PhD student · MIT IDSS
I am a PhD student at MIT IDSS with Prof. Jinhua Zhao. I also work with Prof. Paul Liang.
Before relocating to Cambridge, I spent two years in Hong Kong and received my M.Phil in Operations Research from the Chinese University of Hong Kong (CUHK) under Prof. Anthony Man-Cho So and Prof. Viet Anh Nguyen. I obtained my bachelor’s degree in Mathematics (2019-2023) from Tongji University.
I now study how LLMs acquire and internalize knowledge as they improve themselves.
I’ll be at the INFORMS Annual Meeting in San Francisco in November and NeurIPS 2026 in December. Come say hi!
FrontierCS 2.0 is out!
We released Evolving in Thought Space: a small guidance model learns at test time to steer a frozen, larger executor toward better discoveries. Great collaboration with Ao Qu! Check out the post on X and code. Also see my blog on recent trends in autoresearch.
FrontierOR was accepted to NeurIPS 2026, Evaluations & Datasets Track.
Our paper Optimized Risk-Budget Allocation for Safe Approximations of Joint Chance Constraints is out! Great collaboration with Zhekai!
I received the MIT Presidential Fellowship.
I presented our work on Time Series Imputation at the INFORMS TSL Conference at MIT in Cambridge.
CORAL, our work on autonomous multi-agent evolution for open-ended discovery, was accepted to COLM 2026.
I presented LLMs for Operations Research: Benchmarking and Efficient Algorithms at the POMS Annual Conference in Reno.
I presented our Goods Markets paper at the INFORMS Optimization Society Conference in Atlanta.
AlphaOPT was accepted to KDD 2026 and selected for an oral presentation.
I presented our Goods Markets paper at WINE 2025 in New Brunswick.
Our Goods Markets paper is on arXiv, with a short version accepted to WINE 2025. We study accelerated price adjustment and exact equilibrium recovery.
AlphaOPT was named a Best Paper Award finalist by the INFORMS Data Mining Society.
* Equal contribution.
NeurIPS 2026 TTCL Workshop · Short version
A small guidance model learns strategic changes at test time, while a frozen, larger model turns them into executable solutions.
COLM 2026
Autonomous agents explore, reflect, and collaborate through shared persistent memory to advance open-ended discovery.
NAACL 2025
A probe-free, low-rank intervention in hidden activations that steers language models toward more truthful generations.
Improving optimization modeling
KDD 2026 · Oral
INFORMS Data Mining Society 2025 · Best Paper Finalist
A self-improving experience library strengthens LLM optimization modeling by learning reusable knowledge from solver feedback.
Beyond modeling · Algorithm efficiency
NeurIPS 2026
Moving beyond model formulation, FrontierOR evaluates whether LLMs can design efficient algorithms for large-scale optimization across 180 tasks.
WINE 2025 · Short version
Accelerated price adjustment finds approximate equilibrium prices, and a recovery oracle obtains exact competitive equilibrium in linear and quasi-linear Fisher markets.
WINE 2024 · Short version; Games and Economic Behavior · Major revision
A difference-of-convex formulation yields linear convergence to exact chores equilibrium and a smoothed, subproblem-free method for approximate equilibrium.
Networks · Uncertainty calibration
ICLR 2025 · Spotlight
An ensemble temperature-scaling method that combines multiple inputs and calibration experts to improve uncertainty calibration in graph neural networks.
Networks · Uncertainty allocation
Working paper · 2026
Jointly optimizes decisions and individual risk budgets to reduce conservatism while preserving the safety guarantees of joint chance-constraint approximations.
Time series · Missing data
TSL 2026 · Short version
Combines Hankel lifting and robust regularization for interpretable reconstruction of incomplete time series.
Working paper · Submitted
Working paper · To be submitted
Working paper · To be submitted