Does Weak-to-strong Generalization Happen under Spurious Correlations?

2025/09/28 by Chenruo Liu, Liu, Chenruo, Yijun Dong +3
Computer Science · #FOS: Computer and information sciences #Machine Learning (cs.LG) #Machine Learning (stat.ML) #Neural Networks and Applications

paper · pdf · doi:10.48550/arxiv.2509.24005

openalex publication_date 2025/09/28 · openalex created_date 2025/10/19 · openalex updated_date 2026/07/28

Abstract

We initiate a unified theoretical and algorithmic study of a key problem in weak-to-strong (W2S) generalization: when fine-tuning a strong pre-trained student with pseudolabels from a weaker teacher on a downstream task with spurious correlations, does W2S happen, and how to improve it upon failures? We consider two sources of spurious correlations caused by group imbalance: (i) a weak teacher fine-tuned on group-imbalanced labeled data with a minority group of fraction η_ℓ, and (ii) a group-imbalanced unlabeled set pseudolabeled by the teacher with a minority group of fraction η_u. Theoretically, a precise characterization of W2S gain at the proportional asymptotic limit shows that W2S always happens with sufficient pseudolabels when η_u = η_ℓ but may fail when η_u ≠ η_ℓ, where W2S gain diminishes as (η_u - η_ℓ)² increases. Our theory is corroborated by extensive experiments on various spurious correlation benchmarks and teacher-student pairs. To boost W2S performance upon failures, we further propose a simple, effective algorithmic remedy that retrains the strong student on its high-confidence data subset after W2S fine-tuning. Our algorithm is group-label-free and achieves consistent, substantial improvements over vanilla W2S fine-tuning.

Citations

• From Linear to Nonlinear: Provable Weak-to-Strong Generalization through Feature Learning
• Weak-to-Strong Generalization under Distribution Shifts
• Superclass-Guided Representation Disentanglement for Spurious Correlation Mitigation
• On the Emergence of Weak-to-Strong Generalization: A Bias-Variance Perspective
• Bridging Distribution Shift and AI Safety: Conceptual and Methodological Synergies
• On the Mechanisms of Weak-to-Strong Generalization: A Theoretical Perspective
• Weak-to-Strong Generalization Even in Random Feature Networks, Provably
• Elastic Representation: Mitigating Spurious Correlations for Group Robustness
• Discrepancies are Virtue: Weak-to-Strong Generalization through Lens of Intrinsic Dimension
• Great Models Think Alike and this Undermines AI Oversight
• Spurious Correlations in High Dimensional Regression: The Roles of Regularization, Simplicity Bias and Over-Parameterization
• Representations Shape Weak-to-Strong Generalization: Theoretical Insights and Empirical Predictions
• Relating Misfit to Gain in Weak-to-Strong Generalization Beyond the Squared Loss
• Weak-to-Strong Generalization Through the Data-Centric Lens
• Adaptive Group Robust Ensemble Knowledge Distillation
• High-dimensional Analysis of Knowledge Distillation: Weak-to-Strong Generalization and Scaling Laws
• Provable Weak-to-Strong Generalization via Benign Overfitting
• Weak-to-Strong Reasoning
• Improving Weak-to-Strong Generalization with Reliability-Aware Alignment
• Mind the Graph When Balancing Data for Fairness or Robustness
• Super(ficial)-alignment: Strong Models May Deceive Weak Models in Weak-to-Strong Generalization
• Theoretical Analysis of Weak-to-Strong Generalization
• Quantifying the Gain in Weak-to-Strong Generalization
• The Platonic Representation Hypothesis
• Improving Group Robustness on Spurious Correlation Requires Preciser Group Inference
• Controllable Prompt Tuning For Balancing Group Distributional Robustness
• Co-Supervised Learning: Improving Weak-to-Strong Generalization with Hierarchical Mixture of Experts
• Vision Superalignment: Weak-to-Strong Generalization for Vision Foundation Models
• Superfiltering: Weak-to-Strong Data Filtering for Fast Instruction-Tuning
• A Survey on Autonomous Driving Datasets: Statistics, Annotation Quality, and a Future Outlook
• Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision
• Group Distributionally Robust Knowledge Distillation
• FACTS: First Amplify Correlations and Then Slice to Discover Bias
• Towards Last-layer Retraining for Group Robustness with Fewer Annotations
• Cluster-aware Semi-supervised Learning: Relational Knowledge Distillation Provably Learns Clustering
• Robust Learning with Progressive Data Expansion Against Spurious Correlation
• DINOv2: Learning Robust Visual Features without Supervision
• GPT-4 Technical Report
• Debiased Distillation by Transplanting the Last Layer
• On student-teacher deviations in distillation: does it pay to disobey?
• On Feature Learning in the Presence of Spurious Correlations
• Dimension free ridge regression
• A Kernel-Based View of Language Model Fine-Tuning
• What Knowledge Gets Distilled in Knowledge Distillation?
• FairLex: A Multilingual Benchmark for Evaluating Fairness in Legal Text\n Processing
• Correct-N-Contrast: A Contrastive Approach for Improving Robustness to Spurious Correlations
• A ConvNet for the 2020s
• A ConvNet for the 2020s
• Masked Autoencoders Are Scalable Vision Learners
• Just Train Twice: Improving Group Robustness without Training Group\n Information
• Learning Debiased Representation via Disentangled Feature Augmentation
• Out-of-distribution Generalization in the Presence of Nuisance-Induced Spurious Correlations
• Teacher's pet: understanding and mitigating biases in distillation
• Does Knowledge Distillation Really Work?
• Counterfactual Invariance to Spurious Correlations: Why and How to Pass Stress Tests
• Learning Transferable Visual Models From Natural Language Supervision
• Intrinsic Dimensionality Explains the Effectiveness of Language Model\n Fine-Tuning
• Coping with Label Shift via Distributionally Robust Optimisation
• Noise or Signal: The Role of Image Backgrounds in Object Recognition
• On the Optimal Weighted ℓ₂ Regularization in Overparameterized Linear Regression
• Language Models are Few-Shot Learners
• Out-of-Distribution Generalization via Risk Extrapolation (REx)
• A Survey on Bias and Fairness in Machine Learning
• Invariant Risk Minimization
• Towards Understanding Knowledge Distillation
• Surprises in High-Dimensional Ridgeless Least Squares Interpolation
• Fairness risk measures
• Neural Tangent Kernel: Convergence and Generalization in Neural Networks
• Generalized Cross Entropy Loss for Training Deep Neural Networks with Noisy Labels
• Deep Residual Learning for Image Recognition
• Distilling the Knowledge in a Neural Network
• Microsoft COCO: Common Objects in Context
• The Capabilities and Limitations of Weak-to-Strong Generalization: Generalization and Calibration

Related

• Provable Weak-to-Strong Generalization via Benign Overfitting
• Theoretical Analysis of Weak-to-Strong Generalization
• On the Emergence of Weak-to-Strong Generalization: A Bias-Variance Perspective
• On the Mechanisms of Weak-to-Strong Generalization: A Theoretical Perspective
• Discrepancies are Virtue: Weak-to-Strong Generalization through Lens of Intrinsic Dimension