CV
中文版请访问(https://zuyu3.github.io/Zuyu3-cn.github.io/cv/)
Contact Information
| Name | Leiyu Wang(王雷宇) |
| Professional Title | AI Researcher |
| leiyuwang33@gmail.com | |
| Location | Shanghai, 200240 |
Professional Summary
An AI researcher interested in Embodied Intelligence(specifically, VLA) and the application of VLM/LLM.
Experience
-
2024/7 - 2024/9 Nanjing
HarmonyOS Department Intern
Huawei
Researched Lottie’s background, mainstream use cases, and the ArkTS framework logic, and produced 10+ wiki knowledge summaries, along with an end-to-end AE -> Lottie animation demo. Developed a parser based on the raw Lottie format to reduce key information extraction from tens of thousands of lines to a few hundred, laying the groundwork for delivering Lottie-like atomic animations.
- Internship performance: Excellent
Education
-
2025 - 2028 Shanghghai, China
-
2025 - 2028 Shanghghai, China
Visiting Master
Shanghai Innovation institution
Embodied Intelligence
- Scalable VLA Training leader of the Embodied Intelligence Landmark Project (led by Prof. Cewu Lu)
-
2021 - 2025 Nanjing, China
Bachelor
Nanjing University
Brain Science and Artificial Intelligence
- rank 2/20 in major
- University First-Class Scholarship
- Basic Disciplines Special Scholarship
Publications
-
2024 What makes a good order of examples in in-context learning
ACL Findings
In-context learning is highly sensitive to the order of few-shot examples, but prior ordering heuristics typically rely on additional in-domain (often unlabeled) data and still miss instance-specific differences. We analyze what makes an example order performant at both the corpus and instance levels, and propose DEmO, which adaptively selects a strong order for each test instance without extra data. DEmO filters candidate orders by label fairness and then picks the most influential order per instance using a content-free metric, achieving strong gains over competitive baselines and generalizing well across settings.