My research vision is to make data more usable, then make AI more reliable — by pairing external knowledge graphs with internal agent memory. Rooted in data management, I build retrieval-augmented generation and intelligent assistants that remember, listen, and act reliably, from question answering to scientific discovery. Ongoing work extends agent memory toward long-horizon conversation, ambient speech, and vision-language-action.
My work follows one thread — make data more usable, then make AI more reliable — across three layers: building assistants whose internal memory, perception, and actions stay trustworthy, grounding what they answer in external knowledge graphs, and making the underlying data usable in the first place.
Click an area to filter the publication list below.
Assistants that remember, listen, and act. Memory is the core: how it goes stale, how to audit it, how to build it over long horizons. Ongoing work extends to ambient speech and vision-language-action agents.
Grounding LLM answers in retrieval and knowledge graphs. Co-organized the Meta KDD Cup 2024 on CRAG; KERAG and CacheRAG push retrieval quality and efficiency.
Making structured and scientific data usable by models — from table understanding and annotation to agentic SQL interfaces.
* Equal contribution + Corresponding author