Research
A11yLTLNav: Automatic Detection of Accessibility Navigation Failures
Research Assistant, University of Utah · May – Sep 2026
Advisors: Dr. Yue Jiang and Dr. Ben Greenman
I developed A11yLTLNav, a property-based checker for accessibility navigation failures encountered by blind and low-vision users interacting with generated interfaces.
- Formalized common accessibility failures from prior user studies as Linear Temporal Logic (LTL) properties and tested them on randomly generated website interaction traces.
- Built a comparison web agent that detected accessibility failures through screen-reader feedback and keyboard-only interaction.
- Evaluated the checker on 31 generated websites; human reviewers confirmed 274 of 309 reports (88.7% precision).
RELIEF: Revision-Based Learning from Interactive Experience and Feedback
Research Assistant, Purdue University · May – Oct 2026
Advisor: Dr. Jason Wu
I investigated how user interaction traces provided implicit feedback for improving the usability of generated interfaces. This work contributed to RELIEF, which used feedback-guided interface revisions to construct supervision for fine-tuning code generation models.
- Built a pipeline to generate websites, collect task interaction traces in a user study, diagnose usability problems, and automatically repair websites using the resulting critiques.
- Explored fine-tuning UI generation models using websites revised with usability critiques. RELIEF reconstructed the corresponding generation traces for supervised fine-tuning.
- Explored computer-use agents and coding agents as baselines for simulating user behavior and providing usability feedback.
Paper (PDF) · Publication Details
WebCoEvo: Adversarial Co-Evolution for Web Agents
Research Assistant, University of Michigan · Mar – Jun 2026
Advisor: Prof. Honglak Lee
I developed WebCoEvo, an adversarial co-evolution framework that helped web agents adapt to changing websites by learning generalizable reflection rules from their failures.
- Paired a coding-agent-driven website version generator with a web agent to study adaptation to evolving interfaces.
- Extracted generalizable meta-level reflection rules from agent failure modes, improving performance by 19.6% over an ExpeL baseline.
- Worked on a knowledge-graph pipeline to automatically identify tasks affected by website version changes.
- Containerized multiple website versions with Docker using BrowserGym and AgentLab.