I am a Ph.D. student in Electrical Engineering at the University of Notre Dame. Previously, I received my B.S. in Computer Science from Beijing Jiaotong University.
I study self-improving and reliable LLM agents. My research asks when learning and experience actually improve an agent’s decisions: does acquired knowledge change its actions, and can useful experience transfer across tasks and models? I work on agentic post-training, experience memory, and reliable long-horizon behavior, using tool-using and interactive agents to examine the connection between learning, memory, and action.
My broader research spans LLM adaptation for human mobility, including Mobility-LLM (NeurIPS 2024), and deep learning for spatial sensing and environmental mapping.
My long-term goal is to understand how agents turn experience into lasting, transferable improvements in reasoning and action. My current work connects Learning → Memory → Reliable Action.
I study how post-training and feedback change an agent’s decisions. Know2Act tests when an agent’s written prediction steers its next action and how the fine-tuning data decide this; my search-agent study asks what training against prompt injection actually teaches an agent to resist.
I examine when experience memory provides reusable knowledge across tasks and models. SkillRC uses a matched control to separate the cost of reading a memory from the value of what it says.
I use controlled experiments and failure analysis to evaluate agent behavior under changing conditions. Building on this work, I aim to understand when lasting improvements require coordinated updates to the policy, memory, and execution structure.
Current work on agents, alongside earlier work in spatial sensing and mobility.
Splits an experience memory’s effect on an agent into the cost of reading it and the value of what it says, and finds that the differences it can attribute between models lie in what the memory says.
Shows that training a search agent against injected instructions teaches it to resist instructions that push a wrong answer while still following ones that push the right answer, and barely changes how it treats the same wrong answer stated as a fact.
Tests whether an agent’s written prediction steers its next action, and finds that fine-tuning makes it steer when the training data leave the choice open, while a modest share of contradicted predictions removes this control.
Under review at IEEE ICC 2027 · SAC – Integrated Sensing and Communications
Ultra-wideband localization and deep-learning-based environmental mapping for integrated sensing and communications.
Under review at IEEE ICASSP 2027
A multi-stage deep-learning and signal-processing pipeline for estimating reflective boundaries from noisy acoustic measurements.
Adapts pretrained LLMs to human mobility through visiting-intention memory and travel-preference prompts.
75-MINUTE LECTURE
CSE 60556 LLM · University of Notre Dame
Where should knowledge live in an LLM system? From model parameters and context to external memory, this lecture introduces RAG, examines its limitations, and explores approaches including Self-RAG and GraphRAG.
SlidesTEACHING ASSISTANT
Supported hands-on learning through lab instruction, student support, project debugging, and assessment.
Meet Chocoliz (巧乐兹), my cat.
The familiar face hiding behind my portrait above.
