He Zhu

I am HE ZHU (朱赫), an M.Sc. student in Smart City and Data Sciences at Peking University, advised by Prof. Wenjia Zhang and Prof. Guanhua Chen. I received my B.E. in Computer Science from Southern University of Science and Technology, advised by Prof. Zipei Fan and Prof. Xuan Song.

My research centers on post-training data for LLMs: what makes it good, how to synthesize and select it at scale (FANNO, InstructDiff, Tag-Instruct, AlignDiff), and how to design training objectives that acquire capabilities efficiently without forgetting what the model already knows (ASFT, CFT). I have also applied this perspective to domain-specific foundation models for urban intelligence (PlanGPT, PlanGPT-VL, UrbanClaw). I am currently a research intern at Microsoft Research Asia, GenAI Group, supervised by Dr. Li Dong, and I serve as an Area Chair for EMNLP 2026.

Email  /  CV  /  Google Scholar  /  Github  /  Phone

Open to discussion or collaboration. Feel free to drop me an email if you're interested in my research.

2026.06: I will serve as an Area Chair for EMNLP 2026.
2026.04: InstructDiff was accepted by ACL 2026 Main.
2026.01: ASFT was accepted by ICLR 2026.
2025.11: Joined MSRA GenAI Group as a Research Intern.
2025.10: PlanGPT-VL was accepted by EMNLP Industry 2025.
2025.08: PlanGPT was selected as an ACL Industry 2025 Oral (Top 1%).
2025.05: Three first-author papers (FANNO, Tag-Instruct, PlanGPT) were accepted by ACL 2025.

First Author / Corresponding Author

Anchored SFT studies fine-tuning objectives that acquire new capabilities while preserving a model's original knowledge.

The first systematic study of LLMs for urban planning, covering data, models, benchmarks, and applications.

A domain-adaptive data selection method for efficient LLM fine-tuning.

UPBench evaluates whether LLMs can reason with professional planning judgment across planning knowledge and cognitive levels.

A spatial planning map benchmark for evaluating VLM perception, reasoning, association, and implementation in planning contexts.