Qingyue Zhang  「张清越」

Hi there, thanks for visiting my website! I am a Ph.D. Candidate at SIGS of Tsinghua University, where I am fortunate to be advised by Prof. Shao-Lun Huang. Prior to that, I obtained my bachelor's degree (with honors) in computer science and technology at Xi'an Jiaotong University (2019-2023), and had wonderful time conducting research at the LUD lab, where I worked closely with Shangbin Feng, and was advised by Prof. Minnan Luo and Prof. Qinghua Zheng.

Currently, my research interests lie in applying information and statistical theory to analyze, interpret, and design machine learning algorithms. In particular, my current work focuses on transfer learning and parameter-efficient fine-tuning of large models (e.g., LoRA). I also have prior research experience in social network analysis, anomaly detection, and graph neural networks.

Email:  zhangqingyue2019 [at] gmail [dot] com  /  zhangqy23 [at] mails [dot] tsinghua [dot] edu [dot] cn

CV  /  Google Scholar  /  Twitter  /  Github

• National Scholarship (1.5% in Tsinghua University), Ministry of Education of China, 2025
• Excellent Graduate Student, Xi'an Jiaotong University, 2023
• Dean's List, Xi'an Jiaotong University, 2020, 2021, 2022
• “ShuiYou” Scholarship for Outstanding Students, First Prize (Rank 1/40 in Experimental Class), 2021
• Top Project Runner Up, NUS SoC Summer Workshop, 2021

Currently, my research interests lie in applying information and statistical theory to analyze, interpret, and design machine learning algorithms. In particular, my current work focuses on transfer learning and parameter-efficient fine-tuning of large models (e.g., LoRA). I also have prior research experience in social network analysis, anomaly detection, and graph neural networks.

We propose LoRA-DA, a theoretically grounded and data-aware initialization method for LoRA. Based on asymptotic analysis, our framework decomposes fine-tuning error into a variance term (from Fisher information) and a bias term (from Fisher-gradient), yielding the optimal LoRA initialization. Experiments across multiple benchmarks show that LoRA-DA consistently improves accuracy and convergence speed over previous LoRA initialization methods.

We propose a theoretical framework that answers the question: what is the optimal quantity of source samples needed from each source task to jointly train the target model? Specifically, we introduce a generalization error measure based on K-L divergence, and minimize it based on asymptotic analysis to determine the optimal transfer quantity for each source task. Additionally, we develop an architecture-agnostic and data-efficient algorithm OTQMS to implement our theoretical results for target model training in multi-source transfer learning.

We studied fixed-length hypothesis testing where the number of training samples is set in a fixed ratio to the number of testing samples. We derived the optimal error exponent and its corresponding decision rule, and revealed a phase transition determined by this ratio: when the training portion exceeds one half, the error exponent reaches the classical Chernoff information; when it is below one half, the exponent decreases linearly. This establishes a principled benchmark for machine learning, clarifying how the balance between training and testing samples influences achievable performance.

We adopt contrastive learning and knowledge relational attention network to alleviate the widespread sparsity problem in knowledge graphs.

We present TwiBot-22, the largest graph-based Twitter bot detection benchmark to date, which provides diversified entities and relations in Twittersphere and has considerably better annotation quality.

• Reviewer: ICLR 2026, NeurIPS 2025, ICML2025, AAAI 2025, ICLR 2025, TNNLS-2024, ICML 2024, ICLR 2024, NeurIPS 2023
• Volunteer: ITW2024
• Teaching Assistant: Large Deviations and High-Dimensional Statistics, Tsinghua University, 2025 Spring

Template courtesy: Jon Barron.