Research map / Probabilistic and scientific modeling / Molecules and proteins
Prediction-powered and conformal inference
Gives valid statistical inference using conformal, prediction-powered and multiple-testing methods.
129 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2016–2026) · an approach in Molecules and proteins, keyword cluster “decision · machine · statistical”
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.
25 papers in 2023–24 and 69 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 5.1% to 7.4%.
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
- 2016 · Statistical Inference for Cluster Trees (NeurIPS 2016 · 21 citations)
- 2017 · Task-based End-to-end Model Learning in Stochastic Optimization (NeurIPS 2017 · 122 citations)
- 2018 · Gamma-Poisson Dynamic Matrix Factorization Embedded with Metadata Influence (NeurIPS 2018 · 8 citations)
- 2019 · Meta-Learning Probabilistic Inference for Prediction (ICLR 2019 · 64 citations)
- 2020 · Classification with Valid and Adaptive Coverage (NeurIPS 2020 · 24 citations)
- 2021 · Measuring Dependence with Matrix-based Entropy Functional (AAAI 2021 · 31 citations)
- 2022 · ApproxIFER: A Model-Agnostic Approach to Resilient and Robust Prediction Serving Systems (AAAI 2022 · 10 citations)
- 2023 · Revisiting the Stack-Based Inverse Tone Mapping (CVPR 2023 · 8 citations)
- 2024 · A benchmark for prediction of transcriptomic responses to chemical perturbations across cell types (NeurIPS 2024 · 6 citations)
- 2025 · PerturBench: Benchmarking Machine Learning Models for Cellular Perturbation Analysis (NeurIPS 2025 · 18 citations)
- 2026 · T2Agent: A Tool-augmented Multimodal Misinformation Detection Agent with Monte Carlo Tree Search (AAAI 2026 · 2 citations)
Earliest papers here
- Statistical Inference for Cluster Trees (NeurIPS 2016 · 21 citations)
- A Non-parametric Learning Method for Confidently Estimating Patient's Clinical State and Dynamics (NeurIPS 2016 · 5 citations)
- False Discovery Rate Control and Statistical Quality Assessment of Annotators in Crowdsourced Ranking (ICML 2016 · 4 citations)
What has been tried
82 papers here report 727 measured changes (a component added, removed or swapped, compared with a baseline): helped 66%, no clear effect 11%, hurt 18%, mixed 5%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| remove CPA* adversarial component | Srivatsan transfer · Cosine LogFC rank | 0.15 ± 1 × 10^{-2} -> 0.09 ± 4 × 10^{-3} | Helped | PerturBench: Benchmarking Machine Learning Models for Cellular Perturbation Analysis (NeurIPS 2025 · 18 citations) |
| Add inferred gene regulatory networks from ATAC and RNA data | Unseen-context perturbation prediction · prediction performance | Hurt | A benchmark for prediction of transcriptomic responses to chemical perturbations across cell types (NeurIPS 2024 · 6 citations) | |
| Add instance-specific weights to exponential age weighting | Geo · accuracy | Helped | Instance-Conditional Timescales of Decay for Non-Stationary Learning (AAAI 2024 · 5 citations) | |
| add the counterfactual KL auxiliary loss | Norman, 3σ · ATE-ρ | 0.6505 ± 0.02 -> 0.7119 ± 0.03 | Helped | Cradle-VAE: Enhancing Single-Cell Gene Perturbation Modeling with Counterfactual Reasoning-based Artifact Disentanglement (AAAI 2025 · 3 citations) |
| Vary the Wendland kernel range from 1,000 km to 500 km or 2,500 km | CMIP6 surface temperature and precipitation · model ranking | No clear effect | Validating Climate Models with Spherical Convolutional Wasserstein Distance (NeurIPS 2024 · 3 citations) | |
| Increase max batch size from 4 to 64 | GPQA on NVIDIA H100 · energy per generation | 9713.7 -> 2770.8 | Helped | The ML.ENERGY Benchmark: Toward Automated Inference Energy Measurement and Optimization (NeurIPS 2025 · 2 citations) |
| use SPI-Subset instead of SPI-Whole | ImageNet, label-conditional · prediction set size | Helped | Synthetic-powered predictive inference (NeurIPS 2025 · 1 citation) | |
| Use level-$(1-6\alpha/(n^2\pi^2))$ confidence sequences in round n | Non-partitioned SCD, no change · probability of false alarms | \mathbb{P} (\infty)(\tau<\infty) \leq \alpha | Helped | Reducing sequential change detection to sequential estimation (ICML 2024 · 1 citation) |
| Fine-tune the predictive model during label collection | Pew 2020 election survey — Biden and Trump approval · budget required for matched confidence interval width | Biden: around 40\% vs. no fine-tuning and 30\% vs. PPI; Trump: around 45\% and 35\% | Helped | Active Statistical Inference (ICML 2024 · 1 citation) |
| Reduce the censoring-model training sample size, lowering imputation quality | Synthetic setting 1 · coverage | Hurt | Doubly Robust Conformalized Survival Analysis with Right-Censored Data (ICML 2025 · 0 citations) | |
| Add adaptive shrinkage to PT | Synthetic, f1(x)=x² · MSE (× 10^-3) | 2.642 ± 0.027 -> 0.272 ± 0.003 | Helped | Prediction-Powered Adaptive Shrinkage Estimation (ICML 2025 · 0 citations) |
| Increase the Q-DCP smoothness parameter κ from small to large values | CIFAR100 torus graph · set size | Mixed | Distributed Conformal Prediction via Message Passing (ICML 2025 · 0 citations) |
What reviewers will expect
Datasets and baselines named in the 37 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 37 papers
- galaxies · 2 of 37 papers
- forest · 2 of 37 papers
- healthcare · 2 of 37 papers
- census · 2 of 37 papers
- genes · 2 of 37 papers
- MNIST · 2 of 37 papers
Baselines
- PPI · 5 of 37 papers
- SAMS-VAE · 2 of 37 papers
- GEARS · 2 of 37 papers
- RePPI · 2 of 37 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 94 papers; from OpenAlex and the papers' affiliations).
- Google · 4 papers
- Massachusetts Institute of Technology · 4 papers
- University of Illinois, Urbana Champaign · 4 papers
- Carnegie Mellon University · 3 papers
- Shanghai Jiao Tong University · 3 papers
- University of Michigan - Ann Arbor · 3 papers
- University of California, Berkeley · 2 papers
- Google DeepMind · 2 papers
- Stanford University · 2 papers
- University of Southern California · 2 papers
Use it
Download BibTeX of 38 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
- Statistical analysis of ML practice (122 papers)
- Machine learning for physical simulation (96 papers)
- Counterfactual and statistical explanations (81 papers)
- Sparse concepts and hypothesis testing (78 papers)
- Cold-start recommendation models (80 papers)
Elsewhere in the topic
- Generative protein structure design (142 papers)
- Neuro-symbolic probabilistic layers (120 papers)
- Structure-based 3D molecule generation (107 papers)
- Sparse graphical and tensor models (97 papers)
- Flow matching generative modeling (97 papers)
- Multimodal trajectory prediction (95 papers)
- Probabilistic embeddings and relational inference (93 papers)
- Equivariant and disentangled latent representations (90 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).