Hi, I'm Yifei.

My research focuses on large language models and agentic systems, with an emphasis on reliability, self-improvement, and efficiency. My broader goal is to build capable and trustworthy AI systems for real-world applications.

Open to research internship opportunities in AI agents, reliability, and AI safety.

Education & research

2026 - Present

University of North Carolina at Chapel Hill

Ph.D. in Computer Science. Advised by Prof. Zhun Deng, working on AI agents, reliability, and AI safety.

2023 - 2025

Carnegie Mellon University

M.Sc. in Information Security, CQPA: 4.0/4.0. Worked with Prof. Steven Wu on AI safety and privacy.

Outstanding Student Services Award - Research Assistant (2025) Presented by the Information Networking Institute.

2019 - 2023

Zhejiang University

B.Eng. in Computer Science and Technology. Worked with Prof. Kai Bu on network security and path validation.

Agentic systems Self-evolving agents AI reliability AI safety

Recent work

ReliabilityCOLM 2026 · Equal contribution

Rubrics as an Attack Surface: Stealthy Preference Drift in LLM Judges

Ruomeng Ding*, Yifei Pang*, He Sun, Yizhong Wang, Steven Wu, Zhun Deng

AI PrivacyNeurIPS 2025

Rethinking Exact Unlearning under Exposure: Extracting Forgotten Data under Exact Unlearning in Large Language Models

Xiaoyu Wu, Yifei Pang, Terrance Liu, Steven Wu

AI PrivacyTheory and Practice of Differential Privacy 2025

Winning the MIDST Challenge: New Membership Inference Attacks on Diffusion Models for Tabular Data Synthesis

Xiaoyu Wu, Yifei Pang, Terrance Liu, Steven Wu

Network SecurityIEEE TDSC 2026

SwiftOracle: Orthogonality-driven Private Multipath Validation

Yifei Pang, Anxiao He, Wenjie Hou, Yunyi Teng, Kai Bu, Qianping Gu, Kui Ren

View all publications

Projects

Carnegie Mellon University · 2025

Membership inference on fine-tuned LLMs

Designed and deployed the first project on membership inference attacks for CMU 18-734/17-731, Foundations of Privacy, including its evaluation pipeline, baselines, codebase, and leaderboard.

Vector Institute MIDST Challenge · 2025

Membership inference for synthetic tabular data

Built stable white-box and black-box attacks for diffusion-based tabular synthesis. Our team placed first across all four competition tracks among 71 participants.

Teaching

Fall 2025

Foundations of Privacy

Teaching Assistant · Carnegie Mellon University · 18-734/17-731

Fall 2024

Fundamentals of Telecommunications Networks

Teaching Assistant · Carnegie Mellon University · 14-740

Let's connect

Interested in research internships and collaborations.

If you are working on AI agents, reliability, or AI safety, I would be glad to hear from you.

ppppyf4534@gmail.com