Tai Nguyen

May 2021: Hours before abseiling down Sơn Đoòng.

I’m Tai (Nguyễn Đặng Đức Tài), a graduate Computer Science student at the University of Pennsylvania.

I grew up in Saigon, Vietnam. 🇻🇳

I study methods that allow generalization from fewer examples, such as zero-shot & few-shot learning, and their applications in Information Extraction. I am also broadly interested in robustness and faithfulness of large language models. I currently work closely with Eric Wong.

Previously, I spent 2 memorable years as a data scientist at IBM Systems.
I helped build an analysis tool to support daily diagnosis and development of System-z mainframe servers.

Before discovering CS, I studied Economics and Statistics at Haverford College. I wrote my undergraduate thesis on the impact of home sharing (Airbnb) on welfare in New York City, working with professor Giri Parameswaran.

Some projects I have worked on:

Sports: Finalist for the 2022 NFL Big Data Bowl
Underthesea: an open-source NLP toolkit for Vietnamese 🌊
STEAM for Vietnam: volunteer data science work for a non-profit teaching free programming courses for Vietnamese children

I’m always open to chat and collaborate! Please reach my email at taing [at] seas [dot] upenn [dot] edu.

news

Jan 26, 2023	Wrote and submitted my first papers, both on improving robustness of LLMs 🤞
May 16, 2022	Started my research internship with Muhao Chen (USC) and Tianyi Zhang (Purdue)
Feb 16, 2022	🏆 Announced as a BDB finalist! 🏆
Aug 31, 2021	First day of graduate school at UPenn

publications

In-context Example Selection with Influences

Tai Nguyen, and Eric Wong

2023

Abs arXiv Code

In-context learning (ICL) is a powerful paradigm emerged from large language models (LLMs). Despite its promises, ICL performance is known to be highly sensitive to input examples. In this work, we use in-context influences to analyze few-shot ICL performance directly from the in-context examples. Our proposed influence-based example selection method outperforms most baselines when evaluated on 10 SuperGlue tasks and stably scales with increasing k-shot. The analysis finds up to a 22.2% performance gap between the most positively and negatively influential examples. In a case study, we apply our influence-based framework to quantify the phenomena of recency bias in example ordering for few-shot ICL.
Explanation-based Finetuning Makes Models More Robust to Spurious Cues

7 authors

2023