About
I am an M.S. student at the Shenzhen Graduate School of Peking University, advised by Prof. Wenjia Zhang, and currently an intern with the Seed Speech team at ByteDance Seed, where I work on data for speech foundation models.
My research is data-centric: I study how the choice, synthesis and weighting of training data decide what a language model can do after post-training. Recent work covers data selection (InstructDiff, ACL 2026), stