mlsanityData to 7 Oct 2026

Discrepancies are Virtue: Weak-to-Strong Generalization through Lens of Intrinsic Dimension

arXiv 2502.05075 ↗First posted 7 Feb 2025cs.LG cs.NA math.NAcited by 10

Not on the map. Why?

Abstract (arXiv)

Weak-to-strong (W2S) generalization is a type of finetuning (FT) where a strong (large) student model is trained on pseudo-labels generated by a weak teacher. Surprisingly, W2S FT often outperforms the weak teacher. We seek to understand this phenomenon through the observation that FT often occurs in intrinsically low-dimensional spaces. Leveraging the low intrinsic dimensionality of FT, we analyze W2S in the ridgeless regression setting from a variance reduction perspective. For a strong student-weak teacher pair with sufficiently expressive low-dimensional feature subspaces Vs, Vw, we provide an exact characterization of the variance that dominates the generalization error of W2S. This unveils a virtue of discrepancy between the strong and weak models in W2S: the variance of the weak teacher is inherited by the strong student in Vs cap Vw, while reduced by a factor of dim(Vs)/N in the subspace of discrepancy Vw setminus Vs with N pseudo-labels for W2S. Our analysis further casts light on the sample complexities and the scaling of performance gap recovery in W2S. The analysis is supported by experiments on synthetic regression problems, as well as real vision and NLP tasks.

Key facts. Cited by 10 papers in the map (data through 7 Oct 2026). Builds on 56 of its 86 references that are in the map. First posted on arXiv 7 Feb 2025.

Builds on

56 of 86 references.

References · 56

  1. 2015Deep Residual Learning for Image Recognitioncited by 51,809
  2. 2021Learning Transferable Visual Models From Natural Language Supervisioncited by 33,456
  3. 2015Distilling the Knowledge in a Neural Networkcited by 10,076
  4. 2023DINOv2: Learning Robust Visual Features without Supervisioncited by 7,764
  5. 2018Neural Tangent Kernel: Convergence and Generalization in Neural Networkscited by 2,597
  6. 2019Invariant Risk Minimizationcited by 1,832
  7. 2020Knowledge Distillation: A Surveycited by 1,578
  8. 2020Intrinsic Dimensionality Explains the Effectiveness of Language Model Fine-Tuningcited by 658
  9. 2019Benign Overfitting in Linear Regressioncited by 544
  10. 2018Born Again Neural Networkscited by 539
  11. 2024LESS: Selecting Influential Data for Targeted Instruction Tuningcited by 514
  12. 2019Be Your Own Teacher: Improve the Performance of Convolutional Neural Networks via Self Distillationcited by 491
38 more
  1. 2023Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervisioncited by 445
  2. 2018Bag of Tricks for Image Classification with Convolutional Neural Networkscited by 430
  3. 2017Age Progression/Regression by Conditional Adversarial Autoencodercited by 426
  4. 2018Measuring the Intrinsic Dimension of Objective Landscapescited by 364
  5. 2021Explaining Neural Scaling Lawscited by 353
  6. 2020Kernel and Rich Regimes in Overparametrized Modelscited by 302
  7. 2014Sketching as a Tool for Numerical Linear Algebracited by 285
  8. 2021Does Knowledge Distillation Really Work?cited by 187
  9. 2020Theoretical Analysis of Self-Training with Deep Networks on Unlabeled Datacited by 179
  10. 2020Self-Distillation Amplifies Regularization in Hilbert Spacecited by 171
  11. 2021Towards Understanding Knowledge Distillationcited by 158
  12. 2024Superfiltering: Weak-to-Strong Data Filtering for Fast Instruction-Tuningcited by 134
  13. 2022A Kernel-Based View of Language Model Fine-Tuningcited by 118
  14. 2020Rethinking Bias-Variance Trade-off for Generalization of Neural Networkscited by 107
  15. 2021Towards the Generalization of Contrastive Self-Supervised Learningcited by 99
  16. 2020Classification vs regression in overparameterized regimes: Does the loss function matter?cited by 97
  17. 2024Easy-to-Hard Generalization: Scalable Alignment Beyond Human Supervisioncited by 95
  18. 2022More Than a Toy: Random Matrix Models Predict How Real-World Neural Representations Generalizecited by 87
  19. 2022Connect, Not Collapse: Explaining Contrastive Learning for Unsupervised Domain Adaptationcited by 66
  20. 2025Great Models Think Alike and this Undermines AI Oversightcited by 53
  21. 2021A Theory of Label Propagation for Subpopulation Shiftcited by 48
  22. 2024Theoretical Analysis of Weak-to-Strong Generalizationcited by 46
  23. 2024Quantifying the Gain in Weak-to-Strong Generalizationcited by 40
  24. 2019Exact minimax risk for linear least squares, and the lower tail of sample covariance matricescited by 34
  25. 2024Weak-to-Strong Reasoningcited by 32
  26. 2020When Does Preconditioning Help or Hurt Generalization?cited by 30
  27. 2022Sample Efficiency of Data Augmentation Consistency Regularizationcited by 28
  28. 2024Vision Superalignment: Weak-to-Strong Generalization for Vision Foundation Modelscited by 28
  29. 2024Co-Supervised Learning: Improving Weak-to-Strong Generalization with Hierarchical Mixture of Expertscited by 28
  30. 2022What Knowledge Gets Distilled in Knowledge Distillation?cited by 26
  31. 2024Super(ficial)-alignment: Strong Models May Deceive Weak Models in Weak-to-Strong Generalizationcited by 23
  32. 2024High-dimensional Analysis of Knowledge Distillation: Weak-to-Strong Generalization and Scaling Lawscited by 21
  33. 2024Weak-to-Strong Generalization Through the Data-Centric Lenscited by 21
  34. 2023Understanding Self-Distillation in the Presence of Label Noisecited by 20
  35. 2021Self-training Converts Weak Learners to Strong Learners in Mixture Modelscited by 19
  36. 2023On student-teacher deviations in distillation: does it pay to disobey?cited by 16
  37. 2024Provable Weak-to-Strong Generalization via Benign Overfittingcited by 16
  38. 2024Improving Weak-to-Strong Generalization with Reliability-Aware Alignmentcited by 14

50 most cited of 56 All 56 (JSON)

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2025: 4; 2026: 6.

10 papers.

  1. 2025On the Mechanisms of Weak-to-Strong Generalization: A Theoretical Perspectivecited by 7
  2. 2025Weak-to-Strong Generalization under Distribution Shiftscited by 4
  3. 2025Weak-to-Strong Generalization via Bregman Bias-Variance Decompositioncited by 3
  4. 2026Eliciting Weak-to-Strong Generalization with On-Policy Reverse Distillationcited by 2
  5. 2025Does Weak-to-strong Generalization Happen under Spurious Correlations?cited by 1
  6. 2026Weak-to-Strong Generalization is Nearly Inevitable (in Linear Models)cited by 1
  7. 2026The Mechanism of Weak-to-Strong Generalization: Feature Elicitation from Latent Knowledgecited by 1
  8. 2026On the Blessing of Pre-training in Weak-to-Strong Generalizationcited by 0
  9. 2026Gradient Transformer: Learning to Generate Updates for LLMscited by 0
  10. 2026What Makes a Strong Model? A Unified Spectral Analysis of Knowledge Transfer over High-dimensional Linear Regressioncited by 0