Huy Nghiem
Research Subjects
Huy Nghiem has a total of 27 co-authors. Here are the top co-authors by number of shared publications:
- Fred Morstatter
University of Southern California Information Sciences Institute University of Southern California Information Sciences Institute
2 shared publications - Hal Daumé Iii
Microsoft Research γ University of Maryland
2 shared publications - John Prindle
University of Southern California
2 shared publications Phuong-Anh Nguyen-Le
University of Maryland
2 shared publications- Rachel Rudinger
University of Maryland College Park USA
2 shared publications - Umang Gupta
USC Information Sciences Institute
2 shared publications
Latest Publications
- Open Access
Balancing Safety and Helpfulness in Healthcare AI Assistants through Iterative Preference Alignment
Large Language Models (LLMs) are increasingly used in healthcare, yet ensuring their safety and trustworthiness remains a barrier to deployment. Conversational medical assistants must avoid unsafe compliance without over-refusing benign queries. We present an iterative post-deployment alignment framework that applies Kahneman-Tversky Optimization (KTO) and Direct Preference Optimization (DPO) to refine models against domain-specific safety signals. Using the CARES-18K benchmark for adversarial robustness, we evaluate four LLMs (Llama-3B/8B, Meditron-8B, Mistral-7B) across multiple cycles. Our results show up to 42% improvement in safety-related metrics for harmful query detection, alongside interesting trade-offs against erroneous refusals, thereby exposing architecture-dependent calibration biases. We also perform ablation studies to identify when self-evaluation is reliable and when external or finetuned judges are necessary to maximize performance gains. Our findings underscore the importance of adopting best practices that balance patient safety, user trust, and clinical utility in the design of conversational medical assistants.
Open AccessBalancing Safety and Helpfulness in Healthcare AI Assistants through Iterative Preference Alignment
Large Language Models (LLMs) are increasingly used in healthcare, yet ensuring their safety and trustworthiness remains a barrier to deployment. Conversational medical assistants must avoid unsafe compliance without over-refusing benign queries. We present an iterative post-deployment alignment framework that applies Kahneman-Tversky Optimization (KTO) and Direct Preference Optimization (DPO) to refine models against domain-specific safety signals. Using the CARES-18K benchmark for adversarial robustness, we evaluate four LLMs (Llama-3B/8B, Meditron-8B, Mistral-7B) across multiple cycles. Our results show up to 42% improvement in safety-related metrics for harmful query detection, alongside interesting trade-offs against erroneous refusals, thereby exposing architecture-dependent calibration biases. We also perform ablation studies to identify when self-evaluation is reliable and when external or finetuned judges are necessary to maximize performance gains. Our findings underscore the importance of adopting best practices that balance patient safety, user trust, and clinical utility in the design of conversational medical assistants.
- Open Access
'Rich Dad, Poor Lad': How do Large Language Models Contextualize Socioeconomic Factors in College Admission ?
Large Language Models (LLMs) are increasingly involved in high-stakes domains, yet how they reason about socially sensitive decisions remains underexplored. We present a large-scale audit of LLMs' treatment of socioeconomic status (SES) in college admissions decisions using a novel dual-process framework inspired by cognitive science. Leveraging a synthetic dataset of 30,000 applicant profiles grounded in real-world correlations, we prompt 4 open-source LLMs (Qwen 2, Mistral v0.3, Gemma 2, Llama 3.1) under 2 modes: a fast, decision-only setup (System 1) and a slower, explanation-based setup (System 2). Results from 5 million prompts reveal that LLMs consistently favor low-SES applicants -- even when controlling for academic performance -- and that System 2 amplifies this tendency by explicitly invoking SES as compensatory justification, highlighting both their potential and volatility as decision-makers. We then propose DPAF, a dual-process audit framework to probe LLMs' reasoning behaviors in sensitive applications.
Open Access'Rich Dad, Poor Lad': How do Large Language Models Contextualize Socioeconomic Factors in College Admission ?
Large Language Models (LLMs) are increasingly involved in high-stakes domains, yet how they reason about socially sensitive decisions remains underexplored. We present a large-scale audit of LLMs' treatment of socioeconomic status (SES) in college admissions decisions using a novel dual-process framework inspired by cognitive science. Leveraging a synthetic dataset of 30,000 applicant profiles grounded in real-world correlations, we prompt 4 open-source LLMs (Qwen 2, Mistral v0.3, Gemma 2, Llama 3.1) under 2 modes: a fast, decision-only setup (System 1) and a slower, explanation-based setup (System 2). Results from 5 million prompts reveal that LLMs consistently favor low-SES applicants -- even when controlling for academic performance -- and that System 2 amplifies this tendency by explicitly invoking SES as compensatory justification, highlighting both their potential and volatility as decision-makers. We then propose DPAF, a dual-process audit framework to probe LLMs' reasoning behaviors in sensitive applications.
- Open Access
- 1 January 2025
- Published by Association for Computational Linguistics (ACL)
- p. 21033-21067
‘Rich Dad, Poor Lad’: How do Large Language Models Contextualize Socioeconomic Factors in College Admission ?
Huy Nghiem, Phuong-Anh Nguyen-Le, John Prindle, Rachel Rudinger, Hal Daumé Iii. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.
- 1 January 2025
- Published by Association for Computational Linguistics (ACL)
- p. 21033-21067
Open Access‘Rich Dad, Poor Lad’: How do Large Language Models Contextualize Socioeconomic Factors in College Admission ?
Huy Nghiem, Phuong-Anh Nguyen-Le, John Prindle, Rachel Rudinger, Hal Daumé Iii. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.
