Does Weak-to-strong Generalization Happen under Spurious Correlations?
Not on the map. Why?
Abstract (arXiv)
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 ηℓ, and (ii) a group-imbalanced unlabeled set pseudolabeled by the teacher with a minority group of fraction ηu. Theoretically, a precise characterization of W2S gain at the proportional asymptotic limit shows that W2S always happens with sufficient pseudolabels when ηu = ηℓ but may fail when ηu ne ηℓ, where W2S gain diminishes as (ηu - ηℓ)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.
Key facts. Cited by 1 paper in the map (data through 7 Oct 2026). Builds on 71 of its 83 references that are in the map. First posted on arXiv 28 Sep 2025.
Builds on
71 of 83 references.
References · 71
- 2015Deep Residual Learning for Image Recognitioncited by 51,809
- 2021Learning Transferable Visual Models From Natural Language Supervisioncited by 33,456
- 2020Language Models are Few-Shot Learnerscited by 32,199
- 2023GPT-4 Technical Reportcited by 18,132
- 2014Microsoft COCO: Common Objects in Contextcited by 17,498
- 2015Distilling the Knowledge in a Neural Networkcited by 10,076
- 2023DINOv2: Learning Robust Visual Features without Supervisioncited by 7,764
- 2021Masked Autoencoders Are Scalable Vision Learnerscited by 7,339
- 2022A ConvNet for the 2020scited by 3,545
- 2018Neural Tangent Kernel: Convergence and Generalization in Neural Networkscited by 2,597
- 2019A Survey on Bias and Fairness in Machine Learningcited by 1,911
- 2019Invariant Risk Minimizationcited by 1,832
38 more
- 2018Generalized Cross Entropy Loss for Training Deep Neural Networks with Noisy Labelscited by 994
- 2020Out-of-Distribution Generalization via Risk Extrapolation (REx)cited by 737
- 2020Intrinsic Dimensionality Explains the Effectiveness of Language Model Fine-Tuningcited by 658
- 2021Just Train Twice: Improving Group Robustness without Training Group Informationcited by 522
- 2019Surprises in High-Dimensional Ridgeless Least Squares Interpolationcited by 508
- 2023Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervisioncited by 445
- 2018Confounding variables can degrade generalization performance of radiological deep learning modelscited by 374
- 2024The Platonic Representation Hypothesiscited by 362
- 2020Noise or Signal: The Role of Image Backgrounds in Object Recognitioncited by 342
- 2020Fairness in Machine Learningcited by 194
- 2021Does Knowledge Distillation Really Work?cited by 187
- 2022Correct-N-Contrast: A Contrastive Approach for Improving Robustness to Spurious Correlationscited by 171
- 2021Towards Understanding Knowledge Distillationcited by 158
- 2022On Feature Learning in the Presence of Spurious Correlationscited by 145
- 2024Superfiltering: Weak-to-Strong Data Filtering for Fast Instruction-Tuningcited by 134
- 2022A Kernel-Based View of Language Model Fine-Tuningcited by 118
- 2021Learning Debiased Representation via Disentangled Feature Augmentationcited by 116
- 2020On the Optimal Weighted ℓ2 Regularization in Overparameterized Linear Regressioncited by 97
- 2019Fairness risk measurescited by 85
- 2021Counterfactual Invariance to Spurious Correlations: Why and How to Pass Stress Testscited by 77
- 2020Coping with Label Shift via Distributionally Robust Optimisationcited by 69
- 2024A Survey on Autonomous Driving Datasets: Statistics, Annotation Quality, and a Future Outlookcited by 65
- 2023Towards Last-layer Retraining for Group Robustness with Fewer Annotationscited by 54
- 2025Great Models Think Alike and this Undermines AI Oversightcited by 53
- 2022Dimension free ridge regressioncited by 50
- 2022FairLex: A Multilingual Benchmark for Evaluating Fairness in Legal Text Processingcited by 48
- 2024Theoretical Analysis of Weak-to-Strong Generalizationcited by 46
- 2021Out-of-distribution Generalization in the Presence of Nuisance-Induced Spurious Correlationscited by 41
- 2024Quantifying the Gain in Weak-to-Strong Generalizationcited by 40
- 2023Robust Learning with Progressive Data Expansion Against Spurious Correlationcited by 39
- 2024Weak-to-Strong Reasoningcited by 32
- 2024Vision Superalignment: Weak-to-Strong Generalization for Vision Foundation Modelscited by 28
- 2024Co-Supervised Learning: Improving Weak-to-Strong Generalization with Hierarchical Mixture of Expertscited by 28
- 2022What Knowledge Gets Distilled in Knowledge Distillation?cited by 26
- 2023FACTS: First Amplify Correlations and Then Slice to Discover Biascited by 26
- 2024Super(ficial)-alignment: Strong Models May Deceive Weak Models in Weak-to-Strong Generalizationcited by 23
- 2024High-dimensional Analysis of Knowledge Distillation: Weak-to-Strong Generalization and Scaling Lawscited by 21
- 2024Weak-to-Strong Generalization Through the Data-Centric Lenscited by 21
50 most cited of 71 All 71 (JSON)
Built on by
Citing papers per year: 2026: 1.
1 paper.