Litian Liu 刘力田

I am a researcher in the Large Language Model Team at Qualcomm AI Research. Prior to joining Qualcomm, I earned my PhD from MIT in 2021, advised by Prof. Muriel Médard. I received my M.Eng from Princeton University in 2017, advised by Prof. Mung Chiang. I earned my B.Eng from the Chinese University of Hong Kong in 2016, where I closely worked with Prof. Minghua Chen and completed my thesis under the supervision of Prof. Xiaoou Tang. In 2015, I was an exchange student at MIT, where I was fortunate to work with Prof Vincent Chan.

My projects involve both producterization and research aspects. On the product side, our demo of an on-device voice assistant was highlighted by Qualcomm CEO Cristiano Amon during his keynote talk at the Snapdragon Summit 2023, earning me an Impact Award from Qulacomm Product Marketing. On the research side, my current interest lies in the safe deployment of AI. Particularly, I’m interested in addressing:

知之为知之 不知为不知 是知也. — 《论语· 为政》

This is wisdom: to recognize what you know as what you know, and recognize what you do not know as what you do not know. — The Analects 2.17 [1]

As Confucius (who, like me, is from Shandong, China) noted, true wisdom involves recognizing the limits of our knowledge. This presents a significant challenge for both humans and today’s machine learning models. For example, a trustworthy classifier should raise an alert when encountering samples of classes unseen during training, as the classifier cannot make meaningful predictions. This corresponds to the field of Out-of-Distribution Detection. Similarly, when large language models face queries beyond their knowledge, they should transparently raise an alert instead of producing misleading outputs. This corresponds to the field of Hallucination Mitigation.

If you are interested in the topics, check out my recent publications below. Feel free to reach me at litiliu at qti.qualcomm.com