About Me

Hello! I’m Xutian Chen, a passionate student and researcher in the field of Artificial Intelligence and Machine Learning. I am currently pursuing my Master’s degree in Artificial Intelligence at Beihang University, having graduated with a Bachelor’s degree in Artificial Intelligence from Jinan University.

My research interests lie in the intersection of AI for Bio&Chem, Diffusion Models, LLM Inference Acceleration, and Autonomous Vehicle Path Planning. I am always open to appropriate opportunities in research and tech industries, and I am excited about how GenAI can be applied to revolutionize specific problems in any domain.

My Advisors: During my undergraduate studies at Jinan University, I was fortunate to be advised by Lecturer Deping Li, whose research focuses on intelligent visual perception, including 3D object pose estimation, point cloud processing, visual foundation models, SLAM, and robotic visual perception. At Beihang University, I am fortunate to be advised by Prof. Huijie Zhao, whose research covers multi/hyperspectral intelligent perception, multi-modal perception and data fusion, and 3D visual perception and data processing, and Associate Prof. Na Li of the School of Artificial Intelligence.

Contact: chenxutian@buaa.edu.cn
GitHub: github.com/Blossom0913
Gitee: gitee.com/chenxutian

Remember brick walls let us show our dedication. They are there to separate us from the people who don't really want to achieve their childhood dreams.

关于我

你好，我是 陈旭天，专注于人工智能与机器学习方向的学习与研究。目前我在北京航空航天大学攻读人工智能硕士学位，本科毕业于暨南大学人工智能专业。

我的研究兴趣主要包括 AI for Bio&Chem、扩散模型、大语言模型推理加速 和 自动驾驶路径规划。我持续关注科研与产业实践中的合适机会，也对生成式 AI 在不同领域中的落地应用充满兴趣。

导师信息：在暨南大学本科期间，我有幸在李德平老师指导下开展研究，其方向包括智能视觉感知、三维目标位姿估计、点云处理、视觉基础模型、SLAM 与机器人视觉感知。在北航期间，我有幸接受赵慧洁教授与人工智能学院李娜副教授的指导，研究方向涵盖多/高光谱智能感知、多模态感知与数据融合、三维视觉感知与数据处理等。

联系方式：chenxutian@buaa.edu.cn
GitHub：github.com/Blossom0913
Gitee：gitee.com/chenxutian

Remember brick walls let us show our dedication. They are there to separate us from the people who don't really want to achieve their childhood dreams.

News

• 2026.07: Our paper “A Benchmark Dataset for Rat Social and Aggressive Behavior Classification” was published in Scientific Data.
• 2026.03: Started my internship as a Decision & Planning Intern at Meituan Autonomous Vehicle Department, Beijing.
• 2025.09: Started my Master’s degree in Artificial Intelligence at Beihang University, Beijing.
• 2025.07: Graduated from Jinan University with a Bachelor’s degree in Artificial Intelligence.

新闻

• 2026.07：论文 “A Benchmark Dataset for Rat Social and Aggressive Behavior Classification” 发表在 Scientific Data。
• 2026.03：在北京美团自动车事业部开始决策与规划实习。
• 2025.09：在北京航空航天大学开始攻读人工智能硕士。
• 2025.07：从暨南大学人工智能专业本科毕业。

Education

• 2025.09 - Present: Master of Engineering in Artificial Intelligence, Beihang University, Beijing, China
  • Major Courses: Reinforcement Learning, Deep Learning, Multi-Robot Swarm Intelligence, Algorithm Design and Analysis
• 2021.09 - 2025.07: Bachelor of Engineering in Artificial Intelligence, Jinan University, Zhuhai, China
  • Major Courses: Natural Language Processing, Machine Learning, Principles of Artificial Intelligence, Agile Software Development

教育经历

• 2025.09 - 至今：人工智能工程硕士，北京航空航天大学，中国北京
  • 主要课程：强化学习、深度学习、多机器人群体智能、算法设计与分析
• 2021.09 - 2025.07：人工智能工学学士，暨南大学，中国珠海
  • 主要课程：自然语言处理、机器学习、人工智能原理、敏捷软件开发

