured CBMs	2023+	Image	TransformerStructured	GT labels	GT labels	Distribution	MLP	Discriminative	Joint
Evaluating concept-based reasoning ↗Benchmarks and diagnostic protocols	2024+	Mixed	Misc.	Mixed	Mixed	Cont. & disc.	Misc.	Discriminative	Varies
Concept-based XAI beyond CBMs ↗Related surveys and position papers	2024+	Mixed	Misc.	Mixed	Mixed	Cont. & disc.	Misc.	Disc. & gen.	Varies
What’s in the Bottle? A Survey and Roadmap ↗Knab, Steinmann, Bartelt et al.	2026	Mixed	Misc.	Mixed	Mixed	Cont. & disc.	Misc.	Disc. & gen.	Varies
A closer look at the intervention procedure of concept bottleneck models ↗Shin et al. · ICML	2023	Image	CNNMisc.	Human-in-the-loop	Human-in-the-loop	Continuous	Linear	Discriminative	Joint
Beyond concept bottleneck models: How to make black boxes intervenable? ↗Laguna et al. · NeurIPS	2024	Image	CNNMisc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Joint
Concept bottleneck generative models ↗Ismail et al. · ICLR	2024	Mixed	Misc.Structured	GT Labels	GT Labels	Continuous	Misc.	Generative	Joint
Concept bottleneck model with additional unsupervised concepts ↗Sawada et al. · IEEE Access	2022	Image	CNNMisc.	Mixed / discovered	Mixed / discovered	Continuous	Linear	Discriminative	Joint
Discover-then-name: Task-agnostic concept bottlenecks via automated concept discovery ↗Rao et al. · ECCV	2024	Image	CNNMisc.	Mixed / discovered	Mixed / discovered	Continuous	Linear	Discriminative	Joint
Do concept bottleneck models learn as intended? ↗Margeloiu et al. · arXiv	2021	Image	CNNMisc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Joint
Energy-based concept bottleneck models: Unifying prediction, concept intervention, and probabilistic interpretations ↗Xu et al. · ICLR	2024	Image	CNNMisc.	Human-in-the-loop	Human-in-the-loop	Distribution	Linear	Discriminative	Joint
Incremental residual concept bottleneck models ↗Shang et al. · CVPR	2024	Image	CNNMisc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Sequential
Interpretable concept-based memory reasoning ↗Debot et al. · NeurIPS	2024	Image	CNNMisc.	GT Labels	GT Labels	Continuous	MLP	Discriminative	Joint
Interpretable neural-symbolic concept reasoning ↗Barbiero et al. · ICML	2023	Image	GNNMisc.	Disentanglement	Disentanglement	Continuous	Symbolic	Discriminative	Joint
Interpreting clip with sparse linear concept embeddings (splice) ↗Bhalla et al. · NeurIPS	2024	Image	VLMMisc.	VLM-based	VLM-based	Continuous	MLP	Discriminative	Joint
Language in a bottle: Language model guided concept bottlenecks for interpretable image classification ↗Yang et al. · CVPR	2023	Image	CLIPVLM	LLM-based	VLM-based	Continuous	Linear	Discriminative	Joint
Learning to intervene on concept bottlenecks ↗Steinmann et al. · ICML	2024	Image	CNNMisc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Joint
Object Centric Concept Bottlenecks ↗Steinmann et al. · MICCAI	2025	Image	CNNMisc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Joint
Relational confcept bottleneck models ↗Barbiero et al. · NeurIPS	2024	Image	GNNMisc.	Disentanglement	Disentanglement	Continuous	Linear	Discriminative	Joint
Stochastic concept bottleneck models ↗Vandenhirtz et al. · NeurIPS	2024	Image	CNNMisc.	GT Labels	GT Labels	Distribution	Linear	Discriminative	Joint
TabCBM: Concept-based Interpretable Neural Networks for Tabular Data ↗Espinosa Zarlenga et al. · TMLR	2023	Tabular	CNNMisc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Joint
Adacbm: An adaptive concept bottleneck model for explainable and accurate diagnosis ↗Chowdhury et al. · MICCAI	2024	Image	CNNMisc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Joint
An Analysis of Concept Bottleneck Models: Measuring, Understanding, and Mitigating the Impact of Noisy Annotations ↗Park et al. · NeurIPS	2025	Image	CNNMisc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Joint
