Our paper “Unifying Embodied Foundation Models via a Model-to-Policy Formulation” is on arXiv. It provides a unified formulation for embodied AI paradigms, including vision-language-action models (VLAs) and world-action models (WAMs).
Biography 🙋
EN/CNI am a Ph.D. student in AI & Statistics at the School of Mathematical Sciences, Shanghai Jiao Tong University, supervised by Prof. Mingyang Ren. Before joining SJTU, I completed my bachelor’s degree in Statistics & Computer Science at the School of Statistics, East China Normal University, where I received the Outstanding Undergraduate Thesis Award. In spring 2026, I worked as a research assistant in the Department of Data Science and Artificial Intelligence at The Hong Kong Polytechnic University, mentored by Prof. Binyan Jiang.
My current research focuses on post-training for multimodal foundation models, particularly on-policy distillation for flow matching models. I aim to build post-training algorithms on a statistical machine learning framework, and validate them through large-scale experiments and evals to understand when and why they remain reliable at foundation-model scale.
News 📢
scroll for more →
Just a quick update: our work on HYSET was featured in Andrew Ng’s The Batch newsletter under “Pick the Right Tools for the Right Job”! Really happy to see our research shared with a wider audience, and grateful to Andrew and the DeepLearning.AI team for the thoughtful coverage.
Starting my Ph.D. at the School of Mathematical Sciences,
Shanghai Jiao Tong University, supervised by Prof. Mingyang Ren.
HYSET is on arXiv and under review at
AAAI 2027: set-level tool retrieval for LLM agents, cast as query-conditioned hyperedge
prediction. Code, model, and a live demo are released.
Graduated from East China Normal University with the
Outstanding Undergraduate Thesis Award and the
Outstanding Undergraduate Graduate honor.
HYVINT, our intensity-driven hypergraph
generative model, is on arXiv and under review at AAAI 2027.
Joined the Department of Data Science and Artificial Intelligence, PolyU, as a research
assistant with Prof. Binyan Jiang.
Publications 📄
* corresponding author
Tools Are Not Islands: Set-Level Tool Retrieval for LLM Agents via Query-Conditioned Hyperedge Prediction
Under review, AAAI Conference on Artificial Intelligence (AAAI), 2027
arXiv:2607.25718
HYVINT: Intensity-Driven Hypergraph Generation with Variational Embeddings
Under review, AAAI Conference on Artificial Intelligence (AAAI), 2027
arXiv:2605.16836
Enhancing Mixed Oil Length Prediction through Graph Representation Learning
Under review, Journal of Pipeline Science and Engineering
No publications in this category yet.
Experience 🚀
The Hong Kong Polytechnic University
2026.04 - 2026.05
Research Assistant · Department of Data Science and Artificial Intelligence
Mentor: Prof. Binyan Jiang
Education 🎓
Shanghai Jiao Tong University
2026.09 - present
Ph.D. in AI & Statistics · School of Mathematical Sciences
Supervisor: Prof. Mingyang Ren
East China Normal University
2022.09 - 2026.06
B.S. in Statistics & Computer Science · School of Statistics
Awards & Honors 🏅
Outstanding Undergraduate Thesis Award
East China Normal University
2026.06
Outstanding Undergraduate Graduate
East China Normal University
2026.06
National Second Prize (Team Leader)
National Undergraduate Statistical Modeling Competition
2024.08
Outstanding Graduate
Hefei No.1 High School
2022.06
Finalist, Young Gifted Program
Xi’an Jiaotong University
2019.03
Skills & Languages 🛠️
Artificial Intelligence
Multimodal Foundation Models, Flow Matching / Diffusion Models, Post-Training, LLM Agents, Scaling Laws
Statistics
High-Dimensional Inference, Empirical Process Theory, Optimal Transport, Variational Inference, Uncertainty Quantification, ODE/SDE
Engineering
Python, PyTorch, Flash Attention, FSDP, Megatron-LM, vLLM, SGLang, Slurm
Languages
Mandarin (native), English (professional), Cantonese (conversational), Japanese (beginner)
GitHub activity
Contact 💌
I welcome discussions and potential collaborations on post-training for generative foundation models. Please feel free to contact me at xinyi.hong@sjtu.edu.cn.