mlsanityData to 14 Oct 2026

Theoretical Analysis of Weak-to-Strong Generalization

arXiv 2405.16043 ↗First posted 25 May 2024cs.LG cs.CL stat.MLcited by 48

Not on the map. Why?

Abstract (arXiv)

Strong student models can learn from weaker teachers: when trained on the predictions of a weaker model, a strong pretrained student can learn to correct the weak model's errors and generalize to examples where the teacher is not confident, even when these examples are excluded from training. This enables learning from cheap, incomplete, and possibly incorrect label information, such as coarse logical rules or the generations of a language model. We show that existing weak supervision theory fails to account for both of these effects, which we call pseudolabel correction and coverage expansion, respectively. We give a new bound based on expansion properties of the data distribution and student hypothesis class that directly accounts for pseudolabel correction and coverage expansion. Our bounds capture the intuition that weak-to-strong generalization occurs when the strong model is unable to fit the mistakes of the weak teacher without incurring additional error. We show that these expansion properties can be checked from finite data and give empirical evidence that they hold in practice.

Key facts. Cited by 48 papers in the map (data through 14 Oct 2026). Builds on 47 of its 72 references that are in the map. First posted on arXiv 25 May 2024.

Builds on

47 of 72 references.

References · 47

  1. 2019Sentence-BERT: Sentence Embeddings using Siamese BERT-Networkscited by 9,170
  2. 2022Self-Instruct: Aligning Language Models with Self-Generated Instructionscited by 2,724
  3. 2019CheXpert: A Large Chest Radiograph Dataset with Uncertainty Labels and Expert Comparisoncited by 1,423
  4. 2022MTEB: Massive Text Embedding Benchmarkcited by 640
  5. 2023Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervisioncited by 467
  6. 2019Confident Learning: Estimating Uncertainty in Dataset Labelscited by 413
  7. 2017Snorkel: Rapid Training Data Creation with Weak Supervisioncited by 372
  8. 2021Provable Guarantees for Self-Supervised Deep Learning with Spectral Contrastive Losscited by 313
  9. 2016Data Programming: Creating Large Training Sets, Quicklycited by 246
  10. 2022Is GPT-3 a Good Data Annotator?cited by 208
  11. 2022Large Language Models are Few-Shot Clinical Information Extractorscited by 207
  12. 2021Want To Reduce Labeling Cost? GPT-3 Can Helpcited by 196
35 more
  1. 2021Does Knowledge Distillation Really Work?cited by 190
  2. 2020Theoretical Analysis of Self-Training with Deep Networks on Unlabeled Datacited by 180
  3. 2023AnnoLLM: Making Large Language Models to Be Better Crowdsourced Annotatorscited by 172
  4. 2023Self-Alignment with Instruction Backtranslationcited by 151
  5. 2020Understanding Self-Training for Gradual Domain Adaptationcited by 149
  6. 2018Training Complex Models with Multi-Task Weak Supervisioncited by 106
  7. 2020Fast and Three-rious: Speeding Up Weak Supervision with Triplet Methodscited by 98
  8. 2021WRENCH: A Comprehensive Benchmark for Weak Supervisioncited by 96
  9. 2020Fine-Tuning Pre-trained Language Model with Weak Supervision: A Contrastive-Regularized Self-Training Approachcited by 83
  10. 2023Generalization on the Unseen, Logic Reasoning and Degree Curriculumcited by 73
  11. 2020Self-training Avoids Using Spurious Features Under Domain Shiftcited by 69
  12. 2018Weakly-Supervised Neural Text Classificationcited by 66
  13. 2022A Survey on Programmatic Weak Supervisioncited by 66
  14. 2019Improved Sample Complexities for Deep Networks and Robust Classification via an All-Layer Margincited by 56
  15. 2017Learning From Noisy Singly-labeled Datacited by 50
  16. 2021Self-Training with Weak Supervisioncited by 49
  17. 2021A Theory of Label Propagation for Subpopulation Shiftcited by 48
  18. 2023ChatGPT: Beginning of an End of Manual Linguistic Data Annotation? Use Case of Automatic Genre Identificationcited by 45
  19. 2022Beyond Separability: Analyzing the Linear Transferability of Contrastive Representations to Related Subpopulationscited by 44
  20. 2021End-to-End Weak Supervisioncited by 33
  21. 2019Learning Dependency Structures for Weak Supervision Modelscited by 32
  22. 2022Co-training Improves Prompt-based Learning for Large Language Modelscited by 24
  23. 2022Honest Students from Untrusted Teachers: Learning an Interpretable Question-Answering Pipeline from a Pretrained Language Modelcited by 20
  24. 2021Self-training Converts Weak Learners to Strong Learners in Mixture Modelscited by 19
  25. 2022Training Subset Selection for Weak Supervisioncited by 19
  26. 2020Strength from Weakness: Fast Learning Using Weak Supervisioncited by 18
  27. 2022Shoring Up the Foundations: Fusing Model Embeddings and Weak Supervisioncited by 18
  28. 2022Losses over Labels: Weakly Supervised Learning via Direct Loss Constructioncited by 11
  29. 2022Label Propagation with Weak Supervisioncited by 9
  30. 2022Understanding Programmatic Weak Supervision via Source-aware Influence Functioncited by 8
  31. 2020Self-supervised self-supervision by combining deep learning and probabilistic logiccited by 7
  32. 2023Unified Risk Analysis for Weakly Supervised Learningcited by 5
  33. 2021Disambiguation of weak supervision with exponential convergence ratescited by 4
  34. 2015Improved Cheeger's Inequality and Analysis of Local Graph Partitioning using Vertex Expansion and Expansion Profilecited by 2
  35. 2023Higher-Order Cheeger Inequality for Partitioning with Bufferscited by 1

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2024: 13; 2025: 21; 2026: 14.

