About Me
关于我
I am a final-year undergraduate at Shanghai Jiao Tong University, pursuing a B.Eng. in Automation in the IEEE Honor Class with a minor in Law (Intellectual Property). I am currently a Foundation Model Evaluation Intern at Alibaba Group Holding Limited, where I am mentored by Niantong Li, a core contributor to Qwen-Image. My work at Alibaba includes building evaluation benchmarks, data, and supporting tools for generative models. My work on dynamic routing for Mixture-of-Agents was accepted as an ACL 2026 main-conference long paper.
I am currently preparing applications to master's programmes, including at The Chinese University of Hong Kong, The University of Hong Kong, and Cornell University. I am also seeking internship opportunities with leading foundation-model and multimodal-evaluation teams in China and internationally, and welcome conversations with academic labs. My research interests centre on multi-agent architectures, the pre-training and post-training of foundation models, and the design of robust evaluation systems for large language and multimodal models.
我是上海交通大学自动化专业（IEEE试点班）本科四年级学生，辅修法学（知识产权方向）。目前，我在阿里巴巴控股集团担任大模型评测实习生，由 Qwen-Image 核心贡献者李念潼指导。在阿里期间，我参与面向生成式模型的评测基准、评测数据与相关工具建设。关于混合智能体动态路由的研究工作已作为 ACL 2026 主会长文录用。
我正在申请包括香港中文大学、香港大学和康奈尔大学在内的硕士项目，同时寻找国内外领先基础模型与多模态评测团队的实习机会，也期待与相关研究课题组交流合作。我的研究兴趣主要集中在多智能体架构、基础模型的预训练与后训练，以及面向大语言模型和多模态模型的可靠评测体系建设。
Education
教育背景
Shanghai Jiao Tong University
上海交通大学
Aug 2023 – Jul 2027 (Expected)
2023 年 8 月 – 2027 年 7 月（预计）
B.Eng. in Automation, IEEE Honor Class
全日制工学学士 · 自动化（IEEE试点班）
School: School of Electronic Information and Electrical Engineering. Minor: Law (Intellectual Property).
学院：电子信息与电气工程学院。辅修：法学（知识产权方向）。
Core coursework: Artificial Intelligence Principles and Applications, Machine Learning, Computer Vision, and Robotics.
核心课程：人工智能原理与应用、机器学习、计算机视觉、机器人学。
Honors: Zhiyuan Honors Scholarship, Shanghai Jiao Tong University.
荣誉：上海交通大学致远荣誉奖学金。
English proficiency: IELTS 7.0; CET-6 609; CET-4 630.
英语能力：雅思 7.0；大学英语六级 609；大学英语四级 630。
Publications
学术成果
RouteMoA: Dynamic Routing without Pre-Inference Boosts Efficient Mixture-of-Agents
RouteMoA：无需预推理的动态路由提升混合智能体系统效率
Jize Wang, Han Wu, et al.
ACL 2026 Main Conference (Long Paper)
ACL 2026 主会长文
FilmBench: A Film-Grade Benchmark for Cinematic Video Generation
FilmBench：面向电影级视频生成的高质量评测基准
Teamwork at Alibaba Group
arXiv preprint; submitted to AAAI 2027
arXiv 预印本；AAAI 2027 在投
Research & Industry
科研与产业实践
Alibaba Group Holding Limited
阿里巴巴控股集团
Jun 2026 – Present
Foundation Model Evaluation Intern
大模型评测实习生
Hangzhou
杭州
- FilmBench: Contributed to a film-grade T2V/R2V benchmark submitted to AAAI 2027, evaluating HappyHorse, Seedance 2.0, Kling 3.0/Omni, Veo 3.1, Vidu, and Hailuo on curated cinematic prompts.
- Assessed cinematography, multi-shot continuity, instruction following, and motion aesthetics, translating film-production concepts into consistent annotation guidelines.
