Education
- 2024-Present: Ph.D. in Shanghai Jiao Tong University, advised by Prof. Yu Yu and Prof. Yuncong Hu
- 2020-2024: B.Eng. in Xidian University, advised by Prof. Hui Li
Publications
Human-Readable Fingerprint for Large Language Models
We propose HuRef, a training-free fingerprinting method for LLM copyright protection. We discover that the direction of LLM parameter vectors remains stable across different training stages, and design three invariant terms to uniquely identify a model's base origin. The fingerprint is mapped to a human-readable image via StyleGAN2, with zero-knowledge proofs ensuring honest generation.
zkAgent: Verifiable Agent Execution via One-Shot Complete LLM Inference Proof
We present zkAgent, a system that uses zero-knowledge proofs to verify the entire agent execution including end-to-end LLM inference and tool interactions. By generating a single one-shot inference proof, zkAgent achieves a 294x speedup over prior work (zkGPT), reducing proving time to 1.05s/token for GPT-2 with 512 tokens, making verifiable agent execution practical for real-world deployment.
Experience
- Research Intern, Shanghai Qi Zhi Institute
Research about High Throughput Signature Verification, 2024.3-2024.8
Links
- LatticeLab @ sjtu.edu.cn
- Moxuan Wang @ xidian.edu.cn
- Hao Jiang(d33b4t0) @ iie.ac.cn
- Zihan Zhang(Ghost1032) @ sjtu.edu.cn
- Zhuo Huang(sh1k4ku) @ sjtu.edu.cn
- Yituo He @ sjtu.edu.cn
- Tianyu Zhang @ illinois.edu
- Qi Liu @ sjtu.edu.cn
- Yang Yang @ sjtu.edu.cn
Notes
Interesting Tools
- proxy_tool
One-liner shell helper to redirect proxy envs on remote SSH sessions to your local PC proxy, with auto client-IP detection. - beamer-skill
Give it a paper, and in ~10 minutes with ~6$ of Claude Opus 4.6[1M] tokens, it produces a presentation-ready Beamer slide deck you can almost directly use for your group seminar. [sample slide]