Research map / Label-efficient and robust learning / Recognition and long-tail
Does Weak-to-strong Generalization Happen under Spurious Correlations?
Chenruo Liu, Yijun Dong, Qi Lei
Read the paperCheck your idea against 137,217 papers
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 |
|---|---|---|---|
| retrain strong student on high-confidence subset with GCE loss after W2S fine-tuning (Enhanced-W2S) | vanilla W2S · Waterbirds, BFFHQ, BG-COCO, ImageNet-9 average accuracy | relative accuracy improvement (%): +7.02 / +4.32 / +7.50 / +11.73 mean relative improvement | Improved |
| add Enhanced-W2S to vanilla W2S on a few pairs | vanilla W2S · BFFHQ and BG-COCO (eta_l=eta_o, eta_u=0.5) | relative accuracy improvement (%): -1.37 (Clipb32/MAE BFFHQ); -3.52 (DINOv2/MAE BG-COCO) | Hurt |
| add Enhanced-W2S for worst-group accuracy | vanilla W2S · Waterbirds, BFFHQ, BG-COCO worst-group accuracy | relative worst-group improvement (%): +21.14 / +3.81 / +7.73 | Improved |
| use GCE loss instead of CE loss in retraining (ablation) | Enhanced-W2S · Waterbirds (eta_l=0.5, eta_u=eta_o) | average accuracy (x1e-2): +3.27 over CE variant | Improved |
| use GCE loss instead of CE loss in retraining (ablation) | Enhanced-W2S · ImageNet-9 (eta_l=eta_o, eta_u=0.5) | average accuracy (x1e-2): +0.51 over CE variant | Improved |
| use confidence-based high-confidence subset instead of all unlabeled data (ablation) | Enhanced-W2S · Waterbirds (eta_l=eta_o, eta_u=0.5) | average accuracy (x1e-2): +4.84 over p=1 variant | Improved |
| use confidence-based subset instead of all unlabeled data (ablation) | Enhanced-W2S · ImageNet-9 (eta_l=eta_o, eta_u=0.5) | average accuracy (x1e-2): +5.08 over p=1 variant | Improved |
| increase gap between unlabeled and labeled minority fractions (eta_u vs eta_l) | vanilla W2S · four spurious-correlation benchmarks | W2S gain | Hurt |
| raise minority share of unlabeled data when teacher is group-balanced (eta_l=0.5) | vanilla W2S · four spurious-correlation benchmarks | W2S gain | Improved |
Closest papers by meaning
- On the Mechanisms of Weak-to-Strong Generalization: A Theoretical Perspective (NeurIPS 2025 · similarity 0.91)
- Provable weak-to-strong generalization via benign overfitting (ICLR 2025 · similarity 0.91)
- Theoretical Analysis of Weak-to-Strong Generalization (NeurIPS 2024 · similarity 0.90)
- Discrepancies are Virtue: Weak-to-Strong Generalization through Lens of Intrinsic Dimension (ICML 2025 · similarity 0.90)
- Quantifying the Gain in Weak-to-Strong Generalization (NeurIPS 2024 · similarity 0.89)
- Weak-to-Strong Generalization Even in Random Feature Networks, Provably (ICML 2025 · similarity 0.89)
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.
