I’m a Ph.D. student in Software Engineering at State Key Lab of CAD&CG, Zhejiang University, under the supervision of Prof. Wei Chen.
I’m currently interested in language models and agents. I’m currently an intern with the Qwen Foundation Model Team at Alibaba.
📖 Education
- 2023.09 - 2028.06 (Expected)
Ph.D. student in Software Engineering in Zhejiang University
. Supervisor: Wei Chen. - 2019.09 - 2023.06
B.E. in Software Engineering, Xi’an Jiaotong University
.
📝 Publications
PDF |
Zhaorui Yang*, Bo Pan*, Han Wang*, Yiyao Wang, Xingyu Liu, Luoxuan Weng, Yingchaojie Feng, Haozhe Feng, Minfeng Zhu, Bo Zhang, Wei Chen
- DeepResearch systems typically generate text-only reports, despite visualizations being essential for real-world analysis. We introduced an agentic framework that generates text-chart interleaved reports from scratch, together with a structured chart representation that enables LLMs to produce diverse, high-quality visualizations.
Self-Distillation Bridges Distribution Gap in Language Model Fine-Tuning
Zhaorui Yang, Tianyu Pang, Haozhe Feng, Han Wang, Wei Chen, Minfeng Zhu, Qian Liu
- Fine-tuning LLMs often improves downstream performance at the cost of general capabilities. We attributed this trade-off to the distribution gap between task data and the pretrained model, and proposed Self-Distillation Fine-Tuning (SDFT) to bridge this gap by aligning fine-tuning data with the model’s original distribution.
ProSPy: A Profiling-Driven SQL-Python Agentic Framework for Enterprise Text-to-SQL
Zhaorui Yang, Huawei Zheng, Sen Yang, Yuhui Zhang, Haoxuan Li, Zhizhen Yu, Xuan Yi, Chen Hou, Defeng Xie, Chao Hu, Minfeng Zhu, Dazhen Deng, Haozhe Feng, Danqing Huang, Peng Chen, Wei Chen
- Real-world Text-to-SQL faces heterogeneous schemas, incomplete metadata, SQL dialects, and complex analytical queries. We proposed a profiling-driven SQL–Python agentic framework that combines efficient SQL-based retrieval with flexible Python analysis. ProSPy ranked No. 2 on Spider 2.0-Lite at submission; its improved successor, Tianqiong Data Agent, ranked No. 1 as reported in my September 2026 résumé.
CoSDA: Continual Source-Free Domain Adaptation
PDF |
Haozhe Feng*, Zhaorui Yang*, Hesun Chen*, Tianyu Pang, Chao Du, Minfeng Zhu, Wei Chen, Shuicheng Yan
- Continual source-free domain adaptation suffers from catastrophic forgetting when adapting to a sequence of target domains. We identified a trade-off between adaptation gain and forgetting loss, and proposed CoSDA to balance adaptation and knowledge retention, outperforming prior methods under continuous domain shifts.
💻 Internships
- 2026.04 - Present: Qwen Foundation Model Team, ATH, Alibaba
Contributed to the development of the search teacher model for Qwen 3.8 Max, focusing on evaluation and trajectory analysis pipelines, long-horizon training data construction, and large-scale SFT and RL experiments on hundreds of H200 GPUs. Currently working on reinforcement learning for Qwen’s cyber model, targeting long-horizon CTF and exploitation tasks studied by frontier AI labs. - 2025.07 - 2026.03: Data Computing Platform Department, TEG, Tencent
Led ProSPy from problem formulation and system design to experimentation. It combines data profiling, SQL retrieval, and Python analysis for enterprise Text-to-SQL, reached No. 2 on Spider 2.0-Lite at submission. Its successor, Tianqiong Data Agent, ranks No. 1 now.
🎖 Honors and Awards
- 2022.12 China National Scholarship (Undergraduate).
- 2021.12 China National Scholarship (Undergraduate).
2026-09-09