About Me

👋 Hi! I am Zixuan (Eve) Yi, a third-year PhD student in CS at the University of Pennsylvania, advised by Ryan Marcus and Zack Ives. Before this, I earned my bachelor degree in CS from Tsinghua University.

I’m always happy to chat about research, collaborations, or interesting problems in ML and systems. Feel free to reach out! 😄

News

• ☕️ VLDB’26 (Aug 31 - Sep 4): I’ll be in Boston for VLDB’26. DM me if you want to grab coffee and chat!
• 🧋 COLM’26 (October 6 - 9): I’ll be in SF for COLM’26. Looking forward to meeting researchers working on LLM agents, agentic systems, and related areas!
• 🎉 Data Canvas was accepted to COLM’26!
• 🔦 Theory-Level Autoformalization was selected as a Spotlight at the ICML’26 Position Track.

Research Interests

I am broadly interested in building intelligent systems that are efficient, adaptive, and reliable in real-world environments. My previous works explored how machine learning can make systems more intelligent, and how systems techniques can make AI applications more efficient and reliable.

I’m interested in three broad questions:

• Learning & Adaptation: How can intelligent systems learn from experience and adapt to new tasks, environments, and feedback?
• Efficiency: How can we make learning and decision-making efficient when computation, interaction, or evaluation is expensive?
• Reliability & Control: How can we make intelligent systems easier to understand, steer, and correct when they make mistakes?

Publications

Data Canvas: A Provenance-Guided Harness for Agentic Data Engineering
COLM’26
Zixuan Yi, Yuanming Shao, Shaun Wallace, Zachary Ives, Ryan Marcus
Code and paper coming soon!

When an agent goes wrong, can we identify the responsible step and repair only what it affected?

Agent workflows are difficult to inspect and repair when their outputs come from long, opaque execution traces. Data Canvas structures agent execution into semantic operators and tracks fine-grained provenance, allowing feedback to be traced to the responsible computation, propagated to related outputs, and repaired by replaying only the affected parts of the workflow.

Theory-Level Autoformalization: From Isolated Statements to Unified Formal Knowledge Bases
ICML’26 Position Track (🔦Spotlight)
Marcus J Min, Mike He, Zhaoyu Li, Zixuan Yi, Sharad Malik, Aarti Gupta, Xujie Si, Osbert Bastani [code] [paper]

Autoformalization should build theories, not just translate isolated statements.

Most autoformalization (turning informal math into machine-checkable formal math) assumes that the surrounding definitions, notation, lemmas, and dependencies already exist. We argue for theory-level autoformalization: constructing coherent formal libraries and their dependency structure, rather than treating each statement as an independent translation problem.

LimeQO: Low-Rank Learning for Offline Query Optimization.
SIGMOD’25
Zixuan Yi, Yao Tian, Zachary G. Ives, Ryan Marcus [code][paper][poster]

Can we learn good decisions across an entire workload without exhaustively evaluating every possibility?

Learning good optimization decisions can require many expensive system executions. LimeQO treats an entire workload as a partially observed low-rank matrix, allowing information from a small number of executions to generalize across queries and guide where exploration is most useful, dramatically reducing learning overhead while avoiding regressions. Try it out at LimeQO.

The Unreasonable Effectiveness of LLMs for Query Optimization.
ML4Systems@NeurIPS’24 (🔦Spotlight)
Peter Akioyamen, Zixuan Yi, Ryan Marcus [code][paper][talk]

How much does a pretrained language model already know about the behavior of a SQL query?

Learned query optimizers often rely on specialized representations and database-specific features. LLMSteer shows that pretrained embeddings of raw SQL already contain useful signals for optimization: a lightweight classifier over these representations can steer an existing optimizer without complex model architectures or deep integration with database internals.

Experiences

Microsoft Research Intern @ MSR, 2026 Summer
Developed cost-aware strategies for identifying high-performing execution configurations under expensive and censored feedback.

Google Student Researcher @ SRG, 2024 Summer
Worked on lightweight online learning from historical execution behavior for efficient adaptation on real-world workloads.

Google STEP Intern @ Google Search, 2021 Summer
Built a human-in-the-loop analytics and evaluation pipeline combining expert feedback with NLP models.