Experience

Internship

Decision & Planning Intern | Meituan Autonomous Vehicle Department

2026.03 - present | Beijing, China

• Intersection data closed loop: Maintained and repaired the daily human-driving data pipeline, systematically resolving upstream data gaps, path inconsistencies, Protobuf migration, and dataset-loading issues; backfilled historical gaps and designed an automated collection workflow covering driving modes, turn types, RA-result fusion, filtering, and distribution statistics.
• Large-scale dataset construction: Built an autonomous right-turn intersection dataset spanning 2025.06–2026.05 through SQL extraction, RA mining, quality filtering, and timestamp alignment; distilled 2.789M cases / 13,501 hours of raw data into 1.081M cases / 3,446 hours ready for training, alongside a human-driving dataset.
• Model training and simulation evaluation: Delivered the end-to-end Query Decoder workflow—data reprocessing, feature export, Backbone/Head training, ONNX export, and simulation regression—while adapting to Shared Encoder 6.3; analyzed representative cases for trajectory avoidance, lane selection, and curvature, and improved Mviz visualization for intersection models.

Research Experience

Research on AI4S and AI+Chem | Guangdong Institute of Intelligent Science and Technology

Research Intern | 2025.02 - 2025.09 | Zhuhai, China

• Large-Scale Computational Framework: Engineered a parallel computing framework deployed on a 4-GPU (RTX 2080Ti) cluster, successfully processing 43.8 million molecular docking tasks to accelerate drug discovery pipelines. Source code: Dock Repository.
• Algorithm Research & Benchmarking: Designed an experimental framework and benchmark dataset for rat social and aggressive behavior classification, using DeepLabCut for keypoint labeling; conducted comparative experiments on LightGBM, LSTM, CNN, and GMM models to evaluate performance. Project code: Mouse-Behavior-Classifier-Train. This work was published as co-first author in Scientific Data: “A Benchmark Dataset for Rat Social and Aggressive Behavior Classification”.
• System Development: Contributed to the development of a multi-robot path planning system and a lightweight task management platform, resulting in 2 software copyright registrations (Top-3 Author).

Multi-agent Path Planning | Jinan University

Research Assistant | 2024.03 - 2024.07 | Zhuhai, China

• Performance Analysis: Designed controlled experiments scaling agents from 10 to 50, revealing that runtime increased 135x (2.3s → 312s) and success rate dropped to 67%. Profiling identified conflict detection (70% of runtime) and deep copy operations as primary bottlenecks.

• Algorithm Optimization (C++/MAPF): Reduced conflict detection complexity from O(n²) to O(n) by implementing incremental checks—comparing only the current agent against others instead of full scans. Reduced memory overhead by replacing deep copies with Copy-on-Write, enabling multiple agents to share conflict tree nodes until a write is required.

• Systems Thinking: Applied knowledge of stack/heap allocation and shallow/deep copy semantics to guide optimization decisions. Derived complexity reduction (N² → N) and presented findings with clear problem-solution-impact narrative—demonstrating interpretability in engineering work.

• Platform Integration: Built messaging architecture between AGV fleet and local server for state synchronization under real-world constraints. Reproduced CL-CBS baseline (paper: CL-MAPF), fixed implementation bugs, and delivered faster planning than Hybrid A* baseline within a 3-month cycle.

Competition Experience

ASC2022 Student Supercomputer Challenge

Team Member | 2021.11 - 2022.06 | Zhuhai, China

• Project goal: Under limited compute (8× Tesla V100 16GB) and power constraints, pre-train a 4.7B-parameter Yuan-1.0 language model and achieve a 55% training speedup (from 45h to 28h).
• Memory optimization: Led application of ZeRO-Offload and ZeRO Stage 1 to offload model states (parameters, gradients, optimizer states) to CPU memory, enabling training of a 4.7B model on 8×16GB GPUs and resolving CUDA OOM issues.
• Parallel & acceleration strategy: Deployed Megatron-DeepSpeed and designed a hybrid parallel scheme with 4-way tensor parallelism + 2-way pipeline parallelism, improving throughput from 4.08 → 4.66 samples/s compared with pure 8-way tensor parallelism.
• Engineering optimizations: Adopted mixed precision training (AMP), built PyTorch with Intel MKL, and used DeepSpeed’s CPU Adam to optimize CPU offload computation and communication.
• Convergence tuning: Tuned learning rate scaling and warmup strategies alongside data pipelines and micro-batching, achieving a final training loss of 5.826 in reproduction runs.
• Deliverables: Responsible for parallel strategy evaluation, memory/performance analysis, and engineering implementation; project notes and partial experiment logs at: ASC Student Supercomputer Challenge Proposal

