Undergraduate AI Researcher | World Models & VLN

Seattle, WA · Email · Personal website · Download CV (PDF)

I'm an Informatics undergraduate at the University of Washington, working on embodied AI, world models, and vision-and-language navigation. My research includes human-aware navigation, multimodal speculative decoding, mathematical tool agents, and robot manipulation.

• University of Washington, Seattle — B.S. in Informatics · GPA: 3.73/4.0 · Expected graduation: 2027
• The Chinese University of Hong Kong — Exchange student · Fall 2026

Undergraduate Research Assistant · PI: Zhi-Qi Cheng
Apr 2026–Present · Seattle, WA · Lab

Human-Aware Vision-Language Navigation
Ongoing · Planned 2026 submission · Expected second author

• Built a low-cost batch 3D motion-retargeting agent with Kimodo and Virtual Community for new VLN scenes.
• Automated avatar retargeting and motion-data preparation, including auditing and repairing 300 human-motion samples.
• Train and evaluate VLN models with DAgger-based imitation learning; maintain experiment logs and navigation demos.

Multimodal Speculative Decoding Improvement
Ongoing · Planned 2026 submission · Expected co-first author

• Surveyed multimodal speculative-decoding and sampling methods to identify limitations in the ViSpec baseline.
• Formulated an objective linking visual-compression quality with draft-target model behavior.
• Proposed autoencoder-based supervision for the visual adapter to measure and reduce information loss.

Undergraduate Research Collaborator · Advisor: Ning Yang
Feb 2026–Present · Remote

MATH-LWM-AGENT: Efficient Mathematical Tool Agent
First author · Submitted to AAAI 2027

• Conceived and led a training-free mathematical agent that predicts tool outcomes before execution.
• Designed a residual controller that turns prediction-execution mismatch into state for subsequent reasoning; implemented the runtime with Python, SymPy, and verification tools.
• Led experiment design, ablations, cost accounting, evidence audits, and manuscript preparation.
• With thinking disabled, improved HMMT by 52.13 percentage points. Improved IMO-AnswerBench by 35.85 percentage points and IMO-ProofBench by 1.22/7.
• Residual control reduced model calls per correct answer by 9.41% and joint inference cost by 10.18%.

Research Assistant · PI: Lin Shao
Aug 2026–Present · Shenzhen, China

WAM RL
Ongoing · Planned 2026 submission · Expected third author

• Lead xArm and xHand data collection, cleaning, and organization for robot-manipulation training.
• Lead real-robot deployment on xArm, xHand, and Franka platforms for WAM experiments.
• Participate in real-robot SFT and RL training for WAM, supporting iterative data-to-deployment cycles.

Open-Source Project Maintainer
Dec 2025–Jun 2026 · Seattle, WA

• Contributed code to and maintained Amy J. Ko's open-source educational programming language and IDE.
• Reviewed accessibility and multilingual typography in learner-facing interactions; proposed improvements to the editor and learning workflow.
• Turned findings into scoped implementation tasks and maintainable code changes.

Founder & Content Creator
Jan 2026–Present · China / Online

• Founded a channel about undergraduate applications and international-student life; grew its WeChat Channels audience to approximately 5,000 followers.
• Explain application and university-life topics for students and share firsthand experiences from research-lab work.

Project Lead
Sep 2025–Jun 2026 · Dongguan, China

Multilingual corporate website

• Built a Chinese-English-Vietnamese website and maintain it as an entry point for overseas clients.
• Centralized product, sustainability, production-capacity, and inquiry information; designed the page hierarchy and navigation for product discovery.

ERP management prototype · China operations · 100+ users

• Built and iteratively deployed an ERP prototype used by 100+ employees across the company's China operations.
• Digitized the order-to-reconciliation flow across sales, procurement, quality inspection, warehouse dispatch, and finance.
• Standardized workflows and SOPs with clear ownership and handoffs, then used live operations to identify management gaps and refine controls.

• AI & machine learning: Machine learning, LLM/VLM inference, specialized AI-agent design for research workflows
• Programming & tools: Python, Java, R, LaTeX
• Mathematics & statistics: Calculus, linear algebra, vector geometry, probability, statistics

Scuba diving, surfing, skiing · Valid driver's license