About

Huaiyang Zhong is an Assistant Professor in the Grado Department of Industrial and Systems Engineering at Virginia Tech. He holds a Ph.D. and M.S. in Management Science and Engineering from Stanford University, and a B.S. in Industrial Engineering from the National University of Singapore. His research areas include Data Analytics, Optimization, and Empirical Operations Management. Zhong has contributed to the development of models and decision-making tools in healthcare, public transportation, and agricultural systems, with notable work on vaccine rollout impacts, hepatitis C elimination strategies, and resource allocation. He has been actively involved in teaching courses such as Random Process at Virginia Tech and has served as a teaching assistant at Stanford University. Zhong has also been recognized for his research with awards including the INFORMS Decision Analysis Society Best Student Paper Award runner-up and the Syngenta Crop Challenge first prize. His work has been presented at numerous conferences and he has served as an active reviewer for leading journals in operations management and healthcare systems.

Research topics

• Political Science
• Medicine
• Psychiatry
• Internal medicine
• Virology
• Demography
• Family medicine

Selected publications

• Towards Better Statistical Understanding of Watermarking LLMs

  Figshare · 2026-01-01

  datasetOpen access

  In this paper, we study the problem of watermarking large language models (LLMs). We consider the trade-off between model distortion and detection ability and formulate it as a constrained optimization problem based on the red-green list watermarking algorithm. We show that the optimal solution to the optimization problem enjoys a nice analytical property which provides a better understanding and inspires the algorithm design for the watermarking process. We develop an online dual gradient ascent watermarking algorithm in light of this optimization formulation and prove its asymptotic Pareto optimality between model distortion and detection ability. Such a result guarantees an averaged increased green list probability and henceforth detection ability explicitly (in contrast to previous results). Moreover, we provide a systematic discussion on the choice of the model distortion metrics for the watermarking problem. We justify our choice of KL divergence and present issues with the existing criteria of “distortion-free” and perplexity. Finally, we empirically evaluate our algorithms on extensive datasets against benchmark algorithms.

  PublisherDOI

• Towards Better Statistical Understanding of Watermarking LLMs

  Journal of the American Statistical Association · 2026-01-30

  articleOpen access
  PublisherDOI

• The Enhanced Physics-Informed Kolmogorov-Arnold Networks: Applications of Newton's Laws in Financial Deep Reinforcement Learning (RL) Algorithms

  arXiv (Cornell University) · 2026-02-01

  articleOpen accessSenior author

  Deep Reinforcement Learning (DRL), a subset of machine learning focused on sequential decision-making, has emerged as a powerful approach for tackling financial trading problems. In finance, DRL is commonly used either to generate discrete trade signals or to determine continuous portfolio allocations. In this work, we propose a novel reinforcement learning framework for portfolio optimization that incorporates Physics-Informed Kolmogorov-Arnold Networks (PIKANs) into several DRL algorithms. The approach replaces conventional multilayer perceptrons with Kolmogorov-Arnold Networks (KANs) in both actor and critic components-utilizing learnable B-spline univariate functions to achieve parameter-efficient and more interpretable function approximation. During actor updates, we introduce a physics-informed regularization loss that promotes second-order temporal consistency between observed return dynamics and the action-induced portfolio adjustments. The proposed framework is evaluated across three equity markets-China, Vietnam, and the United States, covering both emerging and developed economies. Across all three markets, PIKAN-based agents consistently deliver higher cumulative and annualized returns, superior Sharpe and Calmar ratios, and more favorable drawdown characteristics compared to both standard DRL baselines and classical online portfolio-selection methods. This yields more stable training, higher Sharpe ratios, and superior performance compared to traditional DRL counterparts. The approach is particularly valuable in highly dynamic and noisy financial markets, where conventional DRL often suffers from instability and poor generalization.