Bayesian concept bottleneck models with llm priors ↗Feng et al. · NeurIPS	2025	Text	LLMMisc.	LLM-based	LLM-based	Distribution	LLM	Discriminative	Joint
PCBEAR: Pose Concept Bottleneck for Explainable Action Recognition ↗Lee et al. · CVPR	2025	Video	CNNMisc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Joint
Coarse-to-fine concept bottleneck models ↗Panousis et al. · NeurIPS	2024	Image	CNNMisc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Joint
Concept Bottleneck Large Language Models ↗Sun et al. · ICLR	2025	Text	LLMMisc.	LLM-based	LLM-based	Continuous	LLM	Generative	Joint
Concept bottleneck with visual concept filtering for explainable medical image classification ↗Kim et al. · MICCAI	2023	Image	CNNMisc.	VLM-based	VLM-based	Continuous	Linear	Discriminative	Joint
Concepts in Motion: Temporal Bottlenecks for Interpretable Video Classification ↗Knab et al. · ICML 2026 workshop	2026	Video	CLIPUnstructured	VLM-based	VLM-based	Continuous	Misc.	Discriminative	Sequential
CONDA: Adaptive Concept Bottleneck for Foundation Models Under Distribution Shifts ↗Choi et al. · ICLR	2025	Mixed	CNNMisc.	GT Labels	GT Labels	Distribution	Linear	Discriminative	Joint
Counterfactual concept bottleneck models ↗Dominici et al. · ICLR	2025	Image	CNNMisc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Joint
DCBM: Data-Efficient Visual Concept Bottleneck Models ↗Prasse et al. · ICML	2025	Image	CNNMisc.	VLM-based	VLM-based	Continuous	Linear	Discriminative	Joint
Measuring leakage in concept-based methods: An information theoretic approach ↗Makonnen et al. · arXiv	2025	Image	CNNMisc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Joint
FaCT: Faithful Concept Traces for Explaining Neural Network Decisions ↗Parchami-Araghi et al. · NeurIPS	2025	Image	CNNMisc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Joint
Deferring Concept Bottleneck Models: Learning to Defer Interventions to Inaccurate Experts ↗Pugnana et al. · NeurIPS	2025	Image	CNNMisc.	Human-in-the-loop	Human-in-the-loop	Continuous	Linear	Discriminative	Joint
Graph integrated multimodal concept bottleneck model ↗Lin et al. · arXiv	2025	Mixed	GNNMisc.	GT Labels	GT Labels	Continuous	GNN	Discriminative	Joint
Disentangled Concepts Speak Louder Than Words: Explainable Video Action Recognition ↗Lee et al. · NeurIPS	2025	Video	TransformerMisc.	Disentanglement	Disentanglement	Continuous	Linear	Discriminative	Joint
Graph of thoughts: Solving elaborate problems with large language models ↗Besta et al. · AAAI	2024	Text	LLMMisc.	LLM-based	LLM-based	Continuous	GNN	Generative	Joint
Tree of thoughts: Deliberate problem solving with large language models ↗Yao et al. · NeurIPS	2023	Text	LLMMisc.	LLM-based	LLM-based	Continuous	LLM	Generative	Joint
Object-centric learning with slot attention ↗Locatello et al. · NeurIPS	2020	Image	TransformerMisc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Joint
Chain-of-thought prompting elicits reasoning in large language models ↗Wei et al. · NeurIPS	2022	Text	LLMMisc.	LLM-based	LLM-based	Continuous	LLM	Generative	Joint
The information bottleneck method ↗Tishby et al. · arXiv	2000	Image	CNNMisc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Joint
Transparency and the black box problem: Why we do not trust AI ↗Von Eschenbach · Philosophy & technology	2021	Other	Misc.Misc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Joint
Attention is all you need ↗Vaswani et al. · NeurIPS	2017	Other	TransformerMisc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Joint
Hi Robot: Open-Ended Instruction Following with Hierarchical Vision-Language-Action Models ↗Shi et al. · ICML	2025	Text	VLMMisc.	VLM-based	VLM-based	Continuous	Linear	Discriminative	Joint
