Ruijia Zhang

PhD Student, Applied Mathematics & Statistics
Johns Hopkins University

I'm a PhD student in Applied Mathematics and Statistics at Johns Hopkins University, where I'm fortunate to be advised by Prof. Soufiane Hayou.
My research is driven by a broad interest in the science of AI: understanding how intelligence emerges, evolves, and scales, from individual neural networks to interacting multi-agent systems.

Selected Research

Three width and depth scaling regimes converging to the same trained network output

Neural network scaling · arXiv preprint

Width and Depth Limits Commute during Training in Linear Residual Networks

Shows that linear residual networks converge to the same limit during training, regardless of how width and depth jointly grow to infinity.

Ruijia Zhang, Soufiane Hayou

paper

Comparison of LoRA, PiSSA, MiLoRA, LoRA-RLPO, and LoRA-RLMO initialization strategies

LLM alignment · ICML 2026

Geometry-Preserving Orthonormal Initialization for Low-Rank Adaptation in RLVR

Joint work with researchers at Meta Superintelligence Labs and xAI

Explains why SVD-initialized LoRA variants can become unstable in RLVR and introduces orthonormal initializations that preserve pretrained geometry.

Ruijia Zhang, Jiacheng Zhu, Hanqing Zhu, Laixi Shi

paper

Journal Publication

Pessimistic Offline RL for Multi-Stage Decision-Making

Accepted at the Journal of the American Statistical Association 🏆 ASA Nonparametric Statistics Section Best Student Paper Award paper

Technical Reports on Scalable LLM Systems

On the Scaling of PEFT: Towards Million Personal Models of Trillion Parameters arXiv:2606.02437
MinT: Managed Infrastructure for Training and Serving Millions of LLMs arXiv:2605.13779

Full Publication List

2026 Width and Depth Limits Commute during Training in Linear Residual NetworksarXiv:2610.30729 · Submitted to ICLR 2027Ruijia Zhang, Soufiane Hayou
2026 Geometry-Preserving Orthonormal Initialization for Low-Rank Adaptation in Reinforcement LearningICML 2026Ruijia Zhang, Jiacheng Zhu, Hanqing Zhu, Laixi Shi
2026 POLAR: A Pessimistic Model-based Policy Learning Algorithm for Dynamic Treatment RegimesAccepted at the Journal of the American Statistical Association🏆 ASA Nonparametric Statistics Section Best Student Paper Award, JSM 2026Ruijia Zhang, Zhenglin Qi, Yue Wu, Xiangyu Zhang, Yanxun Xu
2026 Sparse Target Calibration for Domain Adaptation: A Conditional Distributionally Robust ViewSubmitted to ICLR 2027Ruijia Zhang, Dehong Liu, Yebin Wang, Mitsubishi Electric Research Laboratories
2025 Understanding Inverse Reinforcement Learning under Overparameterization: Non-Asymptotic Analysis and Global OptimalityAISTATS 2025Ruijia Zhang, Mingyi Hong, Alfredo Garcia
2025 Improved Rates of Differentially Private Nonconvex Strongly Concave Minimax Optimization via Gradient DifferencesAAAI 2025Ruijia Zhang, Mingxi Lei, Meng Ding, Zihang Xiang, Jinhui Xu, Di Wang
WIP Learning Optimal Robust Policies under Observational Data with Causal TransportIn preparation for Operations ResearchRuijia Zhang, Luhao Zhang, Michael Lingzhi Li

Industry Experience

Research Intern, Mitsubishi Electric Research Laboratories (MERL) Cambridge, MA · May 2026 — Aug 2026 · Mentor: Dr. Dehong Liu Large language models for anomaly detection and predictive maintenance.

Awards

> 🏆 ASA Nonparametric Statistics Section Best Student Paper Award Joint Statistical Meetings (JSM) 2026
> 🏆 ICCOPT 2025 Student Grant International Conference on Continuous Optimization

Teaching

> TA Time Series Analysis EN.553.639 · JHU · Spring & Fall 2026
> TA Data Science Methods for Large-Scale Graphs EN.553.744 · JHU · Spring 2025
> TA Investment Science EN.553.642 · JHU · Fall 2024 & 2025
> TA Ordinary Differential Equations MAT2002 · CUHK-Shenzhen · Spring 2024
> TA Optimization MAT3007 · CUHK-Shenzhen · Fall 2023

Education

2024 — now PhD, Applied Mathematics & Statistics
Johns Hopkins University
2020 — 2024 BSc, Mathematics & Applied Mathematics (First Class Honors)
The Chinese University of Hong Kong, Shenzhen
2023 Exchange, Mathematics
Technical University of Munich

last updated · 2026-10-09