  PublisherOA PDF

• Towards Better Statistical Understanding of Watermarking LLMs

  Journal of the American Statistical Association · 2026-01-02 · 1 citations

  preprintOpen access

  In this paper, we study the problem of watermarking large language models (LLMs). We consider the trade-off between model distortion and detection ability and formulate it as a constrained optimization problem based on the red-green list watermarking algorithm. We show that the optimal solution to the optimization problem enjoys a nice analytical property which provides a better understanding and inspires the algorithm design for the watermarking process. We develop an online dual gradient ascent watermarking algorithm in light of this optimization formulation and prove its asymptotic Pareto optimality between model distortion and detection ability. Such a result guarantees an averaged increased green list probability and henceforth detection ability explicitly (in contrast to previous results). Moreover, we provide a systematic discussion on the choice of the model distortion metrics for the watermarking problem. We justify our choice of KL divergence and present issues with the existing criteria of ``distortion-free'' and perplexity. Finally, we empirically evaluate our algorithms on extensive datasets against benchmark algorithms.

  PublisherDOI

• Exploring Acylcarnitine Metabolism Using Reverse Metabolomics

  ChemRxiv · 2026-02-25 · 1 citations

  articleOpen access

  Untargeted mass spectrometry (MS) is a valuable tool for studying human metabolism and identifying small molecule disease biomarkers. However, annotation of chemical structures and validation of findings across numerous cohorts remains challenging. Reverse metabolomics employs a structure-driven approach to overcome these issues by searching spectra of known structures against an entire repository of untargeted LC-MS/MS data to see where metabolites of interest are found. This work uses reverse metabolomics to study acylcarnitine (AC) metabolism in humans and other animals. Here, a library of 76 ACs was chemically synthesized then searched against public metabolomics data to explore where metabolites of interest are detected. From this analysis, it was determined that acylcarnitines are most frequently observed in human and mouse samples, with about 90% of all searched AC structures present in both blood and fecal samples from these species. This work identified positive associations between certain AC structures and disease, indicating their capacity as health biomarkers. Machine learning was applied, determining that AC presence and absence data can accurately predict healthy versus unhealthy individuals with good precision and recall, albeit the models lack disease specificity. Overall, our findings suggest that AC profiles can serve as valuable biomarkers for disease detection throughout the entire lifespan and should be examined for their potential beyond current clinical screening protocols.

  PublisherDOI

• The Enhanced Physics-Informed Kolmogorov-Arnold Networks: Applications of Newton's Laws in Financial Deep Reinforcement Learning (RL) Algorithms

  Open MIND · 2026-02-01

  preprintSenior author

  Deep Reinforcement Learning (DRL), a subset of machine learning focused on sequential decision-making, has emerged as a powerful approach for tackling financial trading problems. In finance, DRL is commonly used either to generate discrete trade signals or to determine continuous portfolio allocations. In this work, we propose a novel reinforcement learning framework for portfolio optimization that incorporates Physics-Informed Kolmogorov-Arnold Networks (PIKANs) into several DRL algorithms. The approach replaces conventional multilayer perceptrons with Kolmogorov-Arnold Networks (KANs) in both actor and critic components-utilizing learnable B-spline univariate functions to achieve parameter-efficient and more interpretable function approximation. During actor updates, we introduce a physics-informed regularization loss that promotes second-order temporal consistency between observed return dynamics and the action-induced portfolio adjustments. The proposed framework is evaluated across three equity markets-China, Vietnam, and the United States, covering both emerging and developed economies. Across all three markets, PIKAN-based agents consistently deliver higher cumulative and annualized returns, superior Sharpe and Calmar ratios, and more favorable drawdown characteristics compared to both standard DRL baselines and classical online portfolio-selection methods. This yields more stable training, higher Sharpe ratios, and superior performance compared to traditional DRL counterparts. The approach is particularly valuable in highly dynamic and noisy financial markets, where conventional DRL often suffers from instability and poor generalization.

