My research focuses on enabling humans to better control and collaborate with AI. To achieve this goal, I devise AI agents that can communicate effectively with humans. For a long time, we have been communicating with AI agents in very primitive ways: we teach them with rewards and labels; they convey uncertainty to us through probabilities; most of the time, they perform tasks own their own without sharing information or asking us to help. LLMs offer better ways to control AI through language, but it is unclear whether key elements of human-like communication like inferentiality are present. My goal is to create new foundations for helpful and safe AI by going beyond the current primitive forms of human-AI communication. Specifically, I have been pushing forward three directions:
Learning from natural human feedback: my paper [EMNLP’17] is one of the first to analyze the benefits and risks of using reinforcement learning from human feedback (RLHF) for text generation. Since then, I don’t think teaching agents with only numbers is a good idea. More recently, I developed a theoretically-grounded framework for learning from language feedback [ICML’21].
Learning to express uncertainties and ask questions: It is a mistake to think that only humans should ask AI for help and not the reverse. By asking a question, an agent can: (i) express its uncertainties (not just uncertainty), and (ii) obtain information to expand its task-solving capabilities. So more safety and more utility! I author a series of papers that highlight the challenges and devise solutions for the problem of learning to convey uncertainties and ask good questions [EMNLP’15’, CVPR’19, EMNLP’19, ICML’22].
Modeling humans and the world: I show that current large generative models like GPT-4 implement a very primitive “model of thought” [ToM@ICML’23]. To become more reliable, they need to develop robust models of the world and the humans in it, which allows them to simulate the future and plan appropriate actions. I take an inital step in this direction by improving the pragmatic-reasoning capbility of instruction-generation models [ACL’23].
More facts:
My real name is Nguyễn Xuân Khánh . My first name is usually confused with Khan or Kahn :(
I was born in Việt Nam , a peaceful country (click here for inspiration to visit us).
I am also proud to be a PTNK (Phổ Thông Năng Khiếu) alumnus.
New paper on task-oriented cognitive capabilities. TLDR; we found and improved the deficiency in the pragmatic capability of instruction generation models. Received outstanding paper award at the ToM workshop at ICML 2023.
Aug 17, 2022
I will be organizing InterNLP workshop at NeurIPS 2022. Please submit your papers if interested!
@inproceedings{nguyen2019hanna,author={Nguyen, Khanh and Daum{\'e} III, Hal},title={Help, Anna! Visual Navigation with Natural Multimodal Assistance via Retrospective Curiosity-Encouraging Imitation Learning},booktitle={EMNLP},month={},year={2019},}
Interactive Learning from Activity Description
Khanh Nguyen, Dipendra Misra, Robert Schapire, and 2 more authors
@inproceedings{nguyen2021iliad,title={Interactive Learning from Activity Description},author={Nguyen, Khanh and Misra, Dipendra and Schapire, Robert and Dud{\'\i}k, Miro and Shafto, Patrick},booktitle={ICML},year={2021},}
Posterior calibration and exploratory analysis for natural language processing models
@inproceedings{nguyen15calibration,title={Posterior calibration and exploratory analysis for natural language processing models},author={Nguyen, Khanh and O{'}Connor, Brendan},booktitle={EMNLP},month=sep,year={2015},address={Lisbon, Portugal},publisher={Association for Computational Linguistics},url={https://www.aclweb.org/anthology/D15-1182},doi={10.18653/v1/D15-1182},pages={1587--1598},}
Reinforcement Learning for Bandit Neural Machine Translation with Simulated Human Feedback
Khanh Nguyen, Hal Daumé III, and Jordan Boyd-Graber
Machine translation is a natural candidate problem for reinforcement learning from human feedback: users provide quick, dirty ratings on candidate translations to guide a system to improve. Yet, current neural machine translation training focuses on expensive human-generated reference translations. We describe a reinforcement learning algorithm that improves neural machine translation systems from simulated human feedback. Our algorithm combines the advantage actor-critic algorithm (Mnih et al., 2016) with the attention-based neural encoder-decoder architecture (Luong et al., 2015). This algorithm (a) is well-designed for problems with a large action space and delayed rewards, (b) effectively optimizes traditional corpus-level machine translation metrics, and (c) is robust to skewed, high-variance, granular feedback modeled after actual human behaviors.
@inproceedings{nguyen2017banditnmt,title={Reinforcement Learning for Bandit Neural Machine Translation with Simulated Human Feedback},author={Nguyen, Khanh and Daum{\'e} III, Hal and Boyd-Graber, Jordan},booktitle={EMNLP},month=sep,year={2017},address={Copenhagen, Denmark},publisher={Association for Computational Linguistics},url={https://www.aclweb.org/anthology/D17-1153},doi={10.18653/v1/D17-1153},pages={1464--1474},}
A Framework for Learning to Request Rich and Contextually Useful Information from Humans
@inproceedings{nguyen2022hari,author={Nguyen, Khanh and Bisk, Yonatan and Daum{\'e} III, Hal},title={A Framework for Learning to Request Rich and Contextually Useful Information from Humans},booktitle={ICML},month=jul,year={2022},}