Research map / Optimization / Implicit bias and generalization
Strong Correlations Induce Cause Only Predictions in Transformer Training
Haihan Zhang, Yimu Zhang, Cong Fang
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 |
|---|---|---|---|
| use DeiT-Small instead of ResNet-34 | beard classifier · CelebA beard (Test Set 2, only facial regions) | accuracy: 0.861 -> 0.893 | Improved |
| use DeiT-Small instead of EfficientNet-B4 | beard classifier · CelebA beard (Test Set 2, only facial regions) | accuracy: 0.802 -> 0.893 | Improved |
| use DeiT-Small instead of ResNet-34 | beard classifier · CelebA beard (Test Set 1, facial regions masked) | accuracy: 0.577 -> 0.552 | Improved |
| raise spurious correlation P(z|y) from 0.5 to 0.9 | BERT · Amazon reviews | biased test loss: 0.65 -> 0.68 | Hurt |
| raise spurious correlation P(z|y) from 0.9 to 0.95 | BERT · Amazon reviews | oracle test loss: 0.65 -> 0.77 | Hurt |
| raise training bias strength above 0.9 | DeiT-Small · Waterbirds | test accuracy: ptrain >= 0.9 vs ptrain <= 0.85 | Hurt |
| flip causal label with probability pflip up to 0.2 | DeiT-Small · Waterbirds-like (ptrain 0.7) | test accuracy: over 90% | No clear effect |
| flip causal label with probability pflip 0.4 to 0.5 | DeiT-Small · Waterbirds-like | test accuracy | Hurt |
Closest papers by meaning
- From Shortcut to Induction Head: How Data Diversity Shapes Algorithm Selection in Transformers (NeurIPS 2025 · similarity 0.88)
- From Condensation to Rank Collapse: A Two-Stage Analysis of Transformer Training Dynamics (NeurIPS 2025 · similarity 0.88)
- Training-Induced Escape from Token Clustering in a Mean-Field Formulation of Transformers (NeurIPS 2026 · similarity 0.87)
- Small-scale proxies for large-scale Transformer training instabilities (ICLR 2024 · similarity 0.87)
- Causal Head Gating: A Framework for Interpreting Roles of Attention Heads in Transformers (NeurIPS 2025 · similarity 0.87)
- Attention Sinks Induce Gradient Sinks: Massive Activations as Gradient Regulators in Transformers (NeurIPS 2026 · similarity 0.87)
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.
