University of Pennsylvania
M.S. in Data Science
How can we deploy AI in everyday life, safely and efficiently?
I'm a first-year M.S. student in Data Science at the University of Pennsylvania, after finishing my B.A. in Data Science at the University of Wisconsin–Madison. Since January 2025 I have been doing research at the LIME Lab at USC with Jieyu Zhao, and since August 2026 I have also been part of the KANG Lab at MBZUAI with Jian Kang. Also, I am fortunate to work closely with Linxin Song.
Questions behind my research
- Q1
How can we measure what models and agents can really do, and explain why they fail?
- Q2
When an agent acts on our behalf, how can it understand what we mean, and what its actions will cause?
- Q3
How can we produce data at scale that stays verifiable, and turn it into better agents?
LLM / VLM Evaluation
Benchmarks, failure taxonomies and verifiable judges that reveal where and why models break before they reach users.
Computer-Use Agents
Agents that operate real software on our behalf: how they fail, how to keep them safe, and how to make them reliable.
Data-Centric AI
Better data as the lever: human-crafted benchmarks, verified task synthesis, and error discovery over massive knowledge bases.
Robot Use · Exploring
Extending agents from digital worlds to the physical world: using them to control robots, and what safety and evaluation mean there.
01
Education
University of Wisconsin–Madison
B.A. in Data Science
GPA 4.0 / 4.0Dean's List
02
Experience
KANG Lab, MBZUAI
Aug 2026 – PresentResearch Assistant·Advisor: Jian Kang
- Research focus: evaluation of computer-use agents.
Beijing Lanyue Intelligence
May 2026 – Aug 2026LLM & Agent Algorithm Intern
- Browser-use evaluation. Benchmarked 12 frontier LLMs on web automation in the browser-use framework and built a 3-class, 9-subclass failure taxonomy (reasoning, execution, environment).
- Active-exploration task synthesis. LLM-written web tasks often invent controls and fields. Instead, a computer-use agent explores each site and records a capability card; only cross-page transitions it actually executed become edges of an information-flow DAG, and paths over the graph compose long-horizon, cross-site tasks. 331 sites explored, 75 tasks delivered.
- Verifiable LLM-as-a-Judge. 384 rubrics, each anchored to a verbatim span of the task and scored only against screenshots, the action trajectory and visible page fields, so every failure points to a specific step. On GPT 5.6 Sol, 77.1% of rubrics pass.
30.8 steps / task on average
57.3% task success · GPT 5.6 Sol
97.0% judge–human agreement
LIME Lab, University of Southern California
Jan 2025 – PresentResearch Assistant·Advisor: Jieyu Zhao
- Co-first author of OS-Blind, a 300-task human-authored benchmark of unintended attacks on computer-use agents, spanning 12 harm categories and 8 desktop applications.
- Second author of SEA (COLM 2025): budgeted discovery of knowledge deficiencies in closed-weight LLMs; benchmarked SEA across 8 LLMs against ACD and AutoBencher and analysed discovered errors with topic modeling.
03
Publications
* equal contributionThe Blind Spot of Agent Safety: How Benign User Instructions Expose Critical Vulnerabilities in Computer-Use Agents
We introduce OS-Blind, a benchmark where every user instruction is benign and the harm comes from the environment or the execution outcome. Most computer-use agents exceed 90% attack success rate, and safety alignment mostly fires in the first few steps and rarely re-engages. Multi-agent systems are even more vulnerable: task decomposition strips the context and hides the user's intent from the agents that act.
300 tasks · 8 apps
Claude 4.5 Sonnet: 73.0% ASR
Multi-agent system: 92.7% ASR
Featured on Hugging Face Daily Papers (Apr 15, 2026)
Used for GUI-agent safety evaluation in the UI-Venus-2 Technical Report
Discovering Knowledge Deficiencies of Language Models on Massive Knowledge Base
Stochastic Error Ascent (SEA) finds knowledge errors in closed-weight LLMs under a strict query budget by iteratively retrieving candidates semantically close to observed failures, with hierarchical retrieval and a relation DAG to prune sources.
40.7× more errors than ACD
599× / 9× lower cost per error
FAST-CAD: A Fairness-Aware Framework for Non-Contact Stroke Diagnosis
A domain-adversarial training plus Group-DRO framework that jointly enforces demographic-invariant representations and worst-group robustness for non-contact stroke diagnosis, built on a 12-subgroup multimodal dataset.