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

2025-09-28

AI Summary

This paper investigates whether “weak-to-strong” (W2S) generalization holds—and why it fails—when fine-tuning a strong student model using pseudo-labels generated by a weak teacher, particularly in downstream tasks exhibiting spurious correlations. Theoretical analysis identifies a critical failure mechanism: distributional shift in the minority-group proportion between pseudo-labels and ground-truth labels. To address this, we propose a group-label-free confidence-based subset retraining method that dynamically constructs a robust training set by selecting high-confidence pseudo-labels and modeling subgroup imbalance. Through asymptotic theoretical analysis and extensive experiments across diverse spurious-correlation benchmarks and teacher–student architectures, our method consistently improves W2S generalization and out-of-distribution robustness. It establishes a novel, trustworthy paradigm for knowledge distillation without requiring access to group annotations.

Technology Category

Machine Learning: Calibration & Uncertainty QuantificationNatural Language Processing: Safety and RobustnessComputer Vision: Adversarial Attacks & Robustness

Application Category

Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for rankingSemantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMsWeb Mining and Content Analysis: Robustness and generalizability of Web mining methods

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 $η_ell$, 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 $η_u = η_ell$ but may fail when $η_u e η_ell$, where W2S gain diminishes as $(η_u - η_ell)^2$ 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.

Problem

• Studying weak-to-strong generalization under spurious correlations
• Analyzing failure cases when minority group fractions differ
• Proposing an algorithm to improve weak-to-strong performance

Innovation

• Fine-tuning strong student with weak teacher pseudolabels
• Retraining student on high-confidence data subset
• Group-label-free algorithm improving weak-to-strong generalization