- QwenEditBench: Supported an open-source image-editing benchmark developed with Tongyi Lab, refining the evaluation taxonomy for edit intent, instruction following, visual quality, and failure cases.
- Qwen-Image: Organised, filtered, and quality-checked a subset of image-generation training data against shared delivery standards.
- FilmBench：参与电影级 T2V/R2V 评测基准（AAAI 2027 在投），围绕精选电影化提示词评测 HappyHorse、Seedance 2.0、Kling 3.0/Omni、Veo 3.1、Vidu 和 Hailuo 等视频生成模型。
- 从镜头语言、多镜头连续性、指令遵循和动态美学等维度评估生成结果，并将电影制作概念转化为一致的标注规范。
- QwenEditBench：参与与通义实验室合作开发的开源图像编辑评测基准，完善编辑意图、指令遵循、视觉质量和失败案例的评测分类体系。
- Qwen-Image：根据统一交付规范，整理、筛选并质检部分图像生成训练数据。
RouteMoA: Dynamic Routing for Efficient Mixture-of-Agents
RouteMoA：面向高效混合智能体的动态路由方法
May 2025 – Sep 2025
Core Member
核心贡献者
Co-developed a dynamic-routing MoA framework and its evaluation stack across more than 20,000 QA examples. Compared with MoA and SMoA, the approach maintained stronger performance while reducing large-pool inference cost by 89.8% and latency by 63.6%; accepted as an ACL 2026 long paper.
共同构建动态路由 MoA 框架及覆盖超过 2 万条问答样本的评测栈；与 MoA、SMoA 对比后，在保持更优效果的同时将大模型池推理成本降低 89.8%、延迟降低 63.6%。相关成果作为 ACL 2026 主会长文录用。
SJTU & COMAC
上海交通大学 & 中国商飞
Jun 2024 – Sep 2024
Core Member, SJTU–COMAC Collaborative Project
上海交通大学—中国商飞合作项目核心成员
Built a MobileNetV3-based classifier for 12 types of aircraft-cabin anomalies, combining realistic-noise augmentation and strong-noise validation. Achieved 98.01% average accuracy and CPU inference below 20 ms for 0.4–1 s clips.
面向 12 类飞机客舱异常声音，构建基于 MobileNetV3 的轻量级分类器，并结合真实飞行噪声增强与强噪声测试；平均准确率达 98.01%，0.4–1 秒片段的 CPU 推理耗时低于 20 ms。
Selected Engineering Projects
代表性工程项目
CoppeliaSim, LLM Agents
Built a dual-line flexible-manufacturing simulator with natural-language order parsing and a three-agent planning architecture for coordinated assembly, transfer, inspection, and collision-aware shuttle control.
构建双产线柔性制造仿真系统，支持自然语言订单解析，并以三智能体架构协同完成装配、物料转运、质检与避碰控制。
MCTS, Bayesian Optimization
蒙特卡洛树搜索（MCTS）, 贝叶斯优化
Developed a physics-grounded eight-ball agent that combines geometric shot modelling, MCTS, Bayesian optimisation, and iterative LLM replanning; the strongest specialised agent reached an 86.75% win rate against baselines.
开发面向八球台球的物理建模智能体，结合几何击球建模、MCTS、贝叶斯优化与迭代式 LLM 重规划；最强专用智能体相对基线胜率达 86.75%。
OpenCV, PyTorch, LPRNet
Delivered an end-to-end license-plate recognition workflow—from image preprocessing and localisation to LPRNet inference and a PyQt5 desktop interface—with 86% character-level accuracy.
完成从图像预处理、车牌定位到 LPRNet 推理与 PyQt5 桌面界面的端到端车牌识别流程，单字符识别准确率达 86%。
Skills
技能
Python
C++
JavaScript
PyTorch
OpenCV
FastAPI
LMDeploy
OpenCompass
Git
Linux
LaTeX
Resume
简历
For a fuller account of my academic background, research, and engineering work, see my CV here.
如需了解完整的教育背景、研究经历与工程项目，请在此查看我的简历。