Vision-Language-Action Model with Open-World Embodied Reasoning from Pretrained Knowledge ↗Zhou et al. · arXiv	2025	Text	VLMMisc.	VLM-based	VLM-based	Continuous	Linear	Discriminative	Joint
Object-Centric Representation Learning for Enhanced 3D Semantic Scene Graph Prediction ↗Heo et al. · NeurIPS	2025	Image	GNNMisc.	Disentanglement	Disentanglement	Continuous	GNN	Discriminative	Joint
Learning Object-Centric Representations of Multi-Object Scenes from Multiple Views ↗Li et al. · NeurIPS	2020	Image	CNNMisc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Joint
Erasing Concepts from Diffusion Models ↗Gandikota et al. · Proceedings of the IEEE/CVF	2023	Image	CNNMisc.	GT Labels	GT Labels	Continuous	Linear	Generative	Joint
Ablating Concepts in Text-to-Image Diffusion Models ↗Kumari et al. · Proceedings of the IEEE/CVF	2023	Image	CNNMisc.	GT Labels	GT Labels	Continuous	Linear	Generative	Joint
Do Concept Bottleneck Models Obey Locality? ↗Raman et al. · XAI in Action: Past, Present	2023	Image	CNNMisc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Joint
Locality-aware Concept Bottleneck Model ↗Jeon et al. · UniReps: 2nd Edition of the	2024	Image	CNNMisc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Joint
From Segments to Concepts: Interpretable Image Classification via Concept-Guided Segmentation ↗Eisenberg et al. · arXiv	2025	Image	CNNMisc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Joint
Hybrid Concept Bottleneck Models ↗Liu et al. · CVPR	2025	Image	CNNMisc.	GT Labels	GT Labels	Cont. & Disc.	Linear	Discriminative	Joint
Interactive concept bottleneck models ↗Chauhan et al. · AAAI	2023	Image	CNNMisc.	Human-in-the-loop	Human-in-the-loop	Continuous	Linear	Discriminative	Joint
LogicCBMs: Logic-Enhanced Concept-Based Learning ↗Vemuri et al. · arXiv	2025	Image	GNNMisc.	Disentanglement	Disentanglement	Discrete	Symbolic	Discriminative	Joint
Partially Shared Concept Bottleneck Models ↗Zhao et al. · arXiv	2025	Image	CNNMisc.	GT Labels	GT Labels	Cont. & Disc.	Linear	Discriminative	Joint
Selective Concept Bottleneck Models Without Predefined Concepts ↗Schrodi et al. · TMLR	2025	Image	CNNMisc.	Mixed / discovered	Mixed / discovered	Continuous	Linear	Discriminative	Joint
Show and Tell: Visually Explainable Deep Neural Nets via Spatially-Aware Concept Bottleneck Models ↗Benou et al. · CVPR	2025	Image	CNNMisc.	VLM-based	VLM-based	Continuous	Linear	Discriminative	Joint
Uncertainty-Aware Concept Bottleneck Models with Enhanced Interpretability ↗Zhang et al. · arXiv	2025	Image	CNNMisc.	GT Labels	GT Labels	Distribution	Linear	Discriminative	Joint
V2C-CBM: Building Concept Bottlenecks with Vision-to-Concept Tokenizer ↗He et al. · AAAI	2025	Image	CNNMisc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Joint
Vlg-cbm: Training concept bottleneck models with vision-language guidance ↗Srivastava et al. · NeurIPS	2024	Text	VLMMisc.	VLM-based	VLM-based	Continuous	Linear	Discriminative	Joint
Atlas-Alignment: Making Interpretability Transferable Across Language Models ↗Puri et al. · arXiv	2025	Text	LLMMisc.	LLM-based	LLM-based	Continuous	LLM	Discriminative	Joint
Explainable Visual Anomaly Detection via Concept Bottleneck Models ↗Stropeni et al. · arXiv	2025	Image	CNNMisc.	VLM-based	VLM-based	Continuous	Linear	Discriminative	Joint
Improving intervention efficacy via concept realignment in concept bottleneck models ↗Singhi et al. · ECCV	2024	Image	CNNMisc.	Human-in-the-loop	Human-in-the-loop	Continuous	Linear	Discriminative	Joint
Sub: Benchmarking cbm generalization via synthetic attribute substitutions ↗Bader et al. · Proceedings of the IEEE/CVF	2025	Image	CNNMisc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Joint
Interpretable and Steerable Concept Bottleneck Sparse Autoencoders ↗Kulkarni et al. · arXiv	2025	Image	CNNMisc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Joint
