Hi there! I am a Master's student in Computer Science at ShanghaiTech University, where I am fortunate to be advised by Prof. Kan Ren.
My research focuses on LLM interpretability: understanding what large language models encode in their internal representations and how those representations give rise to their behavior.
Building on this understanding, I am particularly interested in leveraging representations to empower LLMs in agentic scenarios, from steering and enhancing their reasoning to guiding the self-evolution of agent skills.
Interpreting and Controlling LLM Reasoning through Integrated Policy Gradient got accepted to NeurIPS 2026!
arXiv preprint arXiv:2609.39149, 2026
Moves skill evolution beyond the text space by using the agent's internal representations to guide how reusable textual skills are diagnosed and revised.
arXiv preprint arXiv:2605.30076, 2026
A text-guided activation flow matching model for steering LLMs toward fine-grained and compositional behaviors without fixed steering directions.
Conference on Neural Information Processing Systems (NeurIPS) 2026
Localizes the internal components that drive complex LLM reasoning and uses them to enhance or suppress reasoning behaviors.
Exchange Student
Email
Google Scholar
GitHub
Paper