Research map / Label-efficient and robust learning / Recognition and long-tail
The Out-of-Distribution Problem in Explainability and Search Methods for Feature Importance Explanations
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Parallel Local Search instead of Random Search for explanation search | BERT-Base with Counterfactual Training · BoolQ | Comprehensiveness: 31.62 -> 47.80 | Helped |
| Use Parallel Local Search instead of Random Search for explanation search | BERT-Base with Counterfactual Training · SNLI | Sufficiency: 15.38 -> 14.16 | Helped |
| Use Parallel Local Search instead of Random Search for explanation search | BERT-Base with Counterfactual Training · Evidence Inference | Sufficiency: -12.69 -> -20.33 | Hurt |
| Use Parallel Local Search instead of Random Search for explanation search | BERT-Base with Counterfactual Training · FEVER | Comprehensiveness: 18.84 -> 24.07 | Helped |
| Use Parallel Local Search instead of Random Search for explanation search | BERT-Base with Counterfactual Training · MultiRC | Comprehensiveness: 22.19 -> 26.96 | Helped |
| Train on randomized feature-removal counterfactuals | BERT-Base · SNLI, BoolQ, FEVER, Evidence Inference, SST-2, MultiRC | accuracy: 85.84 (0.69) -> 85.08 (0.71); 74.16 (1.62) -> 73.76 (1.62); 89.66 (0.76) -> 89.72 (0.76); 58.81 (3.12) -> 57.35 (3.13); | Mixed |
| Train with counterfactual inputs rather than standard training | BERT-Base · SNLI and SST-2; random-token ablations | accuracy on ablated inputs: +22.9 points | Helped |
| Use Attention Mask or MASK Token replacement rather than the weaker replacement functions | Standard BERT and RoBERTa · SNLI and SST-2; ablation robustness | accuracy: up to 1.61 points; 5.48 points; 2.42 points; 4.72 points | Helped |
| Use Counterfactual Training with Attention Mask, MASK Token, or Slice Out rather than standard training | BERT-Base · SNLI; random search Comprehensiveness | Comprehensiveness: by up to 24.16 points | Hurt |
| Use PLS rather than LIME for explanation search | BERT-Base explanation methods · Across six text classification datasets | Comprehensiveness: up to 20.8 points | Helped |
| Use PLS rather than Random Search for explanation search | BERT-Base explanation methods · Across six text classification datasets | Comprehensiveness: up to 17 points | Helped |
| Use PLS rather than Random Search for explanation search | BERT-Base explanation methods · Across six text classification datasets | Sufficiency: up to 7.6 points | Helped |
Closest papers by meaning
- Discriminative Feature Attributions: Bridging Post Hoc Explainability and Inherent Interpretability (NeurIPS 2023 · similarity 0.89)
- FIMAP: Feature Importance by Minimal Adversarial Perturbation (AAAI 2021 · similarity 0.89)
- Feature Attribution with Necessity and Sufficiency via Dual-stage Perturbation Test for Causal Explanation (ICML 2024 · similarity 0.89)
- CXPlain: Causal Explanations for Model Interpretation under Uncertainty (NeurIPS 2019 · similarity 0.89)
- Regional Explanations: Bridging Local and Global Variable Importance (NeurIPS 2025 · similarity 0.88)
- GEFA: A General Feature Attribution Framework Using Proxy Gradient Estimation (ICML 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.