Research map / Learning theory and algorithms / Prediction and decision losses
Provable post-hoc explanations
Formalize and certify fidelity, robustness and sensitivity of black-box model explanations.
92 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2017–2026) · an approach in Prediction and decision losses, keyword cluster “explanations · shapley · attribution”
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.
28 papers in 2023–24 and 28 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic held steady, from 5.2% to 4.9%.
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 · Axiomatic Attribution for Deep Networks (ICML 2017 · 2,911 citations)
- 2018 · Anchors: High-Precision Model-Agnostic Explanations (AAAI 2018 · 2,125 citations)
- 2019 · Explanations can be manipulated and geometry is to blame (NeurIPS 2019 · 144 citations)
- 2020 · Model Interpretability through the lens of Computational Complexity (NeurIPS 2020 · 38 citations)
- 2021 · Learning explanations that are hard to vary (ICLR 2021 · 69 citations)
- 2022 · Towards Better Understanding Attribution Methods (CVPR 2022 · 36 citations)
- 2023 · Shortcomings of Top-Down Randomization-Based Sanity Checks for Evaluations of Deep Neural Network Explanations (CVPR 2023 · 21 citations)
- 2024 · Integrated Decision Gradients: Compute Your Attributions Where the Model Makes Its Decision (AAAI 2024 · 16 citations)
- 2025 · Explaining, Fast and Slow: Abstraction and Refinement of Provable Explanations (ICML 2025 · 0 citations)
- 2026 · NIMO: a Nonlinear Interpretable MOdel (ICLR 2026 · 1 citation)
Earliest papers here
- Axiomatic Attribution for Deep Networks (ICML 2017 · 2,911 citations)
- Understanding Black-box Predictions via Influence Functions (ICML 2017 · 1,328 citations)
- Streaming Weak Submodularity: Interpreting Neural Networks on the Fly (NeurIPS 2017 · 14 citations)
What has been tried
67 papers here report 556 measured changes (a component added, removed or swapped, compared with a baseline): helped 64%, no clear effect 10%, hurt 21%, mixed 6%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| add adaptive sampling to IG | ImageNet, ResNet101 · Insertion | 0.513 -> 0.532 | Helped | Integrated Decision Gradients: Compute Your Attributions Where the Model Makes Its Decision (AAAI 2024 · 16 citations) |
| Use second-order Taylor expansion instead of first-order only | ImageNet with EfficientNet · Error Rate | 0.02058 -> 0.01165 | Helped | MFABA: A More Faithful and Accelerated Boundary-Based Attribution Method for Deep Neural Networks (AAAI 2024 · 9 citations) |
| Increase the Anchor confidence parameter δ | Classification tasks evaluated with DeBERTa · AOPC | Helped | Accelerating the Global Aggregation of Local Explanations (AAAI 2024 · 2 citations) | |
| use only original features x for allocator training | Allocator feature-set study · datasets not significantly worse than best feature set (α=0.05) | 18.75% | Hurt | Learning Performance Maximizing Ensembles with Explainability Guarantees (AAAI 2024 · 1 citation) |
| use fixed reduction rate 10% without refinement | MNIST · explanation size | 204.4 ± 129.2 -> 507.3 ± 141.3 | Hurt | Explaining, Fast and Slow: Abstraction and Refinement of Provable Explanations (ICML 2025 · 0 citations) |
| Increase the TreeHFD fitting sample size | Analytical case · cumulative MSE | 1.94 -> 0.14 | Helped | Tree Ensemble Explainability through the Hoeffding Functional Decomposition and TreeHFD Algorithm (NeurIPS 2025 · 0 citations) |
| use variance minimization rather than empirical error minimization | Boston housing · MSE | Helped | Minimal Variance Model Aggregation: A principled, non-intrusive, and versatile integration of black box models (ICLR 2025 · 0 citations) | |
| Add interactions to self-context Shapley values | Visual SET-SUM task on synthetic MNIST · insertion accuracy | Helped | Identifying Important Group of Pixels using Interactions (CVPR 2024 · 0 citations) | |
| Apply FORGrad gradient filtering to Saliency | ImageNet validation · Faithfulness | 0.18 -> 0.28 | Helped | Saliency strikes back: How filtering out high frequencies improves white-box explanations (ICML 2024 · 0 citations) |
| Add flexible reference adjustment | Multiple datasets and models · SEV | Helped | Improving Decision Sparsity (NeurIPS 2024 · 0 citations) | |
| apply DSI guarantees during fine-tuning | WikiText-5 · exposure metric | 2.2 -> 0.8 | Helped | Trustworthy Machine Learning through Data-Specific Indistinguishability (ICML 2025) |
| Replace CCA loss with binary cross-entropy using labels 0, 1 and 0.5 | Experiments comparing alternate loss functions · fidelity | Hurt | Model Reconstruction Using Counterfactual Explanations: A Perspective From Polytope Theory (NeurIPS 2024) |
What reviewers will expect
Datasets and baselines named in the 19 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- CIFAR-10 · 2 of 19 papers
Baselines
- Integrated Gradients · 3 of 19 papers
- LIME · 2 of 19 papers
- KernelSHAP · 2 of 19 papers
- SHAP · 2 of 19 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 56 papers; from OpenAlex and the papers' affiliations).
- Hebrew University of Jerusalem · 5 papers
- University of Pennsylvania · 3 papers
- Eberhard-Karls-Universität Tübingen · 3 papers
- Technische Universität Berlin · 3 papers
- CNRS · 2 papers
- The Alan Turing Institute · 2 papers
- Tsinghua University · 2 papers
- TU Berlin · 2 papers
- University of Cambridge · 2 papers
- Google · 2 papers
Use it
Download BibTeX of 39 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
- Shapley value attribution (63 papers)
- Counterfactual explanations and recourse (46 papers)
Elsewhere in the topic
- Adaptive testing and experiments (201 papers)
- Active learning and subset selection (199 papers)
- Distributionally robust optimization (184 papers)
- Data attribution and recourse (160 papers)
- Theory of neural network sensitivity (143 papers)
- Ranking models from comparisons (127 papers)
- Determinantal point process inference (122 papers)
- Conformal prediction sets (116 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).