Zhexi Lu, Hongliang Chi, Nathalie Baracaldo, Swanand Ravindra Kadhe, Yuseok Jeon, Lei Yu
About
Hi! I'm Zhexi, a Ph.D. student in Computer Science at Rensselaer Polytechnic Institute (RPI), advised by Prof. Lei Yu. My research focuses on post-training for large language models and agents. I study how on-policy distillation and reinforcement learning improve reasoning and tool use, and how these methods affect generalization, memorization, privacy, and reliability. I am especially interested in how different learning signals shape model behavior, and whether capability gains are tied to memorization of training data. Recently, I have extended these questions to agents that reason over multiple steps and call tools. My long-term goal is to understand how post-training works, and to use that understanding to build models and agents that are more capable, efficient, and trustworthy.
Post-training
On-policy distillation
Reinforcement learning
LLM agents
Memorization & privacy
News
Sep 2026Two new preprints on arXiv: ACSD and Leaky Students.
May 2026Back at IBM Research for a second summer.
Apr 2026Nominated by RPI for the Google PhD Fellowship.
Feb 2026Presented In-Context Probing at NDSS 2026 in San Diego (talk video).
Jan 2026Leading an IBM–RPI FCRC project on failure analysis for LLM agents.
Dec 2025In-Context Probing (NDSS 2026) is on arXiv.
May 2025Joined IBM Research in Yorktown Heights as a summer visiting student.
Publications
Zhexi Lu, Subhajit Chaudhury, Tejaswini Pedapati, Keerthiram Murugesan, Lei Yu
Preprint
arXiv
Zhexi Lu, Mingzhi Zhu, Stacy Patterson, Lei Yu
Preprint
arXiv
Internalizing State Tracking for Reliable Multi-Turn Tool Agents
Zhexi Lu, Subhajit Chaudhury, Tejaswini Pedapati, Keerthiram Murugesan, Lei Yu
Preprint
FDRShare: A Fully Decentralized and Redactable EHRs Sharing Scheme with Constant-Size Ciphertexts
Zhichao Li, Zhexi Lu, Lingshuai Wang, Qiang Wang, Che Bian
ICA3PP 2023
Experience
IBM Research, Summer Visiting Student
Self-distillation for LLM reasoning
May–Aug 2026Self-distillation for LLM reasoning
IBM Research, Summer Visiting Student
Failure detection in multi-turn agentic interactions
May–Aug 2025Failure detection in multi-turn agentic interactions
Rensselaer Polytechnic Institute, Research Assistant
Membership inference attacks against fine-tuned LLMs
Sep 2024–presentMembership inference attacks against fine-tuned LLMs
Honors
Google PhD Fellowship Nominee, RPI · 2026
Second Prize, National College Student Information Security Competition · 2023
Third Prize, MCM/ICM · 2023
Service
Reviewer: NeurIPS 2025, AsianCCS 2025
Artifact Reviewer: ACM CCS 2025