Research
Research agenda
My research focuses on LLM agents and LLM safety.
I study how agents become more capable in real systems, and how they can remain safe when working with tools, context, other agents, and external environments.
LLM Agents
I study the systems and reasoning methods that allow language agents to use tools, discover capabilities, coordinate with other agents, and complete complex tasks. My recent work covers agent harnesses, active tool discovery, multi-agent reasoning, and agent evaluation.
Representative work SING, A Survey on AI Agent Harness, and MASLegalBench.
LLM Safety
I study safety beyond input and output filtering. My work examines system boundaries, contextual integrity, privacy, policy reasoning, and legal compliance for both language models and agents.
Representative work Isolation for LLM-Agent Safety, MCIP, ContextLens, OmniCompliance-100K, and PrivaCI-Bench.
Research approach
Across both directions, I build benchmarks, datasets, system protocols, and empirical analyses. These tools help reveal where models and agents fail, and whether a proposed method improves the behavior that matters in practice.
Research experience
Working with Haoran Li and advised by Prof. Yangqiu Song.
Advised by Ziyi Liu and Prof. Xiaofang Zhou.