A two-step concept-based approach for enhanced interpretability and trust in skin lesion diagnosis ↗Patrcio et al. · Computational and Structural	2025	Image	CNNMisc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Joint
CBVLM: Training-free Explainable Concept-based Large Vision Language Models for Medical Image Classification ↗Patrcio et al. · arXiv	2025	Text	VLMMisc.	LLM-based	LLM-based	Continuous	LLM	Discriminative	No training
Chat-CBM: Towards Interactive Concept Bottleneck Models with Frozen Large Language Models ↗He et al. · arXiv	2025	Text	LLMMisc.	LLM-based	LLM-based	Continuous	LLM	Generative	No training
Concept bottleneck language models for protein design ↗Ismail et al. · ICLR	2025	Text	LLMMisc.	LLM-based	LLM-based	Continuous	LLM	Generative	Joint
Concept graph embedding models for enhanced accuracy and interpretability ↗Kim et al. · Machine Learning: Science an	2024	Image	GNNMisc.	Disentanglement	Disentanglement	Continuous	GNN	Discriminative	Joint
Concept-centric transformers: Enhancing model interpretability through object-centric concept learning within a shared global workspace ↗Hong et al. · Proceedings of the IEEE/CVF	2024	Image	TransformerMisc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Joint
Continual learning for unsupervised concept bottleneck discovery ↗Lorello et al. · Conference on Lifelong Learn	2024	Image	CNNMisc.	Mixed / discovered	Mixed / discovered	Continuous	Linear	Discriminative	Sequential
Coreset Selection via LLM-based Concept Bottlenecks ↗Mehra et al. · Second Workshop on Visual Co	2025	Text	LLMMisc.	LLM-based	LLM-based	Continuous	LLM	Discriminative	Joint
DeCoDe: Defer-and-Complement Decision-Making via Decoupled Concept Bottleneck Models ↗He et al. · arXiv	2025	Image	CNNMisc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Joint
Editable concept bottleneck models ↗Hu et al. · arXiv	2024	Image	CNNMisc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Joint
Enhancing Interpretable Image Classification Through LLM Agents and Conditional Concept Bottleneck Models ↗Jiang et al. · arXiv	2025	Text	LLMMisc.	LLM-based	LLM-based	Continuous	LLM	Discriminative	Joint
EQ-CBM: A Probabilistic Concept Bottleneck with Energy-based Models and Quantized Vectors ↗Kim et al. · Asian Conference on Computer	2024	Image	CNNMisc.	GT Labels	GT Labels	Distribution	Linear	Discriminative	Joint
A Probabilistic Hard Concept Bottleneck for Steerable Generative Models ↗Martnez-Garca et al. · ICLR	2026	Image	CNNMisc.	GT Labels	GT Labels	Distribution	Linear	Generative	Joint
Explain via any concept: Concept bottleneck model with open vocabulary concepts ↗Tan et al. · ECCV	2024	Image	CNNMisc.	Mixed / discovered	Mixed / discovered	Continuous	Linear	Discriminative	Joint
The caltech-ucsd birds-200-2011 dataset ↗Wah et al. · Reference	2011	Other	Misc.Misc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Joint
Explanation Bottleneck Models ↗Yamaguchi et al. · AAAI	2025	Image	CNNMisc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Joint
Flexible Concept Bottleneck Model ↗Du et al. · arXiv	2025	Image	CNNMisc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Joint
Glancenets: Interpretable, leak-proof concept-based models ↗Marconato et al. · NeurIPS	2022	Image	CNNMisc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Joint
Graph concept bottleneck models ↗Xu et al. · arXiv	2025	Image	GNNMisc.	GT Labels	GT Labels	Continuous	GNN	Discriminative	Joint
Interpretable concept bottlenecks to align reinforcement learning agents ↗Delfosse et al. · NeurIPS	2024	Image	CNNMisc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Joint
Grounding dino: Marrying dino with grounded pre-training for open-set object detection ↗Liu et al. · ECCV	2024	Image	CNNMisc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Joint
