CV

A PDF version of this page is available here: Jesson’s Curriculum Vitae.

Research interests

AI safety and trustworthy machine learning — with a current focus on large language models: alignment, reinforcement learning, agentic RL, and adversarial robustness.

Education

• M.S. in Computer Science, University of Southern California, 2025 – present
• B.Eng. in Computer Science, School of Computer Science, Wuhan University, 2021 – 2025
  • Exchange student, EECS Department, UC Berkeley, Jan. 2024 – May 2024
  • Wuhan University Excellent Exchange Student Scholarship (top 0.4%)

Research experience

• Jan. 2024 – May 2025: Research Intern
  • EECS Department, UC Berkeley
  • Advised by Prof. David Wagner and Postdoc Zhanhao Hu
  • JULI: Jailbreak Large Language Models by Self-Introspection (May 2024 – May 2025). Jailbreaking LLMs by manipulating token log probabilities through a tiny plug-in block, BiasNet. JULI relies solely on the target model’s top-5 predicted token log probabilities, so it works against API-only models in a black-box setting. The released fine-tuned models have over 4,000 downloads on Hugging Face.
  • Reject Option: Eradicating Harmful Content with a Tiny Classifier (Jan. 2024 – May 2024). Detecting and rejecting harmful LLM responses with a classifier of only a few linear layers, matching the performance of Llama Guard.
• Sept. 2022 – Jan. 2024: Research Intern
  • CSE Department, Hong Kong University of Science and Technology
  • Advised by Prof. Qian Zhang
  • MobHAR: Imperceptible Knowledge Transfer for Human Activity Recognition on Mobile Devices (Aug. 2023 – Jan. 2024). A user-centric HAR customization framework built on an adversarial mechanism that enables imperceptible knowledge transfer. Accepted at IMWUT.
  • ARTEMIS: Defending Against Backdoor Attacks via Distribution Shift (Dec. 2022 – Jul. 2023). A backdoor defense that uses distribution shift to close the feature-space gap between poisoned and benign samples. Accepted at IEEE TDSC.

Publications

JULI: Jailbreak Large Language Models by Self-Introspection

Jesson Wang, Zhanhao Hu, David Wagner
International Conference on Learning Representations (ICLR), 2026

MobHAR: Source-free Knowledge Transfer for Human Activity Recognition on Mobile Devices

Meng Xue, Yinan Zhu, Wentao Xie, Zhixian Wang, Yanjiao Chen, Kui Jiang, Qian Zhang
Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies (IMWUT), 2025

ARTEMIS: Defending Against Backdoor Attacks via Distribution Shift

Meng Xue, Zhixian Wang, Qian Zhang, Xueluan Gong, Zhihang Liu, Yanjiao Chen
IEEE Transactions on Dependable and Secure Computing (TDSC), 2025

Teaching

Teaching Assistant

Data Structure, Wuhan University, School of Computer Science, 2024

Skills

• Languages: Python, C++
• Frameworks and tools: PyTorch, Git