Multimodal Agents
Designing systems that can search, perceive, and reason over heterogeneous digital environments, with emphasis on grounded multimodal interaction.
Academic Homepage
Zhe Yang
Master's student in Artificial Intelligence (MSAI) at Nanyang Technological University, working on multimodal and agent-related research with Associate Professor Ziwei Liu at MMLab@NTU.
My current research focuses on multimodal learning, agent systems, and personalized AI. I previously worked on LLM memorization, data privacy, and tokenization-related security risks with Prof. Michael R. Lyu at CUHK.
About
Multimodal AI, agent systems, and grounded personalization
I am currently pursuing the Master of Science in Artificial Intelligence (MSAI) at Nanyang Technological University. Since August 2025, I have been working with Associate Professor Ziwei Liu at MMLab@NTU on papers and research projects, with a current focus on multimodal systems and agentic intelligence.
Before NTU, I received my B.Sc. in Computer Science from The Chinese University of Hong Kong, graduating with a CGPA of 3.785/4.0. My earlier work centered on understanding memorization in large language models, with emphasis on data compressibility, tokenization, and security risks in code LLMs.
Education
- 2025.08 - Present M.S. in Artificial Intelligence, NTU
- 2021.09 - 2025.07 B.Sc. in Computer Science, CUHK
- 2023.07 Summer School, Peking University
Research Interests
- Multimodal learning and grounding
- Agent systems and contextual reasoning
- Personalized AI and memory-centric systems
- LLM memorization, privacy, and security
Research
Current directions
Personalized AI
Building and evaluating systems that recover user-level context from long-horizon, file-system-scale behavioral and multimodal traces.
LLM Memorization and Security
Studying how data properties and tokenization affect memorization, leakage risk, and privacy vulnerabilities in large language models.
Publications
Selected papers
The papers below include recent work currently under review and ongoing research projects. At this stage, the four papers listed here are all in the review cycle.
Under Review
2026
HippoCamp: Benchmarking Contextual Agents on Personal Computers
A benchmark for evaluating contextual agents over realistic multimodal personal file systems, with a focus on search, evidence perception, and multi-step reasoning.
Explore the interactive website to experience what the three personal computers look like, and visit the project page for the benchmark overview.
Under Review
2026
FileGram: Grounding Agent Personalization in File-System Behavioral Traces
A framework for grounding agent memory and personalization in file-system behavioral traces, spanning data generation, benchmarking, and memory architecture.
Under Review
2026
Data Compressibility Quantifies LLM Memorization
This work studies how data compressibility relates to memorization in LLMs and proposes a quantitative perspective on memorization behavior.
Under Review
2026
Understanding Secret Leakage Risks in Code LLMs: A Tokenization Perspective
An investigation into how BPE tokenization contributes to secret memorization and leakage risks in code LLMs through what we term gibberish bias.
Published
2024
RAUIE: A Relation-Augmented Document-level Event Extraction Model Based on UIE
Presented at AINIT 2024. Earlier work on document-level information extraction.
Experience
Research and training
2025.08 - Present
Research with Associate Professor Ziwei Liu, MMLab@NTU
Working on multimodal and agent-oriented papers and research projects while pursuing the M.S. in AI at NTU.
2024.09 - 2025.07
Final-Year Research with Prof. Michael R. Lyu
Focused on LLM memorization, entropy-based characterization, dataset inference, and tokenization-related security risks in code LLMs.
2024.07 - 2024.08
Algorithm Intern, Geovis Technology
Worked on automated data acquisition, knowledge graph modeling, fine-tuning, and graph-based analysis pipelines.
2024.06 - 2024.09
UG Summer Research Internship
Evaluated memorization difficulty in large language models through entropy, perplexity, and memorization-rate analysis on open-source models.
2023.06 - 2023.09
UG Summer Research Internship
Studied adversarial attacks on gender recognition systems and analyzed robustness and fairness issues in facial-recognition APIs.
Honors
Selected awards
HKSAR Government Scholarship
Top 1% in Computer Science.
Dean's List
Top 10% in the Faculty of Engineering, CUHK.
CSE Scholarship Silver Award
Top 1% in the Faculty of Engineering, CUHK.
Contact
Get in touch
If you would like to discuss multimodal learning, agent systems, personalized AI, or LLM memorization and security, feel free to reach out.