Research map / Label-efficient and robust learning / Test-time adaptation
Can You Rely on Your Model Evaluation? Improving Model Evaluation with Synthetic Test Data
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 3S synthetic test data instead of mean-shift estimation on real data | Model performance evaluation under feature shifts · Adult, Support, Bank, Drug, and SEER | mean error: Adult: 5.9 -> 2.6; Support: 19.3 -> 2.0; Bank: 18.6 -> 5.4; Drug: 18.5 -> 5.6; SEER: 3.3 -> 2.7 | Helped |
| Use 3S synthetic test data instead of rejection sampling on real data | Model performance evaluation under feature shifts · Adult, Support, Bank, Drug, and SEER | mean error: Adult: 15.9 -> 2.6; Support: 25.1 -> 2.0; Bank: 18.9 -> 5.4; Drug: 20.1 -> 5.6; SEER: 20.0 -> 2.7 | Helped |
| Use 3S synthetic test data with prior target-feature information | Target-domain performance estimation · SEER to CUTRACT | mean performance estimate error: 0.258 -> 0.023 | Helped |
| Use 3S instead of source rejection sampling | Target-domain performance estimation · SEER to CUTRACT | mean performance estimate error: 0.180 -> 0.023 | Helped |
| Use 3S instead of target-domain performance estimation baselines | Target-domain performance estimation · SEER to CUTRACT | mean performance estimate error: ATC: 0.249 -> 0.023; IM: 0.215 -> 0.023; DOC: 0.201 -> 0.023 | Helped |
| Use 3S rather than the hold-out test set for intersectional evaluation | Subgroup performance estimation · Intersectional subgroups | MAE: 20.29 ± 0.14 -> 11.90 ± 0.19 | Helped |
| Use a deep generative ensemble with K=5 or K=10 rather than baselines for uncertainty intervals | 3S uncertainty estimation · Five datasets | coverage and width | Helped |
| Increase the number of independently trained generators from K=1 to K=5 or K=10 (ablation) | 3S uncertainty estimation · Five datasets | coverage and width | Helped |
| Add synthetic data to the real test set | 3S subgroup evaluation (3S+) · Minority subgroups across five datasets | MAE | Helped |
| Provide target-domain information for more features to condition synthetic generation | 3S target-domain estimation · SEER to CUTRACT | performance estimate error | Helped |
Closest papers by meaning
- Debiasing Synthetic Data Generated by Deep Generative Models (NeurIPS 2024 · similarity 0.85)
- Position: All Current Generative Fidelity and Diversity Metrics are Flawed (ICML 2025 · similarity 0.84)
- Evaluating Model Bias Requires Characterizing its Mistakes (ICML 2024 · similarity 0.84)
- Collapse or Thrive: Perils and Promises of Synthetic Data in a Self-Generating World (ICML 2025 · similarity 0.84)
- Post-Selection Distributional Model Evaluation (NeurIPS 2026 · similarity 0.83)
- Maximizing the Potential of Synthetic Data: Insights from Random Matrix Theory (ICLR 2025 · similarity 0.83)
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.