About Me

I am a first-year Ph.D. student at the Shenzhen International Graduate School, Tsinghua University (THU SIGS), advised by Prof. Xiaozhi Wang. Prior to this, I received my B.Eng. in Software Engineering from the Beijing Institute of Technology (BIT) in 2026.

My research interests broadly lie in Trustworthy AI and Reinforcement Learning. Specifically, I focus on:

• LLM Safety & Alignment: Building efficient and safe AI systems based on Large Language Models.
• RL & Reward Modeling for LLMs Studying reward design, reward hacking dynamics (e.g., in rubric-based RL/RLAIF), and verifiable training mechanisms.

Currently, I am working on the dynamics and mitigations of reward hacking in rubric-based reinforcement learning (introducing CHERRL (Controllable Hacking Environment for Rubric-based RL)).

I have been fortunate to gain extensive research experience during my undergraduate studies. I was a research intern at the State Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences (ICT, CAS), advised by Prof. Shengyu Zhu. During this internship, I proposed Speculative Safety-Aware Decoding (SSD), a novel decoding-time defense mechanism, which was accepted to the EMNLP 2025 Main Conference as a first-author paper (Paper PDF).

Before focusing on Safety-Aware Decoding, I researched defensive mechanisms in Federated Learning advised by Prof. Jinyan Liu, with one paper currently under review. I also worked on LLM Evaluation with Prof. Yang Song (UNCW), where we developed evaluation metrics for RAG-based chatbots, leading to a co-first author paper accepted by ICBAIE 2024.

Beyond research, I am a competitive programmer with a strong interest in algorithms. I have won the Silver Medal in the ACM-ICPC Regional Contest and the Silver Medal in the CCPC National Invitational Contest.

📄 Vita

You can find my detailed Curriculum Vitae here.

🧾 Selected Publications

• Reproducing, Analyzing, and Detecting Reward Hacking in Rubric-Based Reinforcement Learning, arXiv preprint arXiv:2606.04923 (2026)
  PDF

• Speculative Safety-Aware Decoding, EMNLP 2025 Main Conference (2025)
  PDF

💻 Code & Projects

My GitHub contains source code for my research projects and coursework.

• CHERRL (Controllable Hacking Environment for Rubric-based RL): A controllable benchmark and testbed for studying and detecting reward hacking in LLM-as-a-Judge reinforcement learning.
• Speculative Safety-Aware Decoding: Official implementation of my EMNLP 2025 paper.
• Improved A3C with Bandit Models: A course project for Reinforcement Learning at BIT (received full marks), exploring strategy improvements for A3C.
• Note: Some interesting repositories are currently private as I am organizing documentation and cleaning up the codebases.

📝 Blogs

I maintain a personal blog (in Chinese) where I document my learning journey in algorithms, competitive programming, and system design.

✉️ Contact

If you are interested in my research or would like to request paper preprints, please feel free to email me at wangxk0223@gmail.com.