Research map / Vision-language and multimodal / Human and embodied modeling
Language-grounded visual concept learning
Learns visual concepts by tokenization, decomposition, and transfer from language models.
83 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2017–2026) · an approach in Human and embodied modeling, keyword cluster “concepts · brain · fmri”
State History Tried 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.
20 papers in 2023–24 and 39 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 3.8% to 5.8%.
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
- 2017 · Dense Captioning With Joint Inference and Visual Context (CVPR 2017 · 175 citations)
- 2018 · SCAN: Learning Hierarchical Compositional Visual Concepts (ICLR 2018 · 79 citations)
- 2019 · The Neuro-Symbolic Concept Learner: Interpreting Scenes, Words, and Sentences From Natural Supervision (ICLR 2019 · 139 citations)
- 2020 · PaStaNet: Toward Human Activity Knowledge Engine (CVPR 2020 · 145 citations)
- 2021 · Right for the Right Concept: Revising Neuro-Symbolic Concepts by Interacting With Their Explanations (CVPR 2021 · 64 citations)
- 2022 · Multi-Grained Vision Language Pre-Training: Aligning Texts with Visual Concepts (ICML 2022 · 95 citations)
- 2023 · GIVL: Improving Geographical Inclusivity of Vision-Language Models With Pre-Training Methods (CVPR 2023 · 14 citations)
- 2024 · MICA: Towards Explainable Skin Lesion Diagnosis via Multi-Level Image-Concept Alignment (AAAI 2024 · 27 citations)
- 2025 · Design2GarmentCode: Turning Design Concepts to Tangible Garments Through Program Synthesis (CVPR 2025 · 12 citations)
- 2026 · Into the Rabbit Hull: From Task-Relevant Concepts in DINO to Minkowski Geometry (ICLR 2026 · 5 citations)
Earliest papers here
- Dense Captioning With Joint Inference and Visual Context (CVPR 2017 · 175 citations)
- Imagined Visual Representations as Multimodal Embeddings (AAAI 2017 · 82 citations)
- SCAN: Learning Hierarchical Compositional Visual Concepts (ICLR 2018 · 79 citations)
What has been tried
60 papers here report 612 measured changes (a component added, removed or swapped, compared with a baseline): helped 64%, no clear effect 8%, hurt 25%, 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 image-level alignment (ILA) | Derm7pt; PH²; SkinCon · AUC_D (%) | 82.9 -> 84.1; 96.1 -> 97.7; 75.0 -> 75.9 | Helped | MICA: Towards Explainable Skin Lesion Diagnosis via Multi-Level Image-Concept Alignment (AAAI 2024 · 27 citations) |
| Add Learnable Instruction | ORES · M-Evasion Ratio | 28.8% -> 85.6% | Helped | ORES: Open-Vocabulary Responsible Visual Synthesis (AAAI 2024 · 5 citations) |
| Use Semi-NMF rather than random-word grounding | COCO Dog · CLIPScore top-1 | 0.519 ± 0.05 -> 0.610 ± 0.09 | Helped | A Concept-Based Explainability Framework for Large Multimodal Models (NeurIPS 2024 · 3 citations) |
| Add Gaussian alignment | Broden, ResNet-18 · concept accuracy (%) | 67.23 -> 74.89 | Helped | LG-CAV: Train Any Concept Activation Vector with Language Guidance (NeurIPS 2024 · 2 citations) |
| use automatic concept descriptions rather than MILAN annotations, without contextual filtering | ImageNet Validation · Total explanation score (GPT-4o) | 3.14 -> 3.25 | Helped | CoE: Chain-of-Explanation via Automatic Visual Concept Circuit Description and Polysemanticity Quantification (CVPR 2025 · 1 citation) |
| increase LoRA rank from 1 to 2 | ConCon-Chi single-concept context queries · mRR | 51.64 -> 52.31 | Mixed | Improving Personalized Search with Regularized Low-Rank Parameter Updates (CVPR 2025 · 0 citations) |
| Add context-aware attention to the GCD setup | Stanford Action · novel accuracy | 67.8 -> 73.3 | Helped | Open Ad-hoc Categorization with Contextualized Feature Learning (CVPR 2025 · 0 citations) |
| Use reverse weighting rather than uniform weighting | ImageNet-1k training · L0 | .748 (±.006) -> .829 (±.008) | Helped | Interpreting CLIP with Hierarchical Sparse Autoencoders (ICML 2025 · 0 citations) |
| use RL search rather than 0.25 Greedy | Search strategy ablation · Entropy | 2.40 -> 2.80 | Helped | Explainable Concept Generation through Vision-Language Preference Learning for Understanding Neural Networks' Internal Representations (ICML 2025 · 0 citations) |
| add contrastive visual features instead of using non-contrastive visual features | NovelSpecies, Fixed Real Data 5:1 · accuracy | 71.8 -> 73.0 | Helped | Contrastive Visual Data Augmentation (ICML 2025 · 0 citations) |
| use Qwen2.5-VL instead of InternVL2-5 for concept-axis discovery | CelebA-HQ visual concept editing · CLIP | 23.88 -> 23.72 | Hurt | Bridging the gap to real-world language-grounded visual concept learning (NeurIPS 2025 · 0 citations) |
| Add the text anchor loss | Visual concept editing, category edit · CLIP-Score | 0.268 -> 0.308 | Helped | Language-Informed Visual Concept Learning (ICLR 2024 · 0 citations) |
What reviewers will expect
Datasets and baselines named in the 27 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 · 3 of 27 papers
- ImageNet-1K · 3 of 27 papers
- CelebA · 3 of 27 papers
- DTD · 2 of 27 papers
Baselines
- CKA · 2 of 27 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 59 papers; from OpenAlex and the papers' affiliations).
- University of California, Los Angeles · 3 papers
- Amazon · 3 papers
- Adobe · 3 papers
- Arizona State University · 2 papers
- Microsoft · 2 papers
- Zhejiang University · 2 papers
- Massachusetts Institute of Technology · 2 papers
- Carnegie Mellon University · 2 papers
- Korea Advanced Institute of Science and Technology · 2 papers
Use it
Download BibTeX of 37 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)
- Concept bottleneck models (21 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).