经历

实习经历

决策与规划实习生 | 美团自动车事业部

2026.03 - 至今 | 中国北京

• 路口数据闭环：维护并修复日常人驾数据流水线，系统性解决上游数据缺失、路径不一致、Protobuf 迁移、数据集加载等问题；补齐历史缺口，并设计覆盖驾驶模式、转向类型、RA 结果融合、过滤与分布统计的自动化采集流程。
• 大规模数据集构建：通过 SQL 抽取、RA 挖掘、质量过滤与时间戳对齐，构建 2025.06–2026.05 自动驾驶路口右转数据集；将 278.9 万 case / 13,501 小时原始数据提炼为 108.1 万 case / 3,446 小时训练数据，并同步建设人驾数据集。
• 模型训练与仿真评估：完成 Query Decoder 端到端流程（数据重处理、特征导出、Backbone/Head 训练、ONNX 导出、仿真回归），适配 Shared Encoder 6.3；对避让轨迹、选道与曲率等典型案例进行分析，并优化路口模型 Mviz 可视化。

科研经历

AI4S 与 AI+Chem 研究 | 广东省智能科学与技术研究院

科研实习生 | 2025.02 - 2025.09 | 中国珠海

• 大规模计算框架：在 4 GPU（RTX 2080Ti）集群上搭建并行计算框架，完成 4380 万分子对接任务处理，加速药物发现流程。代码仓库：Dock Repository。
• 算法研究与基准：使用 DeepLabCut 进行关键点标注，构建大鼠社交与攻击行为分类实验框架与基准数据集，并对 LightGBM、LSTM、CNN、GMM 等模型开展对比实验。项目代码：Mouse-Behavior-Classifier-Train。成果以共同一作发表于 Scientific Data：“A Benchmark Dataset for Rat Social and Aggressive Behavior Classification”。
• 系统开发：参与多机器人路径规划系统与轻量级任务管理平台开发，获 2 项软件著作权（前三作者）。

多智能体路径规划 | 暨南大学

科研助理 | 2024.03 - 2024.07 | 中国珠海

• 性能分析：设计 10 到 50 智能体的对照实验，发现运行时间增加 135 倍（2.3s → 312s），成功率下降至 67%；剖析定位冲突检测（约占 70%）与深拷贝为主要瓶颈。
• 算法优化（C++/MAPF）：通过增量冲突检查将复杂度从 O(n²) 降至 O(n)，仅比较当前智能体与其他智能体；通过 Copy-on-Write 替代深拷贝以降低内存开销，使多个智能体可共享冲突树节点直到写入发生。
• 系统化思维：结合栈/堆分配与浅拷贝/深拷贝语义指导优化方案，明确推导复杂度下降（N² → N），并形成问题-方案-收益的可解释工程分析。
• 平台集成：在真实约束下完成 AGV 车队与本地服务器消息架构，实现状态同步；复现 CL-CBS 基线（论文：CL-MAPF），修复实现问题，并在 3 个月内实现快于 Hybrid A* 基线的规划性能。

竞赛经历

ASC2022 世界大学生超算竞赛

队员 | 2021.11 - 2022.06 | 中国珠海

• 项目目标：在受限算力（8× Tesla V100 16GB）和功耗约束下，完成 47 亿参数 Yuan-1.0 模型预训练，并实现 55% 训练提速（45h → 28h）。
• 显存优化：主导应用 ZeRO-Offload 与 ZeRO Stage 1，将参数、梯度、优化器状态卸载至 CPU 内存，使 4.7B 模型可在 8×16GB GPU 上训练并解决 CUDA OOM。
• 并行与加速策略：部署 Megatron-DeepSpeed，设计 4 路张量并行 + 2 路流水并行混合方案，相较纯 8 路张量并行，吞吐从 4.08 提升至 4.66 samples/s。
• 工程优化：采用混合精度训练（AMP），使用 Intel MKL 编译 PyTorch，并结合 DeepSpeed CPU Adam 优化 CPU 卸载计算与通信。
• 收敛调优：联合学习率缩放、warmup、数据流水与 micro-batch 策略调参，在复现实验中达到最终训练损失 5.826。
• 产出职责：负责并行策略评估、内存/性能分析与工程实现。项目记录与部分实验日志见：ASC Student Supercomputer Challenge Proposal

