Geyi Yang 杨戈易
Hi, I am Geyi Yang, also known as Gary. I am an undergraduate researcher in Computer Engineering at The Chinese University of Hong Kong, Shenzhen (CUHK-Shenzhen), and was a GLOBE Program visiting student in Computer Science at UC Berkeley in Spring 2026. I am currently advised by Prof. Zhongxiang Dai.
I am broadly interested in LLM-based agents, GUI agents, self-improving agentic systems, harness optimization, and retrieval-augmented generation (RAG). My previous work also covers LLM post-training and dataset construction.
Research
More recently, I have become particularly interested in self-improving agentic systems, especially how agents can turn their own execution traces, failures, and feedback into reliable improvements in future behavior. My current work, GUI-HARVEST, studies how repeated multimodal execution evidence can drive systematic, validated changes to a GUI agent's inference-time harness.
Preprints
GUI-HARVEST: Self-Improving GUI Agents through Evidence-Driven Harness Evolution
Under review at ICLR 2027 · arXiv, 2026
Selected Research
Meme-Qwen-7B-Instruct
Built the full data-to-model pipeline for a Chinese internet-meme conversational model, including 30,000+ collected pairs, an 8,680-example open SFT dataset, and LoRA plus DPO training.
Dynamic Hybrid Lunar Logistics Optimization
Led modeling and programming for an MCM 2026 Finalist team (top 1%), combining dynamic logistics simulation, Differential Evolution, Monte Carlo analysis, and sensitivity analysis.
Experience
Shenzhen Research Institute of Big Data
Research Assistant · Advisor: Prof. Zhongxiang Dai
Cloud Computing Center, Chinese Academy of Sciences
Domain-specific Agentic RAG for mining-blasting QA
Shenzhen Institute of Artificial Intelligence and Robotics for Society
Multimodal digital human synthesis for emotional interaction
Education
The Chinese University of Hong Kong, Shenzhen
B.Eng. in Computer Engineering · GPA 3.73/4.00
University of California, Berkeley
GLOBE Program Visiting Student in Computer Science
Development Projects
Voice Turing Test
A public WeChat Mini Program game and the data-collection platform for an ICLR 2026 study. I led product design and full-stack development; it reached 1,000+ users and collected 5,000+ valid records in its first month.
Full-stack development lead · 2025
Telos: AI-Generated, Ready-to-Learn Courses
A goal-conditioned education product that turns fragmented sources into structured, multimodal, ready-to-learn courses.
Independent full-stack project · 2026 – Present
TaskFlow AI: Visual Task Orchestration
A visual workspace for tasks, dependencies, deadlines, and human-in-the-loop agent workflows.
Independent full-stack project · 2026
Selected Awards
- COMAP Mathematical Contest in Modeling, Finalist (top 1%)
- China International College Students' Innovation Competition, Bronze Award
- Dean's List, CUHK-Shenzhen
- Academic Performance Scholarship, CUHK-Shenzhen
Services
- Teaching Assistant, MAT1001 Calculus I, CUHK-Shenzhen