Yaxuan Wang
I am a second-year PhD student in Computer Science and Engineering at the University of California, Santa Cruz, advised by
Yang Liu. I work on Trustworthy Machine Learning (Fairness and Robustness), and Large Language Model Alignment.
I have been working as a Research Scientist Intern at Accenture Data & AI Research Center since June 2024.
I have a B.Sc. from Sichuan University majored in Computer Science and Technology (Top-notch Program, a class of 15 elite students selected from 400+).
My previous advisor is professor
Qijun Zhao from the Department of Computer Science.
If you are interested in collaborating with me, please feel free to contact me via Email or Linkedin.
More about my research interests:
LLM Alignment, Unlearning, and Preference Optimization
Trustworthy Machine Learning, especially Fairness, Interpretability and Robustness
Reinforcement Learning and Limitation Learning (Inverse Reinforcement Learning)
Representation Learning & Learning from Dynamic and Noisy Data
I am interested in building intelligent agents or models that can learn
from various data sources (e.g., images, audio, and text) and automate decision-making processes
in uncertain environments to solve multiple real-world tasks.
My research lies in the broad area of machine learning and data science and I am open to specific issues. I love the process of giving machines the intelligence to address social and healthcare problems.
Publications
Noise-Resilient Point-wise Anomaly Detection in Time Series Using Weak Segment Labels
Yaxuan Wang, Cheng Hao, Jing Xiong, Qingong Wen, Han Jia, Ruixuan Song, Liyuan Zhang, Zhaowei Zhu, and Yang Liu
KDD 2025: The 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining
LLM Unlearning via Loss Adjustment with Only Forget Data - link
Yaxuan Wang, Jiaheng Wei, Chris Yuhao Liu, Jinlong Pang, Quan Liu, Ankit Parag Shah, Yujia Bao, Yang Liu, Wei Wei
Large Language Model Unlearning via Embedding-Corrupted Prompts
Chris Yuhao Liu, Yaxuan Wang, Jeffrey Flanigan, Yang Liu
NeurIPS 2024: The 38th Annual Conference on Neural Information Processing Systems
Improving Data Efficiency via Curating LLM-Driven Rating Systems - link
Jinlong Pang, Jiaheng Wei, Ankit Parag Shah, Zhaowei Zhu, Yaxuan Wang, Chen Qian, Yang Liu, Yujia Bao, Wei Wei
A Complete Reinforcement-Learning-Based Framework for Urban-Safety Perception - link
Yaxuan Wang, Zhixin Zeng, Qiushan Li, and Yingrui Deng
ISPRS International Journal of Geo-Information 11, Issue 9
Evaluating the Perceived Safety of Urban City via Maximum Entropy Deep Inverse Reinforcement Learning
- link
Yaxuan Wang, Zhixin Zeng, and Qijun Zhao
ACML 2022:The 14th Asian Conference on Machine Learning, Dec 2022, Hyderabad, India
Research
LLM Unlearning using Loss Adjustment based Method
With Prof Yang Liu (University of California, Santa Cruz) and Accenture Data & AI Center
Proposed a new loss objective FLAT to guide the unlearning process without requiring retain data or reference models. The idea is to use f-divergence to maximize the distance between the preferred answer and bad answer.
Responsible for method design, partial theoretical proof and all experimental conduction.
Paper in submission and positively reviewed.
LLM Unlearning via Embedding-Corrupted Prompts
With Prof Yang Liu (University of California, Santa Cruz)
Proprsed a novel perspective on unlearning for LLMs to address both the challenges of knowledge entanglement and unlearning efficiency.
Responsible for part experimental conduction.
Paper has been accepted to NeurIPS 2024.
LLM Generated Misinformation Detection
With Prof Yang Liu (University of California, Santa Cruz)
Designed and trained a classifier to detect misinformation (including Hallucination) within LLM-generated text using simulated datasets.
Responsible for model architecture design, data simulation, and experimental validation.
Currently working on improving the model's generalization to unseen data.
