Yifan Gao

Yifan is a Senior Applied Scientist at Amazon Rufus team working on building large language models for shopping. His work covers developing LLM scaling laws and utilizing findings to determine key pre-training modeling factors, predict benchmark performance, and guide pre-training data recipe. This enabled the team to train billion-level LLMs from scratch, surpass state-of-the-art LLMs in shopping knowledge, and serve Amazon Rufus experience.

Yifan obtained his Ph.D. in Computer Science at the Chinese University of Hong Kong under the supervision of Prof. Irwin King and Prof. Michael R. Lyu.

Research Internship positions in Amazon are available. Feel free to send your CV to me if you are interested in LLM pre-training.

My research interests are in natural language processing and machine learning.

During my PhD study, I teach machines to ask and answer reading comprehension questions driven by two goals:

• Knowledge Assessment: Asking questions is an effective approach to assess the understanding of users towards a passage of text. I work on 1) how to ask questions with different levels of difficulty [IJCAI'19]; 2) how to ask interconnected questions in a conversation for interactiveness and persistence [ACL'19], and 3) how to generate distractors (wrong options) in multiple-choice reading comprehension questions [AAAI'19].
• Information Acquisition: Many questions are inherently underspecified or ambiguous in our day-to-day conversations. In this case, the machine needs to 1) keep asking follow-up questions in a dialog until it gathers enough information to answer those high-level questions [ACL'20] [EMNLP'20] [arXiv'21]; or 2) present a series of disambiguated QA pairs when the query is ambiguous[ACL'21].