Welcome
About me
I am an M.S. student in Artificial Intelligence at the School of Data Science, Fudan University, advised by Prof. Yanwei Fu. I am also a research intern at Shanghai Artificial Intelligence Laboratory, working with Dr. Dongrui Liu.
My research focuses on safe and trustworthy AI agents and data-centric visual learning. I work on evaluating agent behavior, building diagnostic guardrails, and learning across domains with limited supervision.
Before joining Fudan, I received my B.S. in Software Engineering from Dalian University of Technology. I am open to collaboration and discussions — feel free to get in touch.
Current researchExplore research →
Agent safety & trustworthiness
Trajectory evaluation, step-level guardrails, and conflicts in shared environments.
Data-centric visual learning
Retrieval-guided generation and cross-domain few-shot object detection.
Recent updatesView publications →
- Sep 2026ClashBench was released on arXiv, benchmarking agent conflicts over shared resources.
- Aug 2026StepGuard was accepted to EMNLP 2026.
- Jul 2026ATBench was accepted to COLM 2026.
Earlier updates
- Apr 2026ATBench was released on arXiv for trajectory-level agent safety evaluation and diagnosis.
- Jan 2026AgentDoG was released as a diagnostic guardrail framework for AI agent safety and security.
- 2025Domain-RAG was accepted to NeurIPS 2025.
Research
What I work on
Trustworthy agents and visual learning across domains.
Agent safety & trustworthiness
I study how to evaluate and diagnose safety failures in AI agent trajectories, and how guardrails can balance safety with task utility. This includes step-level supervision, realistic evaluation benchmarks, and conflicts between agents sharing resources.
Data-centric visual learning
I explore retrieval-guided compositional image generation for cross-domain few-shot object detection, with an interest in improving visual learning when target-domain annotations are limited.
Also interested in
Trustworthy multimodal foundation models, jailbreak and defense, alignment, interpretability, efficient reasoning, computer-use and mobile agents, and embodied agent safety.
Publications
Selected publications
* denotes equal contribution. Full list on Google Scholar ↗
Peer-reviewed publications3 papers
Preprints3 papers
Experience
Research experience
Nov 2025 – Present
Shanghai Artificial Intelligence Laboratory
Research Intern · Center for Safety and Trustworthiness
Working with Dr. Dongrui Liu.
Worked on data pipelines and benchmark construction for trajectory-level agent safety diagnosis, including scalable synthesis, filtering, and evaluation data design.
Education
Academic background
2024 – 2027
Fudan University
M.S. in Artificial Intelligence · School of Data Science
Advisor: Prof. Yanwei Fu
2020 – 2024
Dalian University of Technology
B.S. in Software Engineering
Contact
Get in touch
I welcome conversations and collaborations on agent safety, trustworthy AI, and visual learning. Email is the best way to reach me.
- GitHub
- github.com/liyu0524 ↗
- Google Scholar
- Publications & citations ↗