Columbia University
September 2025 – May 2029
B.A. in Computer Science and Applied Mathematics
Hello!
I am a sophomore at Columbia University, working on mechanistic interpretability and agentic systems for software engineering. I am particularly interested in representation analysis and agent output verification. I aim to better observe and understand large language models and integrate interpretability methods into agentic systems.
You can reach me at hj2742 [at] columbia.edu.
Education
Beijing National Day School
September 2021 – May 2025
High School
Experience
Research Assistant at ARiSE Lab, Columbia University
September 2025 – Present
Advised by Professor Baishakhi Ray and Professor Simin Chen.
AI-driven Software Engineering. Trustworthy Coding Agents, Monitoring and Verification Systems, Agent Training.
Research Assistant at DitecT Lab, Columbia University
January 2026 – May 2026
Advised by Professor Sharon Di.
LLM post-training. Enhancing LLM-human behavioral consistency for social science research.
Research Assistant at PILLAR Lab, Peking University
May 2026 – Present
Advised by Professor Liangming Pan.
Mechanistic Interpretability. Representation Analysis and Engineering, Natural Language Autoencoders, and SAEs.
Research Engineer, KAV Labs
September 2026 – Present
A startup based at Columbia, enterprise-scale software engineering agents with policy inference and formal verification.
Research
SIGIL: Securing Agentic Code Generation with Intent Code Graph Learning
Under review at ICLR 2027 PDF
Construct LLM-derived semantic graphs linking issue subtasks to code-chunk diffs and classify them (via GNNs or manually extracted features) to flag injected instructions that hijack SWE agents into producing vulnerable patches.
Keywords: Agents for Software Engineering, Prompt Injection, Model Hijack, Graph Neural Networks
Does Steering Break Your Model? A Multi-Dimensional Evaluation Suite for LLM Steering Methods
Under review at ICLR 2027 arXiv
Develop a comprehensive evaluation framework for representation engineering that compares methods across effectiveness, side effects, robustness, generalization, and data efficiency over a range of intervention methods.
Keywords: Representation Engineering, Activation Steering, Side Effects
Award
- International Olympiad in Artificial Intelligence (IOAI)Bronze Medalist, World Rank 99
- China’s National Olympiad in Artificial Intelligence (NOAI)High Distinction, National Rank 5
- China Computer Federation, National Driverless Car ContestDistinction, National Rank 8
- American Mathematics Competition (AMC)130.5 Points, Honor Roll, Global Top 5%
- American Invitational Mathematics Examination (AIME)12 points, Global Top 5%
- International Mathematical Modeling Challenge (IMMC)Meritorious Award
Service
Member, Science Committee of National Olympiad in Artificial Intelligence (China)
November 2025 – October 2026
Designed contest questions and sample exams.
Examiner, USA–North America AI Olympiad
January 2026 – August 2026
Designed and validated contest questions and sample exams.
Programming Skills
Java, Python, PyTorch