About

Biography

I am a Research Fellow at the School of Computing, National University of Singapore, working with Prof. Mohan Kankanhalli. I work on multimedia retrieval and multimodal large language models. A central question in my research is how these models acquire and represent knowledge, how their use of that knowledge can be grounded in evidence, and how external knowledge can be used to improve retrieval and reasoning.

My interest in computing began at thirteen, when I learned Pascal for informatics competitions. I later moved to C++, received a provincial first prize in NOIP, and earned a special-admission opportunity at Shandong University. Competitive programming taught me to look for precise solutions under explicit rules. Research gradually drew me toward problems where neither the rules nor the answers were quite so clean. These days, I write almost everything in Python, but the habit of turning an ambiguous question into something testable has stayed with me.

At Shandong University, I completed my Ph.D. under the guidance of Prof. Liqiang Nie and Prof. Yupeng Hu. Much of my doctoral research asked how models could learn fine-grained video-language correspondence from limited or imperfect supervision. I worked on temporal localization with sparse annotations, as well as cross-modal retrieval problems in which only part of the retrieved content is relevant to a query. Over time, I came to see retrieval as more than ranking similar items: it is also about identifying which pieces of evidence genuinely support a match.

I have also taken part in several research challenges, where our teams earned multiple first-place finishes. On the systems side, I contributed to benchmarking and submitting QSG-NGT, an approximate nearest-neighbor search algorithm, to ANN-Benchmarks. Working on QSG-NGT made me appreciate how much practical retrieval depends on the balance between accuracy, efficiency, and scalability, not just on the quality of the representation.

At NUS, I am continuing this line of work on multimodal large language models. I am interested in how context and retrieved evidence interact with a model's internal knowledge. I also want to understand how to distinguish evidence-grounded predictions from outputs driven by memorization or spurious correlations, and how reliable external knowledge can be integrated to improve multimodal understanding. My work focuses on retrieval, attribution, and verification mechanisms that make both the sources and use of model knowledge more transparent and trustworthy.

Outside research, I enjoy photographing landscapes and ordinary moments with my Fujifilm X-T5. I also write here from time to time, mostly to think through ideas rather than present polished conclusions. I grew up in Fujian and remain fond of its culture. I speak Hokkien, Mandarin, and English, and can understand some Cantonese. I also make time for tea, working out, and climbing whenever I can. I enjoy a good challenge, and I value the people I meet along the way. Feel free to contact me.