Assistant Professor · College of Intelligent Computing
Chang Gung University, Taiwan
長庚大學 智慧運算學院 · 助理教授
I study the reliability of large language models: whether they know the limits of their knowledge, express appropriate confidence, and keep to the facts when users push back. 我研究大型語言模型的可靠性：模型是否清楚自己的知識邊界、表達的信心是否與正確率相符，以及在使用者施壓時能否堅持事實。
I am looking for undergraduate and graduate students interested in LLM reliability, alignment, evaluation, and clinical NLP. Please email me your CV and transcript, and tell me in a few sentences about a paper or problem you find interesting. Prior research experience is not required.
歡迎對大型語言模型可靠性、對齊、評估方法，以及醫療自然語言處理有興趣的大學部與研究所同學。請來信附上履歷與成績單，並用幾句話說明一篇你覺得有趣的論文或問題。不要求先前的研究經驗。
I am an Assistant Professor in the College of Intelligent Computing at Chang Gung University. My research is on the reliability of large language models, with a focus on sycophancy: models agreeing with users even when the user is wrong. I work on how to measure this behavior, how training causes it, and how to reduce it. I also build NLP systems for clinical text with physicians, including psychiatric records and periodontal charting, where a confident but wrong answer can affect patient care. I received my Ph.D. from National Taiwan University.
我是長庚大學智慧運算學院助理教授，研究主題是大型語言模型的可靠性，特別是奉承行為（sycophancy）：即使使用者是錯的，模型仍傾向附和。我研究如何量測這種行為、訓練過程如何造成它，以及如何降低它。我也與醫師合作開發臨床文本的自然語言處理系統，包含精神科病歷與牙周檢查紀錄；在這些場域，一個有自信但錯誤的答案會直接影響病人照護。博士畢業於國立臺灣大學。
My group works on three connected questions. 我的研究圍繞三個彼此相連的問題。
How do we quantify whether a model knows what it knows? Benchmarks and metrics for knowledge boundaries, calibration, and sycophancy, with attention to multi-turn dialogue and Mandarin expert domains.
如何量化模型是否知道自己知道什麼？針對知識邊界辨識、信心校準與奉承行為建立評測基準與指標，特別著重多輪對話與中文專業領域。
What training and inference-time interventions make models more truthful? Preference alignment that resists social pressure, grounded in the model's own internal states.
什麼樣的訓練與推論階段介入能讓模型更誠實？發展能抵抗社會壓力、並以模型內部狀態為依據的偏好對齊方法。
How can LLM systems be used safely in regulated settings? Clinical, legal, and compliance applications where every output can be traced back to its evidence.
如何在受監管的場域安全地使用 LLM 系統？涵蓋醫療、法律與法遵應用，並要求每個輸出都能追溯到依據。
* equal contribution* 共同第一作者
