Hi! I'm Ziyang Xiong

I am currently a master student in EECS at UC Berkeley. For my undergraduate studies, I recently graduated from the University of Michigan with a major in Data Science and completed a dual degree in Electrical and Computer Engineering at Shanghai Jiao Tong University through a joint program. I used to work in Statistics & Optimization for Trustworthy AI (SOTA) lab group to explore more about LLM and Reinforcement Learning. I am also working in Foreseer Group to explore advanced research in LLMs for scientific discovery, focusing on innovative AI methods and their applications in data analysis and knowledge generation. I aslo work with Prof.Z.mao to explore the potential of AI security.

profile
Research Interest: My background and research experience mainly focus on large language models, machine learning, and reinforcement learning. I have extensive hands-on experience in LLM fine-tuning, reinforcement learning, prompt engineering, and multi-modal model applications.

Publications

MapExplorer: New Content Generation from Low-Dimensional Visualizations KDD 2025 Accepted
Xingjian Zhang, Ziyang Xiong, Shixuan Liu, Yutong Xie, Tolga Ergen, Dongsub Shim, Hua Xu, Honglak Lee, Qiaozhu Mei

MASSW: A New Dataset and Benchmark Tasks for AI-Assisted Scientific Workflows NAACL 2025 Accepted
Xingjian Zhang, Yutong Xie, Ziyang Xiong, etc.

Safeguard is a Double-edged Sword: Denial-of-service Attack on Large Language Models ACM CCS-LAMPS Accepted
Qingzhao Zhang, Ziyang Xiong, Z. Morley Mao

Making Small Language Models Efficient Reasoners: Intervention, Supervision, Reinforcement ICML 2025 Workshop LCFM Accepted
Xuechen Zhang, Zijian Huang, Chenshun Ni, Ziyang Xiong, Jiasi Chen, Samet Oymak

Improve LLM Faithfulness with KG

  • Intergrating LLM with KG allows the language model to answer unprecedented questions by leveraging knowledge from external sources.
  • Eliminates the need for model fine-tuning and imposes no constraints on the dataset.
    • Skills: Python, Large Language Model, Knowledge Graph

    Lipstick Search Engine

  • Crawl YouTube for lipstick-related videos and gather information. Train a model to rank lipsticks based on multiple criteria.
  • Provide users with filters such as price, color, and benefits to help them find and recommend the most suitable lipstick.
    • Skills: Python, Information Retrieval, NLP

    Depression Prediction

  • Collect and pre-process data including population depression, various biochemical indicators (various proteins, hormones), socioeconomic factors (education, income), etc.
  • Use various methods, including oversampling, lasso regularization, random forest,SVM,Logistic Regression,KNN and so on to construct a biochemical and socioeconomic prediction model of depression.
    • Skills: Machine Learning, Model Building, python and R

    HealthGoal App

  • Seamlessly integrate with Firebase for database management and tailored health plan recommendations with openai-api.
  • Craft an intuitive health app encompassing features like journaling, goal tracking, and interactive chat.
    • Skills: Dart, Flutter, Firebase

    CV & Experience

    I am currently a master student in EECS at UC Berkeley.

  • During the Summer of 2025, I worked as Model Algorithm Engineer at ByteDance, where I led development of Search Ads title rewriting models using advanced fine-tuning techniques (SFT, DPO, KTO), and keyword extraction processes, achieving production deployment with measurable business impact: CTR +0.467%, advv +1.963%, send +0.477%. Built automated bad case evaluation models achieving 86.7% precision and 89% recall using Qwen3-8B-CoT, establishing robust quality control frameworks, and significantly improving model assessment efficiency. Developed end-to-end video-to-title multimodal models using Qwen2.5-VL-7B, resolving information loss issues in multi-stage processing and enhancing selling point concentration through innovative two-stage Label to Title generation approach.
  • During the Spring of 2025, I worked as a grader for course namedComputer Vision.
  • During the Fall of 2024, I worked as a grader for course named Foundations of LLMs.
  • During the summar of 2024, I worked as a research assistant in the SOTA Lab, where I integrated in-context learning with fine-tuning techniques to improve model performance and conducted research on optimizing the Transformer’s attention mechanism.
  • During the summer of 2023, I worked as a teaching assistant of the course ENGR 100: INTRODUCTION TO ENGINEERING .
  • During the winter of 2022, I worked as an programmer in China Telecom Hubei Branch. During this time, I designed a cloud-based fusion platform for the collection and integration of user device fault data. This platform can automatically analyzed, consolidated, and categorized diverse fault information, forwarding relevant data to respective departments and generating daily reports.