Publications

• Xutian Chen*, Guangyu Li*, Zihan Zhang*, Mingkun Xu, Zuoren Wang, Qianqian Shi. “A Benchmark Dataset for Rat Social and Aggressive Behavior Classification”. Scientific Data, 2026. (*Co-first authors)

More manuscripts are in preparation. Feel free to reach out for collaboration opportunities!

发表论文

• Xutian Chen*，Guangyu Li*，Zihan Zhang*，Mingkun Xu，Zuoren Wang，Qianqian Shi。“A Benchmark Dataset for Rat Social and Aggressive Behavior Classification”。Scientific Data，2026。（*共同一作）

更多稿件正在准备中，欢迎交流与合作。

Honors and Awards

• 2022.06: National Second Prize in ASC2022 (Student Supercomputer Challenge), ranked 22nd among all participants
• 2022.07: Attended the ASC final as visitors at USTC
• 2024: Provincial First Prize (preliminary) in Chinese Mathematics Competition (CMC), Guangdong Division
• 2023: Provincial Second Prize (preliminary) in Chinese Mathematics Competition (CMC), Guangdong Division

荣誉与奖励

• 2022.06：ASC2022（世界大学生超算竞赛）全国二等奖，总排名 22
• 2022.07：以观摩身份参加 ASC 总决赛（中科大）
• 2024：中国大学生数学竞赛（广东赛区）省一等奖（初赛）
• 2023：中国大学生数学竞赛（广东赛区）省二等奖（初赛）

Projects

Tiny-llm

2026.04 – present

• Focused on Attention implementation and lightweight LLM training/inference pipelines
• Tech stack: vLLM, SGLang, Tool Calling
• Reference: https://github.com/skyzh/tiny-llm

PlayTask - Time Management APP

2023.09 – 2023.12 | Personal Project

• Designed and built a time management APP from scratch
• Learned Version Control Systems (VCS), Event Response and basic debugging tools in Android Studio
• Designed UI/UX with ViewPage2, TabLayout and Fragment independently
• Source code: Blossom0913/PlayTask

项目

Tiny-llm

2026.04 – 至今

• 关注 Attention 实现与轻量化 LLM 训练/推理流程
• 技术栈：vLLM、SGLang、Tool Calling
• 参考项目：https://github.com/skyzh/tiny-llm

PlayTask - 时间管理 APP

2023.09 – 2023.12 | 个人项目

• 从零设计并开发时间管理 APP
• 在 Android Studio 中实践 版本控制系统（VCS）、事件响应机制 与基础调试工具
• 独立完成基于 ViewPage2、TabLayout、Fragment 的 UI/UX 设计
• 源码地址：Blossom0913/PlayTask

Skills

Programming Languages

• Proficient: Python, C/C++, Java
• Experienced: Rust, Kotlin

Technical Skills

• Development Tools: Git, SSH, Android Studio
• High Performance Computing: CUDA, MPI, OpenMP
• Distributed Training & Optimization: DeepSpeed, Megatron, ZeRO, ZeRO-Offload
• Precision & Acceleration: Mixed Precision (AMP), Intel MKL optimizations
• Machine Learning: TensorFlow, PyTorch, Fine-Tuning
• Systems: Linux, Shell
• Other: Algorithm Design, Multi-Agent Systems, Computer Vision, Model Inference & Deployment

技能

编程语言

• 熟练：Python、C/C++、Java
• 了解并实践：Rust、Kotlin

技术能力

• 开发工具：Git、SSH、Android Studio
• 高性能计算：CUDA、MPI、OpenMP
• 分布式训练与优化：DeepSpeed、Megatron、ZeRO、ZeRO-Offload
• 精度与加速：混合精度（AMP）、Intel MKL 优化
• 机器学习：TensorFlow、PyTorch、Fine-Tuning
• 系统能力：Linux、Shell
• 其他：算法设计、多智能体系统、计算机视觉、模型推理与部署