Interpreting pretrained language models via concept bottlenecks ↗Tan et al. · Pacific-Asia Conference on K	2024	Text	LLMMisc.	LLM-based	LLM-based	Continuous	LLM	Discriminative	Joint
Learning concise and descriptive attributes for visual recognition ↗Yan et al. · Proceedings of the IEEE/CVF	2023	Image	CNNMisc.	VLM-based	VLM-based	Continuous	Linear	Discriminative	Joint
Neural concept binder ↗Stammer et al. · AAAI	2024	Image	CNNMisc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Joint
Neurosymbolic Diffusion Models ↗van Krieken et al. · NeurIPS	2025	Image	Misc.Misc.	Disentanglement	Disentanglement	Continuous	Symbolic	Generative	Joint
Process-Guided Concept Bottleneck Model ↗Asiyabi et al. · arXiv	2026	Image	CNNMisc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Joint
Restyling unsupervised concept based interpretable networks with generative models ↗Parekh et al. · arXiv	2024	Image	CNNMisc.	Mixed / discovered	Mixed / discovered	Continuous	Linear	Generative	Joint
Which lime should i trust? concepts, challenges, and solutions ↗Knab et al. · AAAI	2025	Image	CNNMisc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Joint
Right for the right concept: Revising neuro-symbolic concepts by interacting with their explanations ↗Stammer et al. · CVPR	2021	Image	CNNMisc.	Disentanglement	Disentanglement	Continuous	Symbolic	Discriminative	Joint
Towards Achieving Concept Completeness for Textual Concept Bottleneck Models ↗Bhan et al. · Joint European Conference on	2025	Text	LLMMisc.	LLM-based	LLM-based	Continuous	Linear	Discriminative	Joint
Towards Better Generalization and Interpretability in Unsupervised Concept-Based Models ↗De Santis et al. · Joint European Conference on	2025	Image	CNNMisc.	Mixed / discovered	Mixed / discovered	Continuous	Linear	Discriminative	Joint
Towards Reasonable Concept Bottleneck Models ↗Kalampalikis et al. · arXiv	2025	Image	CNNMisc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Joint
Zero-shot Concept Bottleneck Models ↗Yamaguchi et al. · arXiv	2025	Image	CNNMisc.	Mixed / discovered	Mixed / discovered	Continuous	Linear	Discriminative	No training
Probabilistic concept bottleneck models ↗Kim et al. · arXiv	2023	Image	CNNMisc.	GT Labels	GT Labels	Distribution	Linear	Discriminative	Joint
Towards Multi-Label Concept Bottleneck Models in Medical Imaging: An Exploratory Survey ↗Mpinda et al. · Medical Imaging with Deep Le	2026	Image	CNNMisc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Joint
Interactive Medical Image Analysis with Concept-based Similarity Reasoning ↗Huy et al. · CVPR	2025	Image	CNNMisc.	Human-in-the-loop	Human-in-the-loop	Continuous	Linear	Discriminative	Joint
Semi-supervised concept bottleneck models ↗Hu et al. · Proceedings of the IEEE/CVF	2025	Image	CNNMisc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Sequential
Learning bottleneck concepts in image classification ↗Wang et al. · CVPR	2023	Image	CNNMisc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Joint
Towards interpretable radiology report generation via concept bottlenecks using a multi-agentic rag ↗Alam et al. · European Conference on Infor	2025	Image	LLMMisc.	LLM-based	LLM-based	Continuous	Linear	Generative	Joint
Concept complement bottleneck model for interpretable medical image diagnosis ↗Wang et al. · arXiv	2024	Image	CNNMisc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Joint
A theoretical design of concept sets: improving the predictability of concept bottleneck models ↗Ruiz Luyten et al. · NeurIPS	2024	Image	CNNMisc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Joint
Causally reliable concept bottleneck models ↗De Felice et al. · arXiv	2025	Image	CNNMisc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Joint