- Open Access
"Define Your Terms" : Enhancing Efficient Offensive Speech Classification with Definition
The propagation of offensive content through social media channels has garnered attention of the research community. Multiple works have proposed various semantically related yet subtle distinct categories of offensive speech. In this work, we explore meta-earning approaches to leverage the diversity of offensive speech corpora to enhance their reliable and efficient detection. We propose a joint embedding architecture that incorporates the input's label and definition for classification via Prototypical Network. Our model achieves at least 75% of the maximal F1-score while using less than 10% of the available training data across 4 datasets. Our experimental findings also provide a case study of training strategies valuable to combat resource scarcity.
Open Access"Define Your Terms" : Enhancing Efficient Offensive Speech Classification with Definition
The propagation of offensive content through social media channels has garnered attention of the research community. Multiple works have proposed various semantically related yet subtle distinct categories of offensive speech. In this work, we explore meta-earning approaches to leverage the diversity of offensive speech corpora to enhance their reliable and efficient detection. We propose a joint embedding architecture that incorporates the input's label and definition for classification via Prototypical Network. Our model achieves at least 75% of the maximal F1-score while using less than 10% of the available training data across 4 datasets. Our experimental findings also provide a case study of training strategies valuable to combat resource scarcity.
- Open Access
- 1 January 2024
- Published by Association for Computational Linguistics (ACL)
- p. 1293-1309
“Define Your Terms” : Enhancing Efficient Offensive Speech Classification with Definition
Huy Nghiem, Umang Gupta, Fred Morstatter. Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.
- 1 January 2024
- Published by Association for Computational Linguistics (ACL)
- p. 1293-1309
Open Access“Define Your Terms” : Enhancing Efficient Offensive Speech Classification with Definition
Huy Nghiem, Umang Gupta, Fred Morstatter. Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.
- Open Access
- 1 January 2024
- Published by Association for Computational Linguistics (ACL)
- p. 5938-5956
HateCOT: An Explanation-Enhanced Dataset for Generalizable Offensive Speech Detection via Large Language Models
Huy Nghiem, Hal Daumé Iii. Findings of the Association for Computational Linguistics: EMNLP 2024. 2024.
- 1 January 2024
- Published by Association for Computational Linguistics (ACL)
- p. 5938-5956
Open AccessHateCOT: An Explanation-Enhanced Dataset for Generalizable Offensive Speech Detection via Large Language Models
Huy Nghiem, Hal Daumé Iii. Findings of the Association for Computational Linguistics: EMNLP 2024. 2024.
- Open Access
- 27 November 2023
- Published by Wolters Kluwer Health in Epidemiology
- Vol. 35 (1) , 51-59
A Decade of Tweets: Visualizing Racial Sentiments Towards Minoritized Groups in the United States Between 2011 and 2021
Background: Research has demonstrated the negative impact of racism on health, yet the measurement of racial sentiment remains challenging. This article provides practical guidance on using social media data for measuring public sentiment. Methods: We describe the main steps of such research, including data collection, data cleaning, binary sentiment analysis, and visualization of findings. We randomly sampled 55,844,310 publicly available tweets from 1 January 2011 to 31 December 2021 using Twitter’s Application Programming Interface. We restricted analyses to US tweets in English using one or more 90 race-related keywords. We used a Support Vector Machine, a supervised machine learning model, for sentiment analysis. Results: The proportion of tweets referencing racially minoritized groups that were negative increased at the county, state, and national levels, with a 16.5% increase at the national level from 2011 to 2021. Tweets referencing Black and Middle Eastern people consistently had the highest proportion of negative sentiment compared with all other groups. Stratifying temporal trends by racial and ethnic groups revealed unique patterns reflecting historical events specific to each group, such as the killing of George Floyd regarding sentiment of posts referencing Black people, discussions of the border crisis near the 2018 midterm elections and anti-Latinx sentiment, and the emergence of COVID-19 and anti-Asian sentiment. Conclusions: This study demonstrates the utility of social media data as a quantitative means to measure racial sentiment over time and place. This approach can be extended to a range of public health topics to investigate how changes in social and cultural norms impact behaviors and policy. A supplemental digital video is available at http://links.lww.com/EDE/C91
- 27 November 2023
- Published by Wolters Kluwer Health in Epidemiology
- Vol. 35 (1) , 51-59
Open AccessA Decade of Tweets: Visualizing Racial Sentiments Towards Minoritized Groups in the United States Between 2011 and 2021
Background: Research has demonstrated the negative impact of racism on health, yet the measurement of racial sentiment remains challenging. This article provides practical guidance on using social media data for measuring public sentiment. Methods: We describe the main steps of such research, including data collection, data cleaning, binary sentiment analysis, and visualization of findings. We randomly sampled 55,844,310 publicly available tweets from 1 January 2011 to 31 December 2021 using Twitter’s Application Programming Interface. We restricted analyses to US tweets in English using one or more 90 race-related keywords. We used a Support Vector Machine, a supervised machine learning model, for sentiment analysis. Results: The proportion of tweets referencing racially minoritized groups that were negative increased at the county, state, and national levels, with a 16.5% increase at the national level from 2011 to 2021. Tweets referencing Black and Middle Eastern people consistently had the highest proportion of negative sentiment compared with all other groups. Stratifying temporal trends by racial and ethnic groups revealed unique patterns reflecting historical events specific to each group, such as the killing of George Floyd regarding sentiment of posts referencing Black people, discussions of the border crisis near the 2018 midterm elections and anti-Latinx sentiment, and the emergence of COVID-19 and anti-Asian sentiment. Conclusions: This study demonstrates the utility of social media data as a quantitative means to measure racial sentiment over time and place. This approach can be extended to a range of public health topics to investigate how changes in social and cultural norms impact behaviors and policy. A supplemental digital video is available at http://links.lww.com/EDE/C91