CV

Research Intern — AI Multibit Robust Watermark

University of California, Santa Barbara

Advisor: Prof. Yuheng Bu; ARR under submission

• Developed robust multi-bit text watermarking for content attribution, enabling identifier recovery after paraphrasing without access to the original text or watermark-specific training.
• Achieved 99.83% watermark-bit injection success and 99.70% exact recovery of 32-bit identifiers in 10,000 resampled trials under the evaluated rewriting protocol.
• Recovered identifiers from all 20 evaluated Wikipedia and arXiv documents after rewriting at a target code rate of 0.2.

Undergraduate Researcher — Defense Against Prompt Injection in LLM Agents

Tsinghua University, College of AI

First author; advisor: Prof. Yinpeng Dong

• Proposed a non-intrusive inference-time defense for indirect prompt injection by detecting intent drift around tool calls and tool responses.
• Built a Qwen3-8B tool-calling agent pipeline with hooks over the final four transformer layers, paired clean and injected trajectories, contrastive probes, and automatic attack-success labeling.
• Constructed a 1,220-task benchmark and a 4,000-case cross-domain suite spanning ten application domains.
• Automated injection generation and genetic search over a 3,072-configuration strategy space; research in progress.

Research Intern — LLM Agent Security

Tencent Xuanwu Lab

• Explored early defense ideas and used Tencent-provided compute resources to support large-scale experiments for the Tsinghua-led project.

AI4Science Research Assistant — Crystal-Structure LLM Post-Training

Tsinghua University, Department of Computer Science and Technology

Advisor: Prof. Yang Liu

• Contributed to a three-stage LLM post-training workflow covering scientific-data preparation, supervised fine-tuning, and reinforcement-learning-based optimization.
• Collected and cleaned crystal-structure data and aligned prompts, model inputs, and target outputs for training and evaluation.
• Designed reasoning-oriented reward functions and explored reward shaping and prompt-alignment strategies.

B.S.; B.Eng.

Tsinghua University

Mathematics and Physics; Software Engineering

• Dual-degree program, GPA 3.74/4.0

University of California, San Diego

UCEAP Exchange Program

• Incoming UCEAP exchange student for the Winter and Spring quarters of 2027 (two quarters).
• Selected for Tsinghua University’s sponsored exchange program, with one seat per department.

Understanding Is Done Early: A Depth Division of Labor in Large Language Models and Its Use for Unbounded-Context Memory

* Co-second authors. Co-developed CoMem, a Qwen3-8B memory architecture that caches intermediate residual states and recomputes query-conditioned upper layers for efficient long-context memory.

FinAgent OS: Multi-Agent Wealth Management System

Group project for the Software Engineering course. GitHub repository

• Built a LangGraph- and FastAPI-based multi-agent system for user profiling, asset screening, market timing, critic review, and compliance auditing across 54 instruments in nine asset classes.
• Implemented Markowitz optimization, risk parity, Monte Carlo simulation, VaR and CVaR, stress testing, backtesting, and performance attribution.
• Integrated AkShare and yfinance data, persistent SQLite sessions, Docker deployment, and GitHub Actions CI while maintaining a 269-test passing baseline.

TASIVE: Transformer-Based Siamese Signature Verification

Lead developer; major project for the Deep Learning course. GitHub repository

• Developed a Swin-T Siamese network with a lightweight MLP projection head, L2-normalized 128-dimensional embeddings, and a scaled cosine binary-cross-entropy objective.
• Designed a strict writer-disjoint train and validation split and evaluated on a fixed all-pairs validation set of 5,748 signature pairs.
• Achieved 99.98% accuracy, 1.0000 ROC-AUC, and 0.0000 EER, outperforming a reproduced SigNet baseline under the same protocol.