48 papers.

Sort:

By area:

  1. 2024What is the Role of Small Models in the LLM Era: A Surveycited by 41
  2. 2024Weak-to-Strong Reasoningcited by 33
  3. 2024Selective Preference Optimization via Token-Level Reward Function Estimationcited by 25
  4. 2024Super(ficial)-alignment: Strong Models May Deceive Weak Models in Weak-to-Strong Generalizationcited by 23
  5. 2025Scaling Laws For Scalable Oversightcited by 23
  6. 2024High-dimensional Analysis of Knowledge Distillation: Weak-to-Strong Generalization and Scaling Lawscited by 22
  7. 2025Weak-for-Strong: Training Weak Meta-Agent to Harness Strong Executorscited by 22
  8. 2024Weak-to-Strong Generalization Through the Data-Centric Lenscited by 21
  9. 2024Provable Weak-to-Strong Generalization via Benign Overfittingcited by 16
  10. 2025Weak-to-Strong Generalization Even in Random Feature Networks, Provablycited by 15
  11. 2024Improving Weak-to-Strong Generalization with Reliability-Aware Alignmentcited by 14
  12. 2024Towards Acyclic Preference Evaluation of Language Models via Multiple Evaluatorscited by 13
36 more
  1. 2024The Road to Artificial SuperIntelligence: A Comprehensive Survey of Superalignmentcited by 13
  2. 2025Discrepancies are Virtue: Weak-to-Strong Generalization through Lens of Intrinsic Dimensioncited by 12
  3. 2025Representations Shape Weak-to-Strong Generalization: Theoretical Insights and Empirical Predictionscited by 10
  4. 2025Incentivizing Strong Reasoning from Weak Supervisioncited by 8
  5. 2026Weak Critics Make Strong Learners: On-Policy Critique Distillation for Scalable Oversightcited by 6
  6. 2025Prediction-Powered Semi-Supervised Learning with Online Power Tuningcited by 5
  7. 2024Weak-to-Strong Generalization beyond Accuracy: a Pilot Study in Safety, Toxicity, and Legal Reasoningcited by 4
  8. 2025EnsemW2S: Enhancing Weak-to-Strong Generalization with Large Language Model Ensemblescited by 4
  9. 2025Weak-to-Strong Generalization via Bregman Bias-Variance Decompositioncited by 4
  10. 2025Weak-to-Strong Generalization under Distribution Shiftscited by 4
  11. 2024Disentangling Latent Shifts of In-Context Learning with Weak Supervisioncited by 3
  12. 2025How to Mitigate Overfitting in Weak-to-strong Generalization?cited by 3
  13. 2025Generalizable Multispectral Land Cover Classification via Frequency-Aware Mixture of Low-Rank Token Expertscited by 3
  14. 2025Weak-to-Strong Generalization with Failure Trajectories: A Tree-based Approach to Elicit Optimal Policy in Strong Modelscited by 3
  15. 2025Context Parametrization with Compositional Adapterscited by 3
  16. 2025Mutual Learning for Hashing: Unlocking Strong Hash Functions from Weak Supervisioncited by 3
  17. 2025The Capabilities and Limitations of Weak-to-Strong Generalization: Generalization and Calibrationcited by 2
  18. 2025Beyond External Monitors: Enhancing Transparency of Large Language Models for Easier Monitoringcited by 2
  19. 2025Does Weak-to-strong Generalization Happen under Spurious Correlations?cited by 2
  20. 2026Weak to Strong: VLM-Based Pseudo-Labeling as a Weakly Supervised Training Strategy in Multimodal Video-based Hidden Emotion Understanding Taskscited by 2
  21. 2026Weak-to-Strong Generalization is Nearly Inevitable (in Linear Models)cited by 2
  22. 2026Eliciting Weak-to-Strong Generalization with On-Policy Reverse Distillationcited by 2
  23. 2024EnsemW2S: Enhancing Weak-to-Strong Generalization with Large Language Model Ensemblescited by 1
  24. 2025Limitations of refinement methods for weak to strong generalizationcited by 1
  25. 2026Self-Improvement as Coherence Optimization: A Theoretical Accountcited by 1
  26. 2026DC-W2S: Dual-Consensus Weak-to-Strong Training for Reliable Process Reward Modeling in Biological Reasoningcited by 1
  27. 2026On the Blessing of Pre-training in Weak-to-Strong Generalizationcited by 1
  28. 2026The Mechanism of Weak-to-Strong Generalization: Feature Elicitation from Latent Knowledgecited by 1
  29. 2025Denoising Mutual Knowledge Distillation in Bi-Directional Multiple Instance Learningcited by 0
  30. 2025Weak-to-Strong Generalization Enables Fully Automated De Novo Training of Multi-head Mask-RCNN Model for Segmenting Densely Overlapping Cell Nuclei in Multiplex Whole-slice Brain Imagescited by 0
  31. 2026COSMOS: Model-Agnostic Personalized Federated Learning with Clustered Server Models and Pseudo-Label-Only Communicationcited by 0
  32. 2026Gradient Transformer: Learning to Generate Updates for LLMscited by 0
  33. 2026Trust Functions: Near-Lossless Weak-to-Strong Generalization by Learning When to Trust the Weak Teachercited by 0
  34. 2026CritICL: Inference-Time Weak-to-Strong Generalization from Small Language Model Failure Modescited by 0
  35. 2026Safeguarding Mutual Correction in Source-Free Domain Adaptation via Cut Statisticscited by 0
  36. 2026CARD: Closing the Teacher-Student Loop for Adversarial Robustness Distillationcited by 0