Weak-to-Strong Generalization

Ask about this topic

Overview

Weak-to-strong generalization asks when a student can learn from an imperfect teacher and outperform it, which could make oversight less dependent on costly expert labels. Methods include filtering unreliable labels with trust scores or consensus, distilling teacher feedback on the student’s own outputs, and using critiques or representation regularization. Theory studies linear and random-feature models, while evaluations span reasoning benchmarks such as AIME and ARC-AGI-2, preference shifts in reward models, and downstream decisions. Results depend on factors such as pretraining, teacher bias, student capacity, and data; a central tension is how to gain from weak supervision without inheriting its blind spots or losing out-of-distribution performance.

Latest papers

CardsList

22 papers2 papers in the last four weeks, down 50% on the four weeks before. 0.0% of all new papers.

  1. Boosting and the Expressive Power of Simple Weak Learners via the γγ-VC Dimension

    Oct 7, 2026Arthur da Cunha, Kasper Green Larsen, Liang-Yu ZouStatistical Learning TheoryWeak-to-Strong Generalization

    Aarhus University

  2. Quadratic Weak-to-Strong Generalization in Random Feature Networks via Random Matrix Theory

    Oct 6, 2026Deborah Oliveira, Elliot PaquetteWeak-to-Strong GeneralizationRandom Feature Methods

    McGill University Mila – Quebec AI Institute

  3. An Active-Bottleneck Mechanism for Weak-to-Strong Generalization

    Sep 27, 2026Mohammad Zeinalpour, Amir NajafiWeak-to-Strong GeneralizationRandom Feature Methods

    Department of Computer Engineering, Sharif University of Technology

  4. Allspark: Weak to Strong Transfer via Alternating Chain of Thought

    Sep 26, 2026Kaizhao Liang, Junxiong Wang, Chen Liang +2Weak-to-Strong GeneralizationTransfer Learning

    UT Austin · Together AI · Microsoft

  5. Scaling Properties of Same-Family On-Policy Distillation

    Sep 26, 2026Yuntai Bao, Qinfeng Li, Guoqing Jiang +5On-Policy DistillationLanguage Model Distillation

    Zhejiang University · Kuaishou Technology

  6. Eliciting Weak-to-Strong Generalization with On-Policy Reverse Distillation

    Sep 8, 2026Youngrok Park, Sangmin Bae, Hojung Jung +6Policy GradientWeak-to-Strong Generalization

    KAIST AI · Microsoft · University of Toronto +3

  7. CritICL: Inference-Time Weak-to-Strong Generalization from Small Language Model Failure Modes

    Aug 27, 2026Yufan Wu, Yinghui He, Zhengyi Hu +4Efficient Language Model ReasoningInference-Time Guidance

    The Ohio State University · Princeton University

  8. Towards a theory of inference-time alignment with unknown rewards

    Aug 15, 2026Steve Hanneke, Hongao Wang, Mingyue XuInference-Time Language Model AlignmentStatistical Learning Theory

    Purdue University

  9. Defensive Boosting for Online Probabilistic Forecasting

    Aug 13, 2026Georgy Noarov, Aaron RothProbabilistic ForecastingWeak-to-Strong Generalization

  10. Weak-to-Strong Learning in Decision Making

    Jul 20, 2026Jingwei Ji, Renyuan XuStochastic OptimizationDecision-Focused Learning

    Management Science and Engineering, Stanford University

  11. Weak-to-Strong Generalization via Direct On-Policy Distillation

    Jul 6, 2026Shiyuan Feng, Huan-ang Gao, Haohan Chi +7RL for Language Model ReasoningWeak-to-Strong Generalization

    1SIA-Lab of Tsinghua AIR and ByteDance Seed · Institute for AI Industry Research (AIR), Tsinghua University · 4Peking University +1

  12. What Makes a Strong Model? A Unified Spectral Analysis of Knowledge Transfer over High-dimensional Linear Regression

    May 31, 2026Wendao Wu, Fangqing Zhang, Haihan Zhang +1High-Dimensional Linear RegressionWeak-to-Strong Generalization

    State Key Lab of General AI, School of Intelligence Science and Technology, Peking University, China · School of Mathematical Sciences, Peking University, China

  13. Trust Functions: Near-Lossless Weak-to-Strong Generalization by Learning When to Trust the Weak Teacher

    May 31, 2026Arda Uzunoglu, Alvin Zhang, Daniel KhashabiData SelectionTraining Data Selection

    Department of Computer Science, Johns Hopkins University, Maryland, United States.

  14. Weak Critics Make Strong Learners: On-Policy Critique Distillation for Scalable Oversight

    May 29, 2026Can Jin, Jiakang Li, Rui Wu +2Scalable OversightWeak-to-Strong Generalization

    Department of Computer Science, Rutgers University, Piscataway, United States.

  15. When In-Distribution Gains Fail: Evaluating Weak-to-Strong Reward Models under Preference Shift

    May 25, 2026Khoi Le, Tri Cao, Phong Nguyen +5Pairwise Preference LearningReward Modeling

    National University of Singapore · VinUniversity · Nanyang Technological University

  16. The Mechanism of Weak-to-Strong Generalization: Feature Elicitation from Latent Knowledge

    May 13, 2026Ryoya Awano, Taiji SuzukiNeural Network GeneralizationRepresentation Learning

    University of Tokyo · Center for Advanced Intelligence Project, RIKEN

  17. Weak-to-Strong Generalization is Nearly Inevitable (in Linear Models)

    May 7, 2026Scott Geng, Dutch Hansen, Jerry LiLogistic RegressionWeak-to-Strong Generalization

    University of Washington, Seattle.

  18. On the Blessing of Pre-training in Weak-to-Strong Generalization

    May 7, 2026Wei Yao, Wang Zhaoyang, Gengze Xu +5Language Model PretrainingNeural Network Generalization

    Renmin University of China · Shanghai Jiao Tong University · Tongji University +1

  19. Evaluating Risks in Weak-to-Strong Alignment: A Bias-Variance Perspective

    Apr 28, 2026Hamid Osooli, Kareema Batool, Rick Gentry +3LLM Safety AlignmentWeak-to-Strong Generalization

    University of Illinois Urbana-Champaign · InstaDeep · Microsoft +1

  20. Resource-constrained Amazons chess decision framework integrating large language models and graph attention

    Mar 11, 2026Tianhao Qian, Zhuoxuan Li, Jinde Cao +2Game-Playing AgentsWeak-to-Strong Generalization

    School of Mathematics, Southeast University, Nanjing, 210096, China · Systems Research Institute of the Polish Academy of Sciences, Warsaw 01-447, Poland · School of Cyber Science & Engineering, Southeast University, Nanjing, 210096, China +2

  21. DC-W2S: Dual-Consensus Weak-to-Strong Training for Reliable Process Reward Modeling in Biological Reasoning

    Mar 9, 2026Chi-Min Chan, Ehsan Hajiramezanali, Xiner Li +6Process Reward ModelsWeak-to-Strong Generalization

    The Hong Kong University of Science and Technology, Hong Kong · Genentech, Inc., South San Francisco, CA, USA.

  22. Improved Scaling Laws via Weak-to-Strong Generalization in Random Feature Ridge Regression

    Mar 5, 2026Diyuan Wu, Lehan Chen, Theodor Misiakiewicz +1Weak-to-Strong GeneralizationScaling Laws

    Institute of Science and Technology Austria (ISTA) · Department of Applied and Computational Mathematics, Yale University · Department of Statistics and Data Science, Yale University