My research bridges theory and practice by developing models of strategic interaction that better capture real-world behavior. My goal is to translate research into insights that inform real-world practice and policy. I combine tools from causal inference, economic theory, reinforcement learning, and agent-based modeling to study how boundedly rational agents learn and interact in multi-agent settings. Currently, I am interested in applying these tools to questions in labor economics and business problems.
I am currently pursuing a Ph.D. in Economics at the University of Washington, mainly advised by Prof. Rachel Heath (UW, Economics). I have also been working with people from different areas, including Prof. Natasha Jaques (UW, Computer Science and Engineering), and Prof. Yanzhen Chen (Hong Kong University of Science and Technology, Information Systems).
I am a proud first-generation college student.
My mom and I have participated in TNR (Trap–Neuter–Return) efforts and helped neuter and feed 7 stray cats in my hometown.
Research
Learn to Match: Two-Sided Matching with Temporally Extended Feedback
Paper Site. GitHub Code. First and Corresponding author. Haijing Zong, Yancheng Liang, Boyang Zhou, and Natasha Jaques. Accepted at EC’26 Online Learning and Economics, Under Review at a CS conference.
Matching markets rarely reveal fit all at once: interviews, repeated interactions, and separations generate information that shapes future decisions. We model this process as a partially observable Markov game and introduce Learn2Match, a multi-agent reinforcement-learning benchmark where decentralized agents decide whom to interview, whom to match with, and when to dissolve a match. In simulations, independent PPO delivers higher cumulative social welfare and lower cumulative regret than the bandit-style CA-ETC baseline, but its higher information-friction loss exposes weaker coordinated exploration. The benchmark points toward hybrid algorithms that combine the adaptability of reinforcement learning, the statistical discipline of bandit methods, and the structural awareness of stable matching.
Partially observable Markov games; multi-agent reinforcement learning
Two-sided AI Adoption in the Hiring Market
Now finished theoretical model development, in preparation for submission to UTD24. With Yancheng Liang and Yanzhen Chen.
AI is reshaping both sides of hiring: applicants use it to strengthen how they present themselves, while firms use it to uncover true ability. We develop a stable-matching model to capture this two-sided adoption. As AI-generated signals become more polished but less informative, firms may intensify screening, creating an arms race that raises costs without necessarily improving matches. The model identifies when this dynamic helps or hurts welfare and asks whether government verification can curb signal inflation and restore more informative hiring.
Game-theoretic modeling; equilibrium and comparative-statics analysis
Gender Price Gap in Online Platforms: Theory and Evidence
Paper link: View paper
Does letting gig workers set their own prices eliminate the gender pay gap? Using observational data from 8,210 experts on Zaihang, China’s largest personal consulting platform, we find that women charge 24.09 RMB less per hour, about 5 percent of the mean wage, even after detailed controls. Yet conditional on price, female experts deliver 0.053 standard deviations more realized quality, and a time-lagged estimate of 0.063 is statistically indistinguishable from the post-learning result. The gap’s persistence as information accumulates points to taste-based discrimination, rather than hidden quality differences or purely statistical discrimination, in this self-priced market.
Causal inference
Labor Trains Its Replacement: Drivers’ Transition to Robotaxis
Under modelling development, with Pengyu Wang (UW), Junyu Cao (UT Austin) and Kan Xu (ASU).
Robotaxi platforms face a striking transition: the human drivers they support today may help create the conditions for replacing those same workers tomorrow. By making this intertemporal trade-off explicit, the model aims to clarify when supporting human labor accelerates, rather than delays its own replacement.
Dynamic Programming; game-theoretic modeling