Research map / Probabilistic and scientific modeling / Causal discovery
Causal representation identifiability
Prove when causal variables can be recovered from observations under interventions or weak supervision.
122 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2017–2026) · an approach in Causal discovery, keyword cluster “counterfactual · causal inference · causal representation”
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.
40 papers in 2023–24 and 52 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic held steady, from 17.2% to 15.6%.
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 · Schema Networks: Zero-shot Transfer with a Generative Causal Model of Intuitive Physics (ICML 2017 · 168 citations)
- 2018 · Learning Abduction Using Partial Observability (AAAI 2018 · 0 citations)
- 2019 · Causal Confusion in Imitation Learning (NeurIPS 2019 · 162 citations)
- 2020 · Explainable Reinforcement Learning through a Causal Lens (AAAI 2020 · 263 citations)
- 2021 · CausalVAE: Disentangled Representation Learning via Neural Structural Causal Models (CVPR 2021 · 248 citations)
- 2022 · Causal Intervention for Subject-Deconfounded Facial Action Unit Recognition (AAAI 2022 · 30 citations)
- 2023 · Layout-Based Causal Inference for Object Navigation (CVPR 2023 · 37 citations)
- 2024 · Causal Representation Learning via Counterfactual Intervention (AAAI 2024 · 7 citations)
- 2025 · Teaching Transformers Causal Reasoning through Axiomatic Training (ICML 2025 · 2 citations)
- 2026 · CausalCLIP: Causally-Informed Feature Disentanglement and Filtering for Generalizable Detection of Generated Images (AAAI 2026 · 1 citation)
Earliest papers here
- Schema Networks: Zero-shot Transfer with a Generative Causal Model of Intuitive Physics (ICML 2017 · 168 citations)
- Learning Abduction Using Partial Observability (AAAI 2018 · 0 citations)
- Causal Confusion in Imitation Learning (NeurIPS 2019 · 162 citations)
What has been tried
86 papers here report 676 measured changes (a component added, removed or swapped, compared with a baseline): helped 65%, no clear effect 8%, hurt 23%, mixed 4%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| remove inductive bias during causal-effect learning using counterfactual intervention | Age, Smile, Pend, Flow · MIC | 0.615 -> 0.626; 0.851 -> 0.873; 0.785 -> 0.997; 0.973 -> 0.992 | Helped | Causal Representation Learning via Counterfactual Intervention (AAAI 2024 · 7 citations) |
| Introduce a hidden variable that drives observed features | Synthetic risk prediction · precision | 0.89 -> 0.72\downarrow | Hurt | Task-Driven Causal Feature Distillation: Towards Trustworthy Risk Prediction (AAAI 2024 · 6 citations) |
| use diffusion-based context-to-label representation transformation | ESL intra + inter-sentence · F1 | 57.7 -> 63.5 | Helped | Mastering Context-to-Label Representation Transformation for Event Causality Identification with Diffusion Models (AAAI 2024 · 5 citations) |
| Set the threshold tl to 0.15 | Random causal models, d=5, K=5 · SNA error and graph recovery accuracy | Helped | Learning Linear Causal Representations from General Environments: Identifiability and Intrinsic Ambiguity (NeurIPS 2024 · 3 citations) | |
| add random edge flipping to sequential-chain training data (TS1/TS2 instead of OCC) | Transitivity, reversed chains · accuracy | OCC w NoPE: 0.41, 0.24, 0.18, 0.13; OCC w RoPE: 0.59, 0.26, 0.22, 0.20 | Helped | Teaching Transformers Causal Reasoning through Axiomatic Training (ICML 2025 · 2 citations) |
| Use an interventional query with a zeroed input error instead of a conditional query | MNIST and FashionMNIST · test accuracy | almost 2% | Helped | Predictive Coding beyond Correlations (ICML 2024 · 2 citations) |
| Reduce the number of informative scenes from Four-Scenes to One-Scene | Static image generation · MCC | 0.3264 -> 0.3197 | Hurt | CausalVerse: Benchmarking Causal Representation Learning with Configurable High-Fidelity Simulations (NeurIPS 2025 · 1 citation) |
| Increase the sparsity ratio of inter-modal causal links | Numerical dataset, Case 2 · MCC | Helped | Causal Representation Learning from Multimodal Biomedical Observations (ICLR 2025 · 1 citation) | |
| Keep a causal coefficient unchanged across environments | Synthetic linear models with Beta noise · MPC | Hurt | Identifiable Latent Polynomial Causal Models through the Lens of Change (ICLR 2024 · 1 citation) | |
| use modular training instead of training all mechanisms jointly | Colored-MNIST frontdoor, P(A|do(D=0)) · TVD | 0.36 -> 0.13 | Helped | Modular Learning of Deep Causal Generative Models for High-dimensional Causal Inference (ICML 2024 · 1 citation) |
| Add sparsity constraints to TeSLA-s | CIFAR10-C · error rate | 12.1 -> 11.7 | Helped | Towards Understanding Extrapolation: a Causal Lens (NeurIPS 2024 · 1 citation) |
| Use the element-wise absolute-difference formulation of the shift statistic L_i instead of the original row-vector comparison | Synthetic shift detection · shift detection performance | Helped | Identifying General Mechanism Shifts in Linear Causal Representations (NeurIPS 2024 · 1 citation) |
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 |
|---|---|---|---|---|
| β-TCVAE | dSprites | IRS | 0.72 – 0.78 | 2 papers, e.g. ICML 2019 and AAAI 2022 |
What reviewers will expect
Datasets and baselines named in the 34 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- CelebA · 4 of 34 papers
- IOI · 2 of 34 papers
- RAVEL · 2 of 34 papers
- cMNIST · 2 of 34 papers
- ISTAnt · 2 of 34 papers
Baselines
- beta-VAE · 2 of 34 papers
- TDRL · 2 of 34 papers
- iVAE · 2 of 34 papers
- TCL · 2 of 34 papers
- CaRiNG · 2 of 34 papers
- IDOL · 2 of 34 papers
- CBM · 2 of 34 papers
- CEM · 2 of 34 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 92 papers; from OpenAlex and the papers' affiliations).
- Carnegie Mellon University · 15 papers
- Columbia University · 5 papers
- Institute of Science and Technology · 5 papers
- Mohamed bin Zayed University of Artificial Intelligence · 5 papers
- University of Amsterdam · 4 papers
- ETHZ - ETH Zurich · 4 papers
- University of Montreal · 4 papers
- Massachusetts Institute of Technology · 4 papers
- Stanford University · 4 papers
- Max Planck Institute for Intelligent Systems · 3 papers
Use it
Download BibTeX of 40 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
- Latent-variable causal estimation (117 papers)
- Causal effect identification (107 papers)
- Counterfactual identification and bounds (92 papers)
- Causal approaches to fairness (46 papers)
Elsewhere in the topic
- Constraint-based structure learning with latents (90 papers)
- Identifiability with hidden confounders (85 papers)
- Interventional causal discovery (68 papers)
- Differentiable and score-based DAG learning (52 papers)
- Causal discovery from time series (51 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).