Publications

In reversed chronological order | * denotes equal contribution

2025

  1. arXiv
    Theoretical Tensions in RLHF: Reconciling Empirical Success with Inconsistencies in Social Choice Theory
    Jiancong Xiao, Zhekun Shi, Kaizhao Liu, Qi Long, and Weijie J. Su
    submitted, under review , 2025
    arXiv
  2. arXiv
    Fundamental Limits of Game-Theoretic LLM Alignment: Smith Consistency and Preference Matching
    Zhekun Shi, Kaizhao Liu, Qi Long, Weijie J. Su, and Jiancong Xiao
    submitted, under review , 2025
    arXiv
  3. ICML 2025
    Restoring Calibration for Aligned Large Language Models: A Calibration-Aware Fine-Tuning Approach
    Jiancong Xiao*, Bojian Hou*, Zhanliang Wang*, Ruochen Jin, Qi Long, Weijie J. Su, and Li Shen
    In Proceedings of the 42nd International Conference on Machine Learning, 2025
    arXiv
  4. arXiv
    Statistical Impossibility and Possibility of Aligning LLMs with Human Preferences: From Condorcet Paradox to Nash Equilibrium
    Kaizhao Liu, Qi Long, Zhekun Shi, Weijie J. Su, and Jiancong Xiao
    submitted, under review , 2025
    arXiv Code
  5. ICLR 2025
    Magnetic Preference Optimization: Achieving Last-iterate Convergence for Language Models Alignment
    Mingzhi Wang, Chengdong Ma, Qizhi Chen, Linjian Meng, Yang Han, Jiancong Xiao, Zhaowei Zhang, Jing Huo, Weijie J Su, and Yaodong Yang
    In The Thirteenth International Conference on Learning Representations, 2025
    arXiv
  6. ICLR 2025
    Preserving Diversity in Supervised Fine-Tuning of Large Language Models
    Ziniu Li, Congliang Chen, Tian Xu, Zeyu Qin, Jiancong Xiao, Zhi-Quan Luo, and Ruoyu Sun
    In The Thirteenth International Conference on Learning Representations, 2025
    arXiv
  7. ICLR 2025
    Fine-Tuning Attention Modules Only: Enhancing Weight Disentanglement in Task Arithmetic
    Ruochen Jin*, Bojian Hou*, Jiancong Xiao*, Weijie J. Su, and Li Shen
    In The Thirteenth International Conference on Learning Representations, 2025
    arXiv Code

2024

  1. arXiv
    On the Algorithmic Bias of Aligning Large Language Models with RLHF: Preference Collapse and Matching Regularization
    Jiancong Xiao, Ziniu Li, Xingyu Xie, Emily Getzen, Cong Fang, Qi Long, and Weijie J. Su
    Minor Revision at Journal of the American Statistical Association (JASA) , 2024
    arXiv Code
  2. NeurIPS 2024 FITML Best Paper Runner-up
    Entropic Distribution Matching in Supervised Fine-tuning of LLMs: Less Overfitting and Better Diversity
    Ziniu Li, Congliang Chen, Tian Xu, Zeyu Qin, Jiancong Xiao, Ruoyu Sun, and Zhi-Quan Luo
    In NeurIPS 2024, Fine-Tuning in Modern Machine Learning: Principles and Scalability, 2024
    HTML
  3. COLT 2024
    Bridging the Gap: Rademacher Complexity in Robust and Standard Generalization
    Jiancong Xiao, Ruoyu Sun, Qi Long, and Weijie J. Su
    In Proceedings of the Thirty Seventh Conference on Learning Theory, 2024
    arXiv HTML
  4. ICML 2024
    Uniformly Stable Algorithms for Adversarial Training and Beyond
    Jiancong Xiao*, Jiawei Zhang*, Zhi-Quan Luo, and Asuman E. Ozdaglar
    In Proceedings of the 41st International Conference on Machine Learning, 2024
    arXiv Code

2023

  1. NeurIPS 2023
    PAC-Bayesian Spectrally-Normalized Bounds for Adversarially Robust Generalization
    Jiancong Xiao, Ruoyu Sun, and Zhi-Quan Luo
    In Advances in Neural Information Processing Systems, 2023
    arXiv HTML
  2. ICML 2023 AdvML
    Improving Adversarial Training for Multiple Perturbations through the Lens of Uniform Stability
    Jiancong Xiao, Zeyu Qin, Yanbo Fan, Baoyuan Wu, Jue Wang, and Zhi-Quan Luo
    In ICML 2023, New Frontiers in Adversarial Machine Learning, 2023
    HTML Code
  3. ICML 2023 AdvML
    PAC-Bayesian Adversarially Robust Generalization Bounds for Deep Neural Networks
    Jiancong Xiao, Ruoyu Sun, and Zhi-Quan Luo
    In ICML 2023, New Frontiers in Adversarial Machine Learning, 2023
    HTML PDF
  4. Ph.D. Thesis
    Understanding Adversarially Robust Generalization: A Learning Theory Perspective
    Jiancong Xiao
    The Chinese University of Hong Kong, Shenzhen, 2023
    PDF

2022

  1. NeurIPS 2022 MLS
    Smoothed-SGDmax: A Stability-Inspired Algorithm to Improve Adversarial Generalization
    Jiancong Xiao*, Jiawei Zhang*, Zhi-Quan Luo, and Asuman E. Ozdaglar
    In NeurIPS 2022, ML Safety, 2022
    HTML
  2. NeurIPS 2022 Spotlight
    Stability Analysis and Generalization Bounds of Adversarial Training
    Jiancong Xiao, Yanbo Fan, Ruoyu Sun, Jue Wang, and Zhi-Quan Luo
    In Advances in Neural Information Processing Systems, 2022
    arXiv HTML Code
  3. arXiv
    Adversrial Rademacher Complexity of Deep Neural Networks
    Jiancong Xiao, Yanbo Fan, Ruoyu Sun, and Zhi-Quan Luo
    Revision at Journal of Machine Learning Research (JMLR) , 2022
    arXiv PDF Code
  4. Pattern Recognition
    Understanding Adversarial Robustness Against On-manifold Adversarial Examples
    Jiancong Xiao, Liusha Yang, Yanbo Fan, Jue Wang, and Zhi-Quan Luo
    Pattern Recognition, 2022
    arXiv HTML