I'm a master's student in Information Technology Strategy (MITS) at Carnegie Mellon University, expected to graduate in December 2027. I received my B.S. in Computer Science from the University of Science and Technology of China (USTC) in June 2026, and spent a visiting semester at UC Berkeley in spring 2025.
我目前在卡内基梅隆大学攻读信息技术战略硕士（MITS），预计 2027 年 12 月毕业。2026 年 6 月本科毕业于中国科学技术大学计算机科学专业，2025 年春季曾在加州大学伯克利分校访学一学期。
My interests include applied machine learning, embodied AI and autonomous driving, and robust AI systems. I'm particularly interested in understanding how models behave in real-world conditions and building systems that work reliably beyond their training environments.
我的兴趣包括应用机器学习、具身智能与自动驾驶，以及鲁棒的 AI 系统。我特别关注模型在真实世界条件下的表现，以及如何构建在训练环境之外依然可靠运行的系统。
I'm looking for Summer 2027 internships in machine learning and software engineering.
我正在寻找 2027 年暑期机器学习或软件工程方向的实习。
Experience经历
- Lenovo联想 Built the data-to-deployment pipeline for vision-language-action (VLA) policies on dual-arm robots, from VR teleoperation data and model fine-tuning to on-robot ROS 2 inference. 为双臂机器人搭建视觉-语言-动作（VLA）策略从数据到部署的完整流程：VR 遥操作数据采集、模型微调，到机器人端的 ROS 2 推理。
- Hong Kong University of Science and Technology (Guangzhou)香港科技大学（广州） Built a language-conditioned trajectory planner in PyTorch, combining T5 text embeddings with a Diffusion Transformer. 使用 PyTorch 构建语言条件轨迹规划器，将 T5 文本嵌入与 Diffusion Transformer 结合。 Developed a two-pass workflow that described an initial trajectory in language, then replanned with that description. Evaluated the refinements in nuPlan closed-loop simulation. 设计两阶段流程：先从初始轨迹提取驾驶行为描述，再以文本为条件重新规划，并在 nuPlan 闭环仿真中评估改进效果。
- Berkeley AI Research (BAIR), UC Berkeley加州大学伯克利分校 BAIR 实验室 Built a Python evaluation pipeline for local and API-hosted vision-language models, covering dataset preparation, controlled visual-prompt experiments, parallel inference, and accuracy analysis. 搭建面向本地及 API 视觉语言模型的 Python 评测流程，涵盖数据准备、视觉提示对照实验、并行推理与准确率分析。 Evaluated nine models across 16 marker variants to study how visual prompts affect performance. Co-authored Visually Prompted Benchmarks Are Surprisingly Fragile. 评测 9 个模型在 16 种标记变体下的表现，研究视觉提示设计对模型性能的影响，并共同署名《Visually Prompted Benchmarks Are Surprisingly Fragile》。
Earlier, I also explored user preferences for AI-generated content and weather-aware image restoration.
更早之前，我还做过 AI 生成内容的用户偏好分析，以及天气感知的图像修复。
Publication论文
- Visually Prompted Benchmarks Are Surprisingly Fragile arXiv:2512.17875, 2025
Miscellaneous其他
Outside of work, I enjoy long-distance running. I've completed several 10K campus races and marathons. Running helps me think through problems and maintain balance.
工作之外，我热爱长跑运动。我完成过多次校园 10 公里跑和马拉松比赛。跑步帮助我思考问题并保持身心平衡。
Medals from 10K races and half marathons 10 公里和半程马拉松奖牌