Beixi (Chloe) Du
MEng in Financial Engineering at Cornell, with a mathematics background and research spanning machine learning, portfolio optimization, and statistical modeling of markets.
BD
MEng in Financial Engineering at Cornell, with a mathematics background and research spanning machine learning, portfolio optimization, and statistical modeling of markets.
About
I am an M.Eng. student in Financial Engineering at Cornell University, where I work on machine learning, sequential decision-making, and control. My current research includes reasoning-aware evaluation for vision-language-action models, recovering LLM-agent policies from noisy human demonstrations, and belief-space control for systems with input-dependent observations. Before Cornell, I earned a B.A. in Mathematics with minors in Computer Science and Economics from McGill University. Broadly, I am interested in machine learning and sequential decision-making, with a particular focus on reinforcement learning and robot learning. I am especially interested in developing methods that improve robustness, generalization, and learning efficiency in complex real-world settings.
Education
Grade: 3.951
Selected coursework: Stochastic Calculus, Applied Time Series, Reinforcement Learning, Applied Introduction to Cryptography, Quantitative Portfolio Strategies, Trading Rates and FX, Optimization, Fixed-Income Securities and Interest-Rate Options, Monte Carlo Simulation, and Operations Research Tools.
Grade: 3.96/4.0 GRE Score: 332
Activities and societies: Mental Health Club
Graduated on the Dean’s Honour List and received the Susan C. Vaughan Scholarship (2023; $2,000).
Relevant coursework: Advanced Calculus, Real Analysis, Ordinary Differential Equations, Linear Algebra, Abstract Algebra, Probability, Statistics, Statistical Computing, Applied Machine Learning, Python, Java, Algorithms and Data Structures, Microeconomics, Macroeconomics, Numerical Analysis, Applied Regression, and Stochastic Processes.
Research & Selected Projects
Built a rolling daily backtest (2022–2023) comparing ten portfolio strategies fed the same LSTM return forecasts and factor-model covariance. Black–Litterman shrinkage of noisy ML views toward market equilibrium delivered the best risk-adjusted performance (gross Sharpe 0.67) with lower drawdown and better diversification; validated with stationary block-bootstrap significance tests and transaction-cost analysis.
Designing a human-preference-trained, reasoning-aware reward model (LoRA-adapted Qwen2-VL) that jointly scores a policy's reasoning trace and action trajectory as a label-efficient automatic evaluator, with Monte-Carlo value estimates for dense per-step credit assignment and Wilson-interval significance testing.
A two-stage pipeline combining graph-based policy extraction with cross-trajectory statistical gap detection; improved agent pass@1 from 67.5% to 90.0% (vs. a 95.0% high-quality ceiling) and lifted pass@3 consistency from 42.5% to 55.0%.
Co-authored a CDC 2026 submission on dual control where the separation principle fails; implemented and benchmarked B-MPC / Sep / Sep-MPC controllers with an input-dependent Kalman filter and L-BFGS optimization.
Authors: Shuo Wen, Beixi Du, Edwin Meriaux, Junming Shi, Chloe Si, David Meger, Doina Precup, and Gregory Dudek.
Introduces a video-understanding benchmark of professional magic performances for evaluating VLM cognitive visual reasoning. The benchmark contains 180 videos and 2,177 open-ended questions across contextual, paradoxical, instant, temporal-summarization, and adversarial false-premise categories.
Manuscripts: Dual Control of Linear Systems from Bilinear Observations (CDC 2026, co-first author).
Experience
Skills
Résumé
Education, research, and quantitative-finance experience in a single PDF.