Learning optimal summaries of clinical time-series with concept bottleneck models ↗Wu et al. · Machine Learning for Healthc	2022	Time Series	CNNMisc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Joint
Interpretability for Time Series Transformers using A Concept Bottleneck Framework ↗van Sprang et al. · arXiv	2024	Time Series	TransformerMisc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Joint
Learning to receive help: Intervention-aware concept embedding models ↗Espinosa Zarlenga et al. · NeurIPS	2023	Image	CNNMisc.	Human-in-the-loop	Human-in-the-loop	Continuous	MLP	Discriminative	Joint
Evidential Concept Embedding Models: Towards Reliable Concept Explanations for Skin Disease Diagnosis ↗Gao et al. · MICCAI	2024	Image	CNNMisc.	GT Labels	GT Labels	Distribution	MLP	Discriminative	Joint
MVP-CBM: Multi-layer Visual Preference-enhanced Concept Bottleneck Model for Explainable Medical Image Classification ↗Wang et al. · arXiv	2025	Image	CNNMisc.	VLM-based	VLM-based	Continuous	Linear	Discriminative	Joint
Aligning Human Knowledge with Visual Concepts Towards Explainable Medical Image Classification ↗Gao et al. · MICCAI	2024	Image	CNNMisc.	VLM-based	VLM-based	Continuous	Linear	Discriminative	Joint
CLIP-QDA: An Explainable Concept Bottleneck Model ↗Kazmierczak et al. · ICML	2024	Image	VLMMisc.	VLM-based	VLM-based	Continuous	Linear	Discriminative	Joint
Addressing Concept Mislabeling in Concept Bottleneck Models Through Preference Optimization ↗Penaloza et al. · ICML	2025	Image	CNNMisc.	Human-in-the-loop	Human-in-the-loop	Continuous	Linear	Discriminative	Joint
Towards multi-dimensional explanation alignment for medical classification ↗Hu et al. · NeurIPS	2024	Image	CNNMisc.	GT Labels	GT Labels	Continuous	Linear	Discriminative	Joint
Faithful Vision-Language Interpretation via Concept Bottleneck Models ↗Lai et al. · ICLR	2024	Text	VLMMisc.	VLM-based	VLM-based	Continuous	Linear	Discriminative	Joint
Interactive Disentanglement: Learning Concepts by Interacting with their Prototype Representations ↗Stammer et al. · CVPR	2022	Image	CNNMisc.	Human-in-the-loop	Human-in-the-loop	Continuous	Linear	Discriminative	Joint
Insight: Interpretable Semantic Hierarchies in Vision-Language Encoders ↗Wittenmayer et al. · arXiv	2026	Text	VLMMisc.	VLM-based	VLM-based	Continuous	Linear	Discriminative	Joint
Semantic bottlenecks: Quantifying and improving inspectability of deep representations ↗Losch et al. · International Journal of Com	2021	Image	CNNMisc.	Disentanglement	Disentanglement	Continuous	Linear	Discriminative	Joint
Controllable Concept Bottleneck Models ↗Lin et al. · arXiv	2026	Image	CNNStructured	Human-in-the-loop	Human-in-the-loop	Continuous	Linear	Discriminative	Post-hoc
Concepts’ Information Bottleneck Models ↗Galliamov et al. · arXiv	2026	Image	CNNStructured	GT labels	GT labels	Distribution	Linear	Discriminative	Joint
Rethinking Concept Bottleneck Models: From Pitfalls to Solutions ↗Tapli et al. · arXiv	2026	Image	ViTVLM	Mixed / discovered	Mixed / discovered	Continuous	MLP	Discriminative	Joint
Matryoshka Concept Bottleneck Models ↗Chen et al. · arXiv	2026	Image	CNNStructured	GT labels	GT labels	Continuous	Linear	Discriminative	Joint
Post-hoc Stochastic Concept Bottleneck Models ↗Hoffmann et al. · ICLR 2026 workshop	2026	Image	CNNStructured	GT labels	GT labels	Distribution	Linear	Discriminative	Post-hoc
Simulating Concept Bottlenecks with Vision-Language Models ↗Galliamov et al. · ICLR 2026 workshop	2026	Image	VLMUnstructured	LLM-based	VLM-based	Structured	MLP	Discriminative	Joint
Partially Shared Concept Bottleneck Models ↗Zhao et al. · AAAI 2026	2026	Image	VLMStructured	VLM-based	VLM-based	Continuous	Linear	Discriminative	Joint
Scaling Inherently Interpretable Language Models ↗Guide Labs Team et al. · arXiv	2026	Text	TransformerUnstructured	LLM-based	Mixed	Cont. & disc.	Transformer	Generative	Joint