  DOI

• Towards Better Statistical Understanding of Watermarking LLMs

  Figshare · 2026-01-01

  datasetOpen access

  In this paper, we study the problem of watermarking large language models (LLMs). We consider the trade-off between model distortion and detection ability and formulate it as a constrained optimization problem based on the red-green list watermarking algorithm. We show that the optimal solution to the optimization problem enjoys a nice analytical property which provides a better understanding and inspires the algorithm design for the watermarking process. We develop an online dual gradient ascent watermarking algorithm in light of this optimization formulation and prove its asymptotic Pareto optimality between model distortion and detection ability. Such a result guarantees an averaged increased green list probability and henceforth detection ability explicitly (in contrast to previous results). Moreover, we provide a systematic discussion on the choice of the model distortion metrics for the watermarking problem. We justify our choice of KL divergence and present issues with the existing criteria of “distortion-free” and perplexity. Finally, we empirically evaluate our algorithms on extensive datasets against benchmark algorithms.

  PublisherDOI

• Comparative Risk of Developing Interstitial Cystitis With Childhood Gastrointestinal, Urological, Autoimmune, or Psychiatric Disorders

  Neurourology and Urodynamics · 2025-06-25 · 2 citations

  articleOpen accessSenior authorCorresponding

  AIMS: Interstitial cystitis (IC) is a chronic urological condition associated with significant discomfort, posing diagnostic and therapeutic challenges. Although its etiology remains unclear, early-life conditions such as gastrointestinal (GI) disorders, urological anomalies (UA), psychiatric disorders (PD), and autoimmune diseases (AD) have been hypothesized as potential risk factors for developing IC in adulthood. This study aims to investigate these associations by conducting a retrospective cohort analysis utilizing data from the TriNetX US Collaborative Network, encompassing over 118 million patient records. METHODS: The study and control groups were established across four categories of childhood disorders, with IC incidence monitored over a 14-year period. Statistical methodologies, including propensity score matching and Kaplan-Meier survival analysis, were employed to compare outcomes between cohorts. RESULTS: Findings indicate that childhood GI and UA conditions significantly elevate the risk of IC in adulthood, with irritable bowel syndrome (IBS) and urinary tract infections (UTIs) exhibiting risk ratios of 2.9 and 3.2, respectively. Gender disparities were also noted, with females exhibiting higher incidences of diseases included, particularly UA and AD during adolescence. Additionally, individuals with these early-life conditions demonstrated a higher prevalence of comorbidities, underscoring the complex interplay of health factors contributing to IC pathogenesis. CONCLUSIONS: These findings suggest that childhood GI and UA conditions may serve as predictive markers for IC, emphasizing the need for targeted early interventions and preventative care strategies. By identifying at-risk populations, this study provides valuable insights into early detection and management approaches, potentially mitigating the long-term burden of IC on affected individuals. TRIAL REGISTRATION: This paper includes an observational retrospective study. No clinical trial has been conducted.

  PublisherOA PDFDOI

• Wheels on the Bus: Impact of Vaccine Rollouts on Demand for Public Transportation

  Production and Operations Management · 2025-09-01

  article1st authorCorresponding

  The coronavirus disease 2019 (COVID-19) pandemic led to sharp declines in public transit ridership, resulting in budget shortfalls and service cuts that disproportionately affect vulnerable riders who lack alternative transportation options. Understanding the effect of vaccination on public transportation demand is important for planning for demand recovery as vaccination efforts progress. This study examines this effect, with a particular focus on heterogeneous effects across population groups. Estimating the impact of vaccination is challenging due to a lack of fine-grained data and potential endogeneity issues. To overcome these hurdles, we exploit the distinctive features of the COVID-19 vaccination progress to identify an instrumental variable. By merging the U.S. COVID-19 vaccination data with county-level mobility data, we construct a sample that links vaccination rates to the demand for public transportation and follow the instrumental variable approach to estimate the impact. We show that higher vaccination rates led to increases in public transportation demand, as reflected in mobility data from transit stations. Furthermore, we find the effect of vaccination on public transit demand is 50% greater in counties with a larger uninsured population and 80% greater in counties with a higher share of residents without a college degree. Our findings demonstrate that as vaccination efforts progress, public transit agencies should proactively strengthen their infrastructure to accommodate the anticipated rise in ridership. A strategic, forward-looking approach will be critical to ensuring a sustainable and effective recovery of public transportation systems.