Time Series Anomaly Detection Using Partial Instance-level Labels
With Prof Yang Liu (University of California, Santa Cruz) and Staff Engineer Qingsong Wen (DAMO Academy)
Proposed a novel noise-tolerant anomaly detection approach that efficiently identifies anomaly points using provided partial instance-level labels (More details in the future paper)
roposed method achieved satisfactory results in all selected time series datasets compared to the previous model.
Responsible for idea formation, model design, experimental conduction, and paper writing.
Paper has been accepted to KDD 2025.
Research on Urban Safety Perception using Reinforcement Learning and Inverse Reinforcement Learning
With Prof Qijun Zhao (Sichuan University)
We aim to leverage reinforcement learning-based decision framework to obtain human-like safety perceptions and use inverse reinforcement learning (IRL) to
recover the reward function that can explain the evaluation pattern and help experts quantitatively analyze perceptual features.
Experimental results using our crowdsourced dataset showed satisfactory prediction performance (at least 3% improvement in F1 score) and excellent visual interpretability. It also showed that IRL has promising prospects in this field.
I was one of the main executors of this project, and we completed the writing and publication of two papers (one as first author, the other as a co-author).
Autonomous Robot Shooting and Movement in Specific Maps
With Prof Jiancheng Lv Prof Qijun Zhao (Sichuan University)
Based on visual perception and cost map, we used behaviour tree to realize intelligent decision-making of
the automatic robot, which can move and shoot on a specific map.
Responsible for the construction of the robot decision-making framework. Also as
the project leader to control the research progress and the funding application.
Our project won the ICRA RMUA International Third Prize.
Projects
Curriculum Design
| Course Name |
Project Name |
| Java Programming |
Airplane War Game |
| Representative Learning |
Occluded face image classification based on SRC algorithm and voting mechanism |
| Introduction to Deep Learning |
Non-framework neural network solves very similar problems of SVHN |
| Software Engineering |
Cross-platform talent management system |
| Digital Image Processing |
Portrait Cutout System |
| Introduction to Pattern Recoginiton |
Classification problem based on MNIST and traditional machine learning algorithms |
| Robotics Programming with ROS |
Robot following system based on ROS |
Previous Projects
2021.10-2022.06 Application research of artificial intelligence technology based on CBCT to automatically
determine alveolar bone density
2020.11-2021.10 Portable vision integrated real-time detection and tracking system for rare animals -
report
2020.11-2021.10 Commodity evaluation system based on natural language processing -
report |
code
2020.03-2020.06 "You Must Ask" WeChat Mini Program -
code
2019.11-2020.10 Evaluation system for graduation requirements under the background of engineering certification
Awards
2024, Anti-Racist Research Fellowship, University of California, Santa Cruz
2023, Dean's Fellowship, University of California, Santa Cruz
2023, Awarded the 2023 'Computer Science Polaris Star' at Sichuan University, an honor voted by peers and faculty as the only student in the department.
2023, Scholarship in Honor of Chinese Contemporary Scientists (top 0.03% Students University-wide)
2022, The RoboMaster 2022 University AI Challenge - International Third Prize
2021, National Scholarship(Top 0.2% Nationwide) & CK Power Scholarship & School-level first-class scholarship
2021, school-level first-class scholarship, Sichuan University (top 2% Students University-wide)
2021, The 14th China University Student Computer Design Competition - National First Prize
2021,
The RoboMaster University Championship - South Division First Prize
2021, The RoboMaster University League
- Provincial First Prize
2020, Outstanding Student of Sichuan University
2020, Outstanding Cadre of Sichuan University Library Volunteer Team
2020, The Eighth Sichuan University Student Engineering Training Comprehensive Ability Competition
- First Prize at School Level
2019, Second Prize of College-level English Speech Contest of Electrical Engineering
Activities
Worked part-time at Xuezhi and DJI.
Actively participated in social welfare activities, such as library volunteers, correcting essays
for children and participating in elementary school technology festival to tutor students.
Member of the women's basketball team of the School of Computer Science.
Served as vice president of 3D Printing Association.