LoopedTrans-formers2026Looped Transformers with Layer Normalization Provably Learn the Power MethodLyumin Wu, Chenyang Zhang, Yuan CaoarXiv preprint
Trans-formersTrained2026Transformers Trained via Gradient Descent Can Provably Learn a Class of Teacher ModelsChenyang Zhang, Qingyue Zhao, Quanquan Gu, Yuan CaoarXiv preprint
Trans-formersEfficiently2026Transformers Efficiently Perform In-Context Logistic Regression via Normalized Gradient DescentChenyang Zhang, Yuan CaoarXiv preprint
Trans-formerLearns2025Transformer Learns Optimal Variable Selection in Group-Sparse ClassificationChenyang Zhang, Xuran Meng, Yuan CaoarXiv preprint
TowardsUnder-standing2025Towards Understanding Generalization in DP-GD: A Case Study in Training Two-Layer CNNsZhongjie Shi, Puyu Wang, Chenyang Zhang, Yuan CaoarXiv preprint
GradientDescent2025Gradient Descent Robustly Learns the Intrinsic Dimension of Data in Training Convolutional Neural NetworksChenyang Zhang, Peifeng Gao, Difan Zou, Yuan CaoarXiv preprint
On theRobust-ness2025On the Robustness of Transformers against Context Hijacking for Linear ClassificationTianle Li, Chenyang Zhang, Xingwu Chen, Yuan Cao, Difan ZouarXiv preprint
TheImplicitBias2024The Implicit Bias of Adam on Separable DataChenyang Zhang, Difan Zou, Yuan CaoarXiv preprint