XS Xingyu Song

PhD Student · The University of Tokyo

Xingyu Song

I study flow-based generative models and the deterministic and stochastic dynamics behind modern AI systems.

GitHub Google Scholar CV

About

Building simple systems with subtle machinery.

I am a first-year PhD student at the RCAST AI Lab, Department of Advanced Interdisciplinary Studies, Graduate School of Engineering, The University of Tokyo, supervised by Prof. Naoya Takeishi and Prof. Takehisa Yairi.

My research explores flow-based generative models: how probability flows can generate the world we imagine. Recently, I have been drawn to the intrinsic connections between deterministic and stochastic dynamics.

I like building things that reflect the beauty and philosophy of computer science and AI: simplicity that hides subtle complexity.

Research

Current focus areas

01

Generative Dynamics

Flow-based generative modeling, probability transport, bridge matching, and stochastic process structure.

02

Physical Simulation

Training measures, multi-scale partitions, neural surrogates, and bias reduction for discretized simulations.

03

Embodied Vision

Action recognition, synthetic data generation, 3D pose representation, and game-engine-based augmentation.

Publications

Selected and recent work

Grouped by research area, newest first within each section.

Generative Modeling & Physical Simulation

  1. Deterministic Decomposition of Stochastic Generative Dynamics Xingyu Song, Yuan Mei, Naoya Takeishi. arXiv preprint, 2026.

    Decomposes generative drift into transport and osmotic components and introduces Bridge Matching for controllable sampling.

  2. M3: Reframing Training Measures for Discretized Physical Simulations Yuan Mei, Xingyu Song, Xiaowen Song, Naoya Takeishi. arXiv preprint, 2026.

    Balances supervision across multi-scale Morton partitions to reduce measure-induced bias in neural surrogate models.

Natural Language Processing

  1. ReAgent: Reversible Multi-Agent Reasoning for Knowledge-Enhanced Multi-Hop QA Zhao Xinjie, Fan Gao, Xingyu Song, Yingjian Chen, Rui Yang, Yanran Fu, Yuyang Wang, Yusuke Iwasawa, Yutaka Matsuo, Irene Li. EMNLP, 2025.
  2. JiraiBench: A Bilingual Benchmark for Evaluating Large Language Models' Detection of Human Risky Health Behavior Content in Jirai Community Yunze Xiao, Tingyu He, Lionel Z. Wang, Yiming Ma, Xingyu Song, Xiaohang Xu, Mona T. Diab, Irene Li, Ka Chung Ng. EACL, 2026.
  3. Exploring the Role of Knowledge Graph-Based RAG in Japanese Medical Question Answering with Small-Scale LLMs Yingjian Chen, Feiyang Li, Xingyu Song, Tianxiao Li, Zixin Xu, Xiujie Chen, Issey Sukeda, Irene Li. arXiv preprint, 2025.

Computer Vision & Action Recognition

  1. 4A: An Animation-based Augmentation Approach for Action Recognition from Discontinuous Video Xingyu Song, Zhan Li, Shi Chen, Xin-Qiang Cai, Kazuyuki Demachi. ECAI, 2024 (Oral).
  2. Quater-GCN: Enhancing 3D Human Pose Estimation with Orientation and Semi-supervised Training Xingyu Song, Zhan Li, Shi Chen, Kazuyuki Demachi. ECAI, 2024 (Oral).
  3. GTAutoAct: An Automatic Datasets Generation Framework Based on Game Engine Redevelopment for Action Recognition Xingyu Song, Zhan Li, Shi Chen, Kazuyuki Demachi. arXiv preprint, 2024.

Intelligent Surveillance & Security

  1. Advancement and Development of Graph-Based Reasoning Method for Human Malicious Behaviors Identification Based on Graph Contrastive Representation Learning Zhan Li, Xingyu Song, Shi Chen, Kazuyuki Demachi. Neurocomputing, 2026.
  2. Armed Boundary Sabotage: A Case Study of Human Malicious Behaviors Identification with Computer Vision and Explainable Reasoning Methods Zhan Li, Xingyu Song, Shi Chen, Kazuyuki Demachi. Computers and Electrical Engineering, 121, 109924, 2025.
  3. Abnormal Detection in Nuclear Security Videos Based on Label-Specific Autoencoders and Reconstruction Errors Comparison Zhan Li, Xingyu Song, Shi Chen, Kazuyuki Demachi. Nuclear Engineering and Technology, 57(3), 103239, 2025.
  4. Data, Language and Graph-Based Reasoning Methods for Identification of Human Malicious Behaviors in Nuclear Security Zhan Li, Xingyu Song, Shi Chen, Kazuyuki Demachi. Expert Systems with Applications, 236, 121367, 2024.
  5. Game Engine Based Data Augmentation with In-game Customization and Modeling for Malicious Behaviors Identification in Nuclear Security Xingyu Song, Zhan Li, Shi Chen, Kazuyuki Demachi. INMM/ESARDA Joint Annual Meeting, 2023.
  6. Malicious Behaviors Identification in Nuclear Security Based on Visual Relationships Extraction and Knowledge Reasoning Zhan Li, Xingyu Song, Shi Chen, Kazuyuki Demachi. INMM/ESARDA Joint Annual Meeting, 2023.

CV

Education, honors, and internships

Education

  1. PhD, Graduate School of Engineering, The University of Tokyo.
  2. M.Eng., Graduate School of Engineering, The University of Tokyo.
  3. Research Student, School of Fundamental Science and Engineering, Waseda University.
  4. B.Sc., College of Hongshen / College of Computer Science, Chongqing University.

Honors and Awards

  1. SPRING GX Fellowship, The University of Tokyo.

Internships

  1. Sichuan Hwadee Information Technology Co., Ltd, Chengdu, China.
  2. National Key Laboratory of Software Development Environment, Beihang University, Beijing, China. Supervised by Prof. Zhiming Zheng.
  3. ChinaSoft International, Chongqing, China. CQUHub.

Contact

Open to research conversations and collaborations.

github.com/xingyu-song Google Scholar profile