About Me

Building LLM agents for the messiness of the real world.

I study long-horizon decision-making under sparse, noisy, and delayed feedback. I build simulation environments and develop post-training methods to make agents more capable, reliable, and adaptive in real-world settings.

LLM AgentsPost-training & RLEvaluation & Attribution

I am pursuing a two-year M.S. in Computer Science at Yale University. I hold dual bachelor’s degrees in Data Science from the University of Michigan and Electrical and Computer Engineering from Shanghai Jiao Tong University.

Experience

Alibaba Inc.

Research Intern · AliStar Top Talent Program

May 2026 — Present

Business Arena: Benchmarking LLM Agents in a Realistic Marketplace

We build a long-horizon, realistic business world to measure whether agents can operate end-to-end businesses.

Snap Inc. & Yale University

Student Researcher

Sep 2025 — Feb 2026

FlexRec: Adapting LLM-based Recommenders for Flexible Needs via Reinforcement Learning

Real user needs are diverse, and one item can carry different value when the need changes. We introduce FlexRec, a post-training framework that aligns LLM recommenders across multiple needs.

Carnegie Mellon University

Machine Learning Research Intern

Jan 2025 — May 2025

DATE-LM: Benchmarking Data Attribution Evaluation for Large Language Models

Data attribution promises to explain how training data shapes LLM behavior, but existing evaluations are fragmented. We introduce DATE-LM, a unified benchmark for comparing attribution methods across practical LLM applications.

UIUC & USC

Machine Learning Research Intern

Jun 2024 — Dec 2024

Detecting and Filtering Unsafe Training Data via Data Attribution with Denoised Representation

Small amounts of unsafe training data can meaningfully change model behavior, while fixed moderation categories can miss emerging risks. We use data attribution to connect harmful behavior back to influential training examples and support targeted filtering.

University of Michigan & UIUC

Student Researcher

2024

dattri: A Library for Efficient Data Attribution

Data attribution methods are useful but difficult to implement and compare consistently. We introduce dattri, an open-source PyTorch library for developing, benchmarking, and deploying efficient data-attribution methods.

University of Michigan

Student Researcher

2024

Bridging AI and Science: Implications from a Large-Scale Literature Analysis of AI4Science

AI and scientific research are advancing quickly, but useful methods and real scientific needs do not always meet. We map the AI4Science literature at scale to reveal these gaps and surface opportunities for cross-disciplinary collaboration.