Research map / Label-efficient and robust learning / Test-time adaptation
Improving Out-of-Distribution Robustness via Selective Augmentation
Huaxiu Yao, Yu Wang, Sai Li, Linjun Zhang, Weixin Liang, James Zou, Chelsea Finn
This paper tried a common change: Add mixup or CutMix
use vanilla mixup instead of LISA's selective pairing · Camelyon17, FMoW, RxRx1, Amazon, MetaShift · dataset evaluation metric: 71.2 $\pm$ 5.3\% vs. 77.1 $\pm$ 6.5\%; 34.2 $\pm$ 0.45\% vs. 35.5 $\pm$ 0.65\%; 26.5 $\pm$ 0.5\% vs. 31.9 $\pm$ 0.8\%; 5 Mixed
What 547 accepted papers report for this change
- 416 improved
- 28 no clear effect
- 14 mixed
- 89 hurt
- Lightweight Federated Incremental Learning via Decoupled Replay ICML 2026 · add mixup to uploaded features · CIFAR100 (α=1.0) · A(f): 33.20 -> 35.61Improved
- Quantifying and Optimizing Simplicity via Polynomial Representations ICML 2026 · add embedding mixup · GLUE RTE, MRPC, CoLA · accuracy / Matthews corr.: MRPC 86.74 -> 86.30; CoLA 62.31 -> 61.04; RTE unchanged 70.28Hurt
- Fix the Mind, Not the Move: Interpretable AI Assistance via Knowledge-Gap Localization ICML 2026 · add auxiliary mixup generalization loss · Breakfast · System recall: 0.221 -> 0.762Improved
See this change in detailMatched by the wording of each paper's own result, read by AI from the full text. Papers report gains more often than failures, so this shows what gets published, not the chance a change works.
AI-extracted. The measured effects below were read by AI from the paper's full text (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 selective interpolation across same-label/different-domain or same-domain/different-label pairs rather than vanilla mixup (ablation) | LISA · CMNIST | worst-group accuracy: 3.1% -> 73.3% | Improved |
| use selective interpolation across same-label/different-domain or same-domain/different-label pairs rather than vanilla mixup (ablation) | LISA · Waterbirds | worst-group accuracy: 56.2% -> 89.2% | Improved |
| use selective interpolation across same-label/different-domain or same-domain/different-label pairs rather than vanilla mixup (ablation) | LISA · CelebA | worst-group accuracy: 46.4% -> 89.3% | Improved |
| use selective interpolation across same-label/different-domain or same-domain/different-label pairs rather than vanilla mixup (ablation) | LISA · CivilComments | worst-group accuracy: 67.2% -> 72.6% | Improved |
| interpolate samples from the same label but different domains (ablation) | LISA · CMNIST, Waterbirds, CelebA | worst-group accuracy | Improved |
| interpolate samples from the same domain but different labels (ablation) | LISA · CMNIST, Waterbirds, CelebA | worst-group accuracy | Improved |
| restrict interpolation to samples sharing both label and domain (ablation) | mixup for OOD robustness · CMNIST, Waterbirds, CelebA, CivilComments | worst-group accuracy: LISA 73.3% vs. 24.0% (CMNIST); 89.2% vs. 68.0% (Waterbirds); 89.3% vs. 58.3% (CelebA); 72.6% vs. 69.2% (CivilComments) | Hurt |
| use vanilla mixup instead of LISA's selective pairing (ablation) | LISA · Camelyon17, FMoW, RxRx1, Amazon, MetaShift | dataset evaluation metric: 71.2 $\pm$ 5.3\% vs. 77.1 $\pm$ 6.5\%; 34.2 $\pm$ 0.45\% vs. 35.5 $\pm$ 0.65\%; 26.5 $\pm$ 0.5\% vs. 31.9 $\pm$ 0.8\%; 5 | Mixed |
| increase the training-to-test domain distance | LISA versus best baseline · MetaShift | worst-group accuracy gain over best baseline: +1.5%, +1.2%, +3.6%, +9.6% | Improved |
| balance groups to remove spurious correlations (ablation) | LISA · CMNIST, Waterbirds, CelebA | accuracy: ERM 73.67%, 88.07%, 86.11%; LISA 73.18%, 87.05%, 87.22% | No clear effect |
Closest papers by meaning
- Sufficient Invariant Learning for Distribution Shift (CVPR 2025 · similarity 0.87)
- Selective Mixup Helps with Distribution Shifts, But Not (Only) because of Mixup (ICML 2024 · similarity 0.86)
- Out of the Ordinary: Spectrally Adapting Regression for Covariate Shift (ICML 2024 · similarity 0.85)
- Tolerant Algorithms for Learning with Arbitrary Covariate Shift (NeurIPS 2024 · similarity 0.85)
- Learning Conditional Invariances through Non-Commutativity (ICLR 2024 · similarity 0.84)
- What Does It Take to Build a Performant Selective Classifier? (NeurIPS 2025 · similarity 0.84)
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.
