CV
Education, research experience, teaching, and academic service.
General Information
| Full Name | Xiaozhuang Song (宋晓壮) |
| xiaozhuangsong1 [AT] link [DOT] cuhk [DOT] edu [DOT] cn | |
| Research | AI for Science; structured representations and generation; LLM-guided reasoning, search, and planning |
| Profiles | |
| Full CV | |
Education
-
2023–2027 Ph.D. in Computer Science
The Chinese University of Hong Kong, Shenzhen - Sep 2023 — Mar 2027
- Expected graduation · School of Data Science
- Advisor: Tianshu Yu
-
2019–2022 Master's in Computer Science and Engineering
Southern University of Science and Technology - Sep 2019 — Jul 2022
- Multi-task learning, transfer learning & spatiotemporal modeling
- Advisor: James J. Q. Yu
-
2015–2019 B.Eng. in Digital Media Technology
Beijing University of Posts and Telecommunications - Sep 2015 — Jul 2019
- Undergraduate studies
Research Experience
-
2025–Present Shanghai, China
Research Intern
Shanghai AI Laboratory Core member of the multi-step retrosynthesis planning project.- Jul 2025 — Present
- Designed AOT*, integrating LLM-generated routes with AND-OR tree search, retrieval, reward assignment, and multi-level reaction verification.
- Developed RetroGEF for single-step retrosynthesis with dynamic graph edit generation.
- Contributed to RIGS, Latent JEPA, and MolEmb, covering controllable generation, latent chemical reasoning, and molecular representations.
- Integrated template retrieval, candidate generation, multi-step search, and the MoSyn route evaluation framework.
-
2024–2025 Shanghai, China
Research Intern
Shanghai AI Laboratory Research on language model reasoning over structured knowledge.- Dec 2024 — Feb 2025
- Designed ReKG-MCTS, a training-free reasoning framework combining language models with Monte Carlo tree search (Findings of ACL 2025).
- Modeled reasoning as sequential decisions with UCB selection, path expansion, LLM-guided rollout, and value backpropagation.
-
2021–2023 Beijing, China
Research Intern · Machine Learning Area
Microsoft Research Asia Mentors: Shun Zheng and Jiang Bian.- Jun 2021 — Jun 2023
- Proposed meta-learning on task combinations for efficient multi-task grouping (NeurIPS 2022).
- Participated in prototype-based tabular representation learning via space calibration (ICLR 2024 Spotlight).
- Contributed a dynamic programming and imitation learning approach to the NeurIPS 2022 CityLearn Challenge, placing fifth.
Research Interests
-
Representation
- Learning useful representations of tabular, temporal, graph, and molecular data, through multi-task learning, space calibration, and multimodal embedding.
-
Generation
- Modeling graph edits and molecular geometry with diffusion and flow-based methods, from conformer generation and fluid dynamics to single-step retrosynthesis.
-
Reasoning & planning
- Combining LLMs with explicit graph and tree structures to explore reasoning paths, reuse intermediate results, and build executable synthesis routes.
Selected Publications
- AOT*: Efficient Synthesis Planning via LLM-Empowered AND-OR Tree Search — Findings of ACL 2026.
- RetroGEF: Dynamic Graph Edit Flow for Single-Step Retrosynthesis — preprint, 2026.
- ReKG-MCTS: Reinforcing LLM Reasoning on Knowledge Graphs via Training-Free Monte Carlo Tree Search — Findings of ACL 2025.
- Physically-Informed Flow Matching with Graph Neural Networks for Complex Fluid Dynamics — AAAI 2026.
- Efficient and Effective Multi-task Grouping via Meta Learning on Task Combinations — NeurIPS 2022.
- View all publications and preprints.
Teaching
-
Spring 2026 CSC3100 · Data Structures
CUHK-Shenzhen - Teaching Assistant
-
Fall 2025 CSC1001 · Introduction to Computer Science
CUHK-Shenzhen - Leading TA · 700-student course
-
Fall 2024 STA2001 · Probability and Statistical Inference
CUHK-Shenzhen - Teaching Assistant
-
Spring 2024 CSC3150 · Operating Systems
CUHK-Shenzhen - Teaching Assistant
-
Fall 2023 DDA3020 · Machine Learning
CUHK-Shenzhen - Teaching Assistant
-
Spring 2021 CS201 · Digital Logic
SUSTech - Teaching Assistant
Academic Service
| Since | 2022 |
| Conferences | NeurIPS, ICML, ICLR, WWW, AISTATS, CVPR, KDD, AAAI, ACL, COLM, IJCAI |
| Journals | TPAMI, TKDE |
Honors and Other Experience
- NeurIPS 2022 CityLearn Challenge: Fifth place · dynamic programming and imitation learning for energy control
- CCF-BDCI 2020: Top 10% · road spatiotemporal forecasting
- Invention patents: Co-inventor of six publicly disclosed patents during master's studies
Technical Skills
| Research & modeling | Python, PyTorch, DeepSpeed, scientific generative modeling |
| Research engineering | Slurm, Linux, cloud clusters (Azure, Alibaba Cloud, Volcengine), Shell, LaTeX |
| Development workflows | Agentic coding tools and research-to-system integration |