  PublisherDOI

• Understanding needs and perspectives on remote technological interventions for perinatal mood and anxiety disorders: experiences of Black or African American women (Preprint)

  2025-08-25

  preprintOpen access

  <sec> <title>BACKGROUND</title> Perinatal mood and anxiety disorders (PMADs) affect one in four women, contributing to maternal health complications and adversely affecting child development outcomes. While remote intervention systems can increase care access, lower costs, and reduce stigma, most are limited to either the pregnancy or postpartum phase—failing to provide continuous support across the entire perinatal period. Moreover, these systems are rarely designed with the specific cultural, social, and structural needs of racial and ethnic minority women, who face disproportionate barriers to mental health care. Given the evolving and complex nature of mental health challenges before and after childbirth, there is an urgent need for a remote intervention system that delivers effective, inclusive support throughout all phases of the perinatal journey. </sec> <sec> <title>OBJECTIVE</title> This study aims to: 1) identify the challenges and needs of Black or African American women experiencing PMAD; 2) explore the benefits and barriers influencing their willingness to use remote intervention systems; and 3) determine essential features for such a system. </sec> <sec> <title>METHODS</title> A mixed-methods approach was used, combining an online survey with semi-structured interviews from perinatal women who identify as Black or African American. The survey assessed mental health challenges, support preferences (in-person, remote, or both), and desired intervention features, with chi-square and ANOVA tests performed for quantitative analysis. Follow-up Zoom interviews were conducted, and responses were analyzed using thematic coding. </sec> <sec> <title>RESULTS</title> While frequency of PMAD was not significant across the perinatal phases, a majority of support was sought during the second and third trimesters, suggesting the need for targeted interventions during these phases. Participants identified four key challenges: emotional, physical, financial, systemic and social barriers to support. The preference for remote or hybrid (in-person and remote) support was high during the second (72.4%) and third trimesters (79.2%). Major benefits of remote systems included easier appointment scheduling (90.3%), reduced travel time (90.3%), and lower costs (80.6%). However, primary barriers were missing in-person interaction (64.5%), and time constraints due to family responsibilities (67.7%). Key features preferred for remote systems included: communication with healthcare providers, cognitive behavioral therapy-based counseling services, symptom monitoring for anxiety or depression, calming music, positive affirmations or motivational quotes, and educational resources. Participants favored a customizable smartphone-based system, with sessions requiring minimal time commitment between 15-60 minutes, and flexible scheduling particularly during the postpartum period. </sec> <sec> <title>CONCLUSIONS</title> Findings suggest that mental health intervention systems for Black or African American women should primarily focus on the second and third trimesters, as participants seek more support in these phases. Further the interventions should be holistic, incorporating physical health tracking, emotional well-being tools, and education tools. </sec>

  PublisherOA PDFDOI

Frequent coauthors

• Jagpreet Chhatwal

  Massachusetts General Hospital

  28 shared

• Yueran Zhuo

  Mississippi State University

  19 shared

• Madeline Adee

  19 shared

• Lindsey Hiebert

  8 shared

• Alec Aaron

  Massachusetts General Hospital

  8 shared

• Gallican N. Rwibasira

  Ministry of Health

  7 shared

• Tiannan Zhan

  Northwestern University

  7 shared

• Janvier Serumondo

  UNSW Sydney

  7 shared

Labs

• Huaiyang Zhong LabPI

Education

• PhD, Management Science and Engineering

  Virginia Tech

Awards & honors

• Runner-up, INFORMS Decision Analysis Society Best Student Pa…
• Winner of 2015 INFORMS Syngenta Crop Challenge
• Nominated for Society of Medical Decision Making Lee B. Lust…