Portrait of Xiangyu Zhou

Seeking a Summer 2027 internship. I'd be glad to talk about user simulation, human-centered AI, or programming tools. Get in touch.

I work at the intersection of HCI, AI, and programming languages. My research builds computational models of people: what they intend, and how they act. Today I build user simulators that evaluate interfaces automatically, generating realistic human interaction on a design before anyone runs a user study. Before that, I built no-code programming systems that let non-programmers specify tasks through demonstrations and natural language.

I received my B.S. in Computer Engineering from the University of Washington, where I worked with Chenglong Wang and Rastislav Bodik.

News

Show older news
  • Apr 2025Started a new line of work on computational user simulation for interface evaluation.
  • Jan 2024Presented Arborist, our bottom-up synthesizer for web automation, at POPL 2024 in London.
  • Sep 2022Started my Ph.D. at the University of Michigan.
  • Jun 2022Our PLDI 2022 paper on Sickle received a Distinguished Paper Award.

Research

One question runs through my work: what will a person actually do? I started by inferring what people want from a few demonstrations, so they could program without code. Now I predict how people move through an interface, so designers can find usability problems before a study. In both, I turn human behavior into a model a computer can search, optimize, or simulate.

  1. Now · 2025 to present

    Simulating users to evaluate interfaces

    Can we see where people will struggle with a design before recruiting anyone?

    Usability problems usually surface late, in user studies that are slow and costly to repeat. I build computational user models in an agentic framework that generate synthetic human interaction on a given interface. The models are grounded in human-factors literature and our own user studies, and validated against human data.

    CHI 2026 · Poster

    Simulating Human Cursor Trajectories for Path-Sensitive GUI Evaluation

    Xiangyu Zhou, Steve Oney

    Some interactions only work if the cursor stays on a path. A cascading menu closes if you slip off it, and a lasso selection fails if you cut a corner. Earlier models predict only completion time for these tasks. Our simulator produces full cursor trajectories by treating movement as receding-horizon optimization (Model Predictive Contouring Control) that balances speed, smoothness, and staying in bounds. Its paths closely match human data on tunnels, cascading menus, and lasso selection.

    Under review

    Pace by Gaze: A Gaze-Informed Simulator of Constrained Cursor Movement

    Xiangyu Zhou, Beichen Xue, Yue Jiang, Steve Oney

    People look ahead, and the hand follows. This closed-loop simulator models both: regression models fit to data from our user studies predict the eye-gaze horizon, and a Model Predictive Control policy drives the arm and hand to get there. A personalization layer fits the simulator to individual users and personas from short calibration data, recovering per-user behavioral parameters.

  2. How I got here. Building tools that learn from what users show them taught me how much a system depends on a good model of the person on the other side. That led me from inferring intent to simulating behavior.

  3. Earlier · 2020 to 2025

    No-code programming by demonstration

    Can people who don't code still get a computer to do what they mean?

    Many people need to automate a web workflow or analyze a table but cannot write the program. I built systems that let them specify tasks by example, combining structured demonstrations, program synthesis, and LLMs. I developed search algorithms fast enough for interactive use and implemented the stack end to end, from DSL design to interpreters and synthesis engines in Python and Rust.

    PLDI 2022 Distinguished Paper Award

    Synthesizing Analytical SQL Queries from Computation Demonstration

    Xiangyu Zhou, Rastislav Bodik, Alvin Cheung, Chenglong Wang

    Analytical SQL, with grouping, partitioning, and window functions, is hard for novices, and giving full input-output examples is tedious. In Sickle, users instead demonstrate how a few output cells are computed by dragging cells from the input table. A new abstraction, abstract data provenance, prunes the search. Sickle solves 76 of 80 real-world tasks and is 22.5× faster on average than prior approaches.

    POPL 2024

    Efficient Bottom-Up Synthesis for Programs with Local Variables

    Xiang Li*, Xiangyu Zhou*, Rui Dong, Yihong Zhang, Xinyu Wang

    Web automation programs loop over data and bind local variables, which earlier bottom-up synthesizers could not evaluate or prune. Our key idea, lifted interpretation, evaluates all programs from a grammar at once and enumerates every binding context for local variables. Our tool, Arborist, solves 93.9% of a challenging web automation benchmark, beating the prior state of the art.

Publications

* equal contribution · Full list on Google Scholar

Teaching & Service

Teaching
Graduate Student Instructor, University of Michigan: EECS 481 Software Engineering (2 semesters), EECS 183 Elementary Programming Concepts (3 semesters)
Service
Artifact Evaluation Committee, PLDI '25, POPL '25, OOPSLA '25 · Reviewer, CHI '26 Extended Abstracts