University of California, Santa Cruz
I am a third-year Ph.D. Candidate in Computer Science and Engineering at the University of California, Santa Cruz, advised by Yang Liu. My research studies how knowledge and behavior are learned, represented, and modified in large language models, and how these insights can help models and agents improve through feedback and experience. I am currently especially interested in LLM post-training, reinforcement learning, agentic RL, model behavior (alignment and memorization), credit assignment, and self-improving agents. Since June 2026, I have been a Research Intern at Zoom, working on LLM agents and agentic RL. Previously, I worked as a Research Scientist Intern at Accenture Data & AI Research Center, focusing on high-impact research in the Generative AI space.
A central thread of my work is understanding, controlling, and improving LLM behavior: using unlearning to study memorization and knowledge storage, developing methods for selective behavioral control, and building agents that can learn from mistakes, interaction, and long-horizon feedback. My long-term goal is to develop reliable AI systems that can understand their own behavior and improve over time.
Warning
Problem: The current name of your GitHub Pages repository ("Solution: Please consider renaming the repository to "
http://".
However, if the current repository name is intended, you can ignore this message by removing "{% include widgets/debug_repo_name.html %}" in index.html.
Action required
Problem: The current root path of this site is "baseurl ("_config.yml.
Solution: Please set the
baseurl in _config.yml to "News
Research Direction
My research asks how LLMs learn and store knowledge, how learned knowledge and behavior can be deliberately controlled, and how models and agents can improve from feedback, failures, and interaction.
My current work studies agentic reinforcement learning for multi-turn search agents, including GRPO, process-level rewards, online RL, and credit assignment across long-horizon reasoning and tool-use trajectories. Earlier work on unlearning motivates this direction by asking what models memorize, how knowledge is retrieved or suppressed, and how interventions change observable behavior.
Unlearning provides a controlled intervention for studying knowledge acquisition, representation, retrieval, forgetting, and leakage.
Training-time and inference-time interventions test when stored knowledge can be selectively modified, suppressed, or expressed.
Current work studies how post-training and reinforcement learning objectives shape long-horizon agent behavior and enable learning from experience.
Selected Publications (view all )
PromptBridge: Cross-Model Prompt Transfer for Large Language Models Behavior Transfer
Yaxuan Wang, Quan Liu, Zhenting Wang, Zichao Li, Wei Wei, Yang Liu, Yujia Bao
The Conference on Language Modeling (COLM) 2026
Studies whether optimized behavior can be preserved and transferred across heterogeneous LLMs as model capability and cost constraints change.
PromptBridge: Cross-Model Prompt Transfer for Large Language Models Behavior Transfer
Yaxuan Wang, Quan Liu, Zhenting Wang, Zichao Li, Wei Wei, Yang Liu, Yujia Bao
The Conference on Language Modeling (COLM) 2026
Studies whether optimized behavior can be preserved and transferred across heterogeneous LLMs as model capability and cost constraints change.
Observations and Remedies for Large Language Model Bias in Self-Consuming Performative Loop Model Behavior
Yaxuan Wang, Zhongteng Cai, Yujia Bao, Xueru Zhang, Yang Liu
The 64th Annual Meeting of the Association for Computational Linguistics (ACL) - Main Conference 2026
Studies LLM bias in self-consuming performative loops and remedies for improving model behavior under repeated feedback.
Observations and Remedies for Large Language Model Bias in Self-Consuming Performative Loop Model Behavior
Yaxuan Wang, Zhongteng Cai, Yujia Bao, Xueru Zhang, Yang Liu
The 64th Annual Meeting of the Association for Computational Linguistics (ACL) - Main Conference 2026
Studies LLM bias in self-consuming performative loops and remedies for improving model behavior under repeated feedback.
DRAGON: Guard LLM Unlearning in Context via Negative Detection and Reasoning Behavior Control
Yaxuan Wang, Chris Yuhao Liu, Quan Liu, Jinlong Pang, Wei Wei, Yujia Bao, Yang Liu
The Fourteenth International Conference on Learning Representations (ICLR) 2026
Studies inference-time detection and control of undesired knowledge expression, separating what a model stores from what it retrieves or says in context.
DRAGON: Guard LLM Unlearning in Context via Negative Detection and Reasoning Behavior Control
Yaxuan Wang, Chris Yuhao Liu, Quan Liu, Jinlong Pang, Wei Wei, Yujia Bao, Yang Liu
The Fourteenth International Conference on Learning Representations (ICLR) 2026
Studies inference-time detection and control of undesired knowledge expression, separating what a model stores from what it retrieves or says in context.
LLM Unlearning via Loss Adjustment with Only Forget Data Post-training
Yaxuan Wang, Jiaheng Wei, Chris Yuhao Liu, Jinlong Pang, Quan Liu, Ankit Shah, Yujia Bao, Yang Liu, Wei Wei
The Thirteenth International Conference on Learning Representations (ICLR) 2025
Develops a loss-adjustment approach for targeted LLM unlearning using only forget data, highlighting the trade-off between selective forgetting and retained model utility.
LLM Unlearning via Loss Adjustment with Only Forget Data Post-training
Yaxuan Wang, Jiaheng Wei, Chris Yuhao Liu, Jinlong Pang, Quan Liu, Ankit Shah, Yujia Bao, Yang Liu, Wei Wei
The Thirteenth International Conference on Learning Representations (ICLR) 2025
Develops a loss-adjustment approach for targeted LLM unlearning using only forget data, highlighting the trade-off between selective forgetting and retained model utility.
Noise-Resilient Point-wise Anomaly Detection in Time Series Using Weak Segment Labels Anomaly Detection
Yaxuan Wang, Hao Cheng, Jing Xiong, Qingsong Wen, Han Jia, Ruixuan Song, Liyuan Zhang, Zhaowei Zhu, Yang Liu
The 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD) 2025
Studies noise-resilient point-wise anomaly detection in time series using weak segment-level labels.
Noise-Resilient Point-wise Anomaly Detection in Time Series Using Weak Segment Labels Anomaly Detection
Yaxuan Wang, Hao Cheng, Jing Xiong, Qingsong Wen, Han Jia, Ruixuan Song, Liyuan Zhang, Zhaowei Zhu, Yang Liu
The 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD) 2025
Studies noise-resilient point-wise anomaly detection in time series using weak segment-level labels.
Large Language Model Unlearning via Embedding-Corrupted Prompts Unlearning
Chris Yuhao Liu, Yaxuan Wang, Jeffrey Flanigan, Yang Liu
The 38th Annual Conference on Neural Information Processing Systems (NeurIPS) 2024
Studies unlearning as a controlled way to intervene on memorized knowledge and test when apparent forgetting is robust or recoverable.
Large Language Model Unlearning via Embedding-Corrupted Prompts Unlearning
Chris Yuhao Liu, Yaxuan Wang, Jeffrey Flanigan, Yang Liu
The 38th Annual Conference on Neural Information Processing Systems (NeurIPS) 2024
Studies unlearning as a controlled way to intervene on memorized knowledge and test when apparent forgetting is robust or recoverable.