Hihaody đź‘‹
I’m Zhuohui (Jerry) Chen, a third-year Ph.D. student at the College of Information, University of Maryland, College Park, advised by Prof. Zubin Jelveh. I study how to make large language models trustworthy enough for high-stakes decisions: secure as agents, reliable as judges and annotators, and useful as forecasters.
Before my Ph.D., I spent three years as a Research Analyst in the Central Bank Operations Division of the International Monetary Fund’s Monetary and Capital Markets Department. There I built forecasting models that central banks use, co-authored IMF working papers on tail-risk and reserve-demand forecasting, and joined 25+ technical assistance and financial sector assessment missions. My background is unusually mixed: computer science, English and translation studies, and public administration. That mix is why I care about language, measurement, and real-world use in equal parts.
Especially in LLM agents, evaluation, and AI for finance or forecasting.
Research
My current work follows three connected threads:
- LLMs for forecasting and judgment. Can LLM agents make judgmental forecasts, the kind human experts make by combining models, data, and context? I’m building an agentic framework to study this, drawing on my forecasting work at the IMF.
- Trustworthy and secure LLM agents. Agents that read tool outputs, web pages, and documents can be hijacked by prompt injection. I study detectors that use the model’s own internal signals, and alignment methods that hold up when preference data is poisoned.
- LLMs as measurement instruments. When social scientists use LLMs to code or annotate text, how far can we trust the labels? I use controlled synthetic data and human-annotation benchmarks to find where LLM coders succeed and where they fail.
I also work on multilingual and code-switched NLP, including pragmatic realism in synthetic code-switched data and catastrophic forgetting during code-switched fine-tuning. More on the Research page.
News
- 2026 Received the Do Good Accelerator Fellowship for the second year.
- Dec 2025 Finished Huber DPO, a robust DPO loss that resists poisoned preference data.
- 2025 Received the Do Good Accelerator Fellowship.
- Oct 2024 IMF handbook chapter Liquidity Forecasting—Part II: The Statistical Component published.
- Sep 2024 Started my Ph.D. at UMD with a Dean's Fellowship.
- Jul 2024 Our reserve-demand model was used in Box 7 of the IMF's 2024 U.S. Article IV report to size the Fed's reserve needs.
- May 2024 IMF Working Paper Forecasting Tail Risk via Neural Networks with Asymptotic Expansions released.
- Sep 2023 IMF Working Paper Modeling the Reserve Demand to Facilitate Central Bank Operations released and presented at the 2023 International Symposium on Forecasting.