Research map / Label-efficient and robust learning / Recognition and long-tail
Concept bottleneck models
Predicts through human-interpretable concept layers, including interactive, probabilistic and language-guided variants.
30 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2020–2026) · an approach in Recognition and long-tail, keyword cluster “explanations · concept · interpretable”
State History Tried Baselines Checklist Preprints Who Export
Where it stands
Not enough benchmark data Fewer than four papers report the same benchmark, so there is no record to follow.
13 papers in 2023–24 and 13 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic held steady, from 1.0% to 1.0%.
Open questions the data points to
- Papers report PCBM on CIFAR-100 (accuracy) anywhere from 43.2 to 56: comparisons against it depend on whose number you take.
- Papers report PCBM on CIFAR-10 (accuracy) anywhere from 70.5 to 84.5: comparisons against it depend on whose number you take.
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2020 · Concept Bottleneck Models (ICML 2020 · 149 citations)
- 2022 · Concept Embedding Models: Beyond the Accuracy-Explainability Trade-Off (NeurIPS 2022 · 36 citations)
- 2023 · Language in a Bottle: Language Model Guided Concept Bottlenecks for Interpretable Image Classification (CVPR 2023 · 151 citations)
- 2024 · On the Concept Trustworthiness in Concept Bottleneck Models (AAAI 2024 · 10 citations)
- 2025 · Discovering Fine-Grained Visual-Concept Relations by Disentangled Optimal Transport Concept Bottleneck Models (CVPR 2025 · 5 citations)
- 2026 · SL-CBM: Enhancing Concept Bottleneck Models with Semantic Locality for Better Interpretability (AAAI 2026 · 2 citations)
Earliest papers here
- Concept Bottleneck Models (ICML 2020 · 149 citations)
- Concept Embedding Models: Beyond the Accuracy-Explainability Trade-Off (NeurIPS 2022 · 36 citations)
- Addressing Leakage in Concept Bottleneck Models (NeurIPS 2022 · 13 citations)
What has been tried
18 papers here report 217 measured changes (a component added, removed or swapped, compared with a baseline): helped 65%, no clear effect 11%, hurt 20%, mixed 3%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| add OT assignment to the baseline | CUB · classification accuracy | 80.92 -> 81.75 | Helped | Discovering Fine-Grained Visual-Concept Relations by Disentangled Optimal Transport Concept Bottleneck Models (CVPR 2025 · 5 citations) |
| use only the x-y energy branch rather than the full ECBM | CUB · Class accuracy | 0.825 -> 0.812 | Helped | Energy-Based Concept Bottleneck Models: Unifying Prediction, Concept Intervention, and Probabilistic Interpretations (ICLR 2024 · 4 citations) |
| Use a concept representation instead of the direct CLIP embedding | CIFAR-10 · generalization performance | Mixed | A theoretical design of concept sets: improving the predictability of concept bottleneck models (NeurIPS 2024 · 4 citations) | |
| increase memory from 25% to larger fractions of available intervention data | CUB (Aug.) · concept and class accuracy | Helped | Learning to Intervene on Concept Bottlenecks (ICML 2024 · 3 citations) | |
| Reduce the validation set size | Synthetic bottleneck · intervention effectiveness | Hurt | Beyond Concept Bottleneck Models: How to Make Black Boxes Intervenable? (NeurIPS 2024 · 3 citations) |
Baseline numbers that disagree
The same baseline on the same benchmark and metric, reported more than 5% apart by different papers here. Training budgets, data splits and reimplementations differ, so state which number you compare with.
| Baseline | Benchmark | Metric | Reported | Papers |
|---|---|---|---|---|
| PCBM | CIFAR-100 | accuracy | 43.2 – 56 | 2 papers, e.g. ICLR 2023 and CVPR 2023 |
| PCBM | CIFAR-10 | accuracy | 70.5 – 84.5 | 2 papers, e.g. ICLR 2023 and CVPR 2023 |
What reviewers will expect
Datasets and baselines named in the 8 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- CUB · 4 of 8 papers
- AwA2 · 4 of 8 papers
- CelebA · 4 of 8 papers
- CIFAR-100 · 3 of 8 papers
- OAI · 2 of 8 papers
- ImageNet · 2 of 8 papers
Baselines
- CEM · 3 of 8 papers
- LaBo · 2 of 8 papers
- CoopCBM · 2 of 8 papers
- ProbCBM · 2 of 8 papers
- CBM · 2 of 8 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 26 papers; from OpenAlex and the papers' affiliations).
- University of Cambridge · 3 papers
- Google · 2 papers
- University of California, San Diego · 2 papers
- Hong Kong University of Science and Technology · 2 papers
- University of Montreal · 2 papers
Use it
Download BibTeX of 30 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Self-explaining and visual explanation networks (108 papers)
- Prototype and concept dictionary models (104 papers)
- Post-hoc explanation analysis (86 papers)
- Feature attribution methods (51 papers)
Elsewhere in the topic
- Multilingual language representation pretraining (507 papers)
- Contrastive representation theory (495 papers)
- Dense features and foundation priors (482 papers)
- Correspondence and causal representation learning (447 papers)
- Data augmentation and generalization bias (409 papers)
- Partial and multi-label learning (345 papers)
- Loss design and classifier theory (273 papers)
- Group-robust classification (251 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).