Input modality, encoder, embeddingConcept semantics, grounding, activationsOutput predictor and taskTraining joint, independent, sequential, or none

Search titles, authors, or ideas

Filter by contribution

Showing 6 anchor workstaxonomy tags are intentionally cross-cutting

2020foundation

Concept Bottleneck Models

Koh et al.

Introduces the canonical two-stage architecture and the idea of test-time concept intervention.

x → c → ysupervised

paper ↗

2021evaluation

Promises and Pitfalls of Black-Box Concept Learning Models

Mahinpei et al. · ICML XAI workshop

Stress-tests the gap between concept accuracy, task performance, and the interpretability claim.

pitfallsfaithfulness

read paper ↗

2022training

Addressing Leakage in Concept Bottleneck Models

Havasi et al.

Shows how information can bypass the intended bottleneck and proposes ways to reduce concept leakage.

leakageshortcuts

find paper ↗

2022concepts

Concept Embedding Models: Beyond the Limitations of Concept Bottleneck Models

Espinosa Zarlenga et al.

Revisits the bottleneck representation to make concept embeddings more expressive than scalar activations.

embeddingsexpressivity

find paper ↗

2023concepts

Label-free Concept Bottleneck Models

Oikarinen et al.

Explores how vision-language models can supply concept supervision when human concept labels are unavailable.

label-freeVLM

find paper ↗

2023intervention

Post-hoc Concept Bottleneck Models

Yüksekgönül et al.

Asks how a concept interface can be added after a strong predictor has already been trained.

post-hocediting

find paper ↗

2023+intervention

Intervenable and editable CBMs

A growing line of work

Moves beyond explaining a prediction toward letting humans correct, edit, or supervise reasoning at inference time.

human-in-loopeditable

search thread ↗

2023+concepts

Discovering concepts from data

Prototype-, clustering-, and VLM-guided methods

Relaxes the assumption that experts must enumerate every concept, but raises new questions about meaning and bias.

discoverygrounding

search thread ↗

2023+training

Beyond complete concept sets

Incomplete, probabilistic, and structured CBMs

Studies what happens when the listed concepts do not fully explain the target or depend on one another.

completenessuncertainty

search thread ↗

2024+evaluation

Evaluating concept-based reasoning

Benchmarks and diagnostic protocols

Builds more demanding tests around faithfulness, robustness, intervention gains, sparsity, and human usefulness.

metricsrobustness

search thread ↗

2024+context

Concept-based XAI beyond CBMs

Related surveys and position papers

Places bottleneck models alongside concept-based explanations, prototypes, concept factorization, and feature attribution.

contextrelated XAI

survey thread ↗

2026this survey

What’s in the Bottle? A Survey and Roadmap

Knab, Steinmann, Bartelt et al.

Unifies the field around input, concept, output, and training modules—and maps the challenges between them.

taxonomyroadmap

read paper ↗

Development of the field

From the canonical model to new settings.

The literature begins with supervised visual prediction and progressively changes the concept representation, concept source, intervention mechanism, and task setting.

2020

Concept Bottleneck Models

Koh et al. introduce the two-stage CBM formulation and demonstrate it on x-ray grading and bird identification, with concepts mediating the final prediction.

2021—22

Reliability challenges

Subsequent work examines concept quality, information leakage, label noise, and the relation between concept accuracy and task performance.

2023—25

Extensions of the bottleneck

Methods modify concept representations, concept sources, intervention mechanisms, and training procedures, including label-free and multimodal variants.

2026—

Generative and multimodal settings

Recent work applies concept-based interfaces to language and generative models and to modalities beyond the original image-classification setting.

Challenges and open problems

Questions that remain unresolved.

The points below are only a fraction of the challenges discussed in the paper. The survey groups open problems around concept-module design, model reliability, and the validation and scope of CBMs.

01 · SEMANTICS

Concept acquisition and semantics

Which concepts should be selected, who defines them, and how should their meaning be specified? Expert concepts are costly; discovered concepts may be unstable or difficult to interpret.

02 · GROUNDING

Concept grounding

Does an activation correspond to the relevant evidence in the input, or can the model obtain the correct label through a shortcut?

03 · REPRESENTATION

Representation and completeness

Concept representations must be expressive enough for the task while remaining human-understandable. Incomplete concept sets may omit task-relevant information.

04 · LEAKAGE

Leakage and enforcement

Residual pathways and expressive predictors can carry information outside the intended concept path, weakening selective interpretability.

05 · INTERVENTION

Intervention reliability

A useful intervention requires calibrated concept predictions and a predictable downstream effect, not only a change in a concept score.

06 · VALIDATION

Validation and scope

Evaluation must distinguish predictive accuracy, concept quality, faithfulness, intervention behavior, and performance across domains and distribution shifts.

Citation

Cite the survey.

Use the BibTeX entry below when referring to the survey and its taxonomy.

BibTeX

@article{
knab2026whats,
title={What{\textquoteright}s in the Bottle? A Survey and Roadmap of Concept Bottleneck Models},
author={Patrick Knab and David Steinmann and Christian Bartelt and Kristian Kersting and Bernt Schiele and Thomas Seidl and Udo Schlegel and Wolfgang Stammer},
journal={Transactions on Machine Learning Research},
issn={2835-8856},
year={2026},
url={https://openreview.net/forum?id=IF5vnqxBEW},
note={}
}

The matrix is a selected, searchable view of the categorization presented in the survey.