Research map / Vision-language and multimodal / Human and embodied modeling
Concept bottleneck models
Builds interpretable classifiers that predict human-understandable concepts using vision-language guidance.
21 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2020–2026) · an approach in Human and embodied modeling, keyword cluster “concepts · brain · fmri”
State History Tried Checklist Preprints Export
Where it stands
Not enough benchmark data Fewer than four papers report the same benchmark, so there is no record to follow.
7 papers in 2023–24 and 12 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 1.3% to 1.8%.
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 · Bongard-LOGO: A New Benchmark for Human-Level Concept Learning and Reasoning (NeurIPS 2020 · 21 citations)
- 2022 · Mapping Language Models to Grounded Conceptual Spaces (ICLR 2022)
- 2023 · What’s Left? Concept Grounding with Logic-Enhanced Foundation Models (NeurIPS 2023 · 3 citations)
- 2024 · VLG-CBM: Training Concept Bottleneck Models with Vision-Language Guidance (NeurIPS 2024 · 21 citations)
- 2025 · Hybrid Concept Bottleneck Models (CVPR 2025 · 6 citations)
- 2026 · Partially Shared Concept Bottleneck Models (AAAI 2026 · 2 citations)
Earliest papers here
- Bongard-LOGO: A New Benchmark for Human-Level Concept Learning and Reasoning (NeurIPS 2020 · 21 citations)
- Mapping Language Models to Grounded Conceptual Spaces (ICLR 2022)
- What’s Left? Concept Grounding with Logic-Enhanced Foundation Models (NeurIPS 2023 · 3 citations)
What has been tried
11 papers here report 127 measured changes (a component added, removed or swapped, compared with a baseline): helped 54%, no clear effect 8%, hurt 29%, mixed 9%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| set confidence threshold T to 0.10 rather than 0.20 | CUB200 · ANEC-5 | 75.73% -> 75.75% | Helped | VLG-CBM: Training Concept Bottleneck Models with Vision-Language Guidance (NeurIPS 2024 · 21 citations) |
| Add dynamic concepts to the static concept bank (ratio 0 to 0.2) | Average across 10 datasets · accuracy | 84.27 -> 85.47 | Helped | Hybrid Concept Bottleneck Models (CVPR 2025 · 6 citations) |
| Add PIM | BRCA N_IND=1 OOD · AUC | 0.914{ ±0.015} -> 0.932{ ±0.011} | Helped | Free Lunch in Pathology Foundation Model: Task-specific Model Adaptation with Concept-Guided Feature Enhancement (NeurIPS 2024 · 6 citations) |
| Increase unlabeled image set from 1k to 200k | CIFAR10, CUB, DTD, RESISC45 · accuracy | CIFAR10: 97.6 -> 98.0; CUB: 80.3 -> 83.0; DTD: 73.1 -> 78.0; RESISC45: 90.2 -> 92.6 | Helped | V2C-CBM: Building Concept Bottlenecks with Vision-to-Concept Tokenizer (AAAI 2025 · 4 citations) |
| add hierarchical linkage between high-level concepts and patch-level attributes | CUB · accuracy | 59.62 -> 73.20 | Helped | Coarse-to-Fine Concept Bottleneck Models (NeurIPS 2024 · 2 citations) |
| add explanation distillation | Aircraft · CLIP-Score | 0.4730 -> 0.7151 | Helped | Explanation Bottleneck Models (AAAI 2025 · 1 citation) |
| Apply PCA with 100 components before clustering | CUB · top-1 accuracy | 81.8 -> 81.9 | No clear effect | DCBM: Data-Efficient Visual Concept Bottleneck Models (ICML 2025 · 0 citations) |
| use iterative refinement instead of stopping after initial discretization | Clustering benchmarks · F1 | 0.53 -> 0.64 | Helped | Explaining Datasets in Words: Statistical Models with Natural Language Parameters (NeurIPS 2024 · 0 citations) |
What reviewers will expect
Datasets and baselines named in the 5 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- ImageNet · 4 of 5 papers
- CUB-200-2011 · 3 of 5 papers
- CIFAR-10 · 3 of 5 papers
- CIFAR-100 · 3 of 5 papers
- FGVC Aircraft · 2 of 5 papers
- CUB · 2 of 5 papers
- DTD · 2 of 5 papers
- Food-101 · 2 of 5 papers
Baselines
- LaBo · 4 of 5 papers
- linear probe · 4 of 5 papers
- Label-free CBM · 2 of 5 papers
- CDM · 2 of 5 papers
- DCLIP · 2 of 5 papers
- DN-CBM · 2 of 5 papers
- LF-CBM · 2 of 5 papers
Recent preprints
Loading recent arXiv preprints…
Use it
Download BibTeX of 21 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
- fMRI visual decoding (106 papers)
- Language-grounded visual concept learning (83 papers)
Elsewhere in the topic
- Vision-language pretraining (182 papers)
- Medical vision-language models (130 papers)
- Scene graph generation (126 papers)
- Language-grounded scene text and structure (124 papers)
- Human-like visual cognition in VLMs (97 papers)
- Multimodal and affective interaction modeling (87 papers)
- Multimodal retrieval-augmented modeling (85 papers)
- Composed and text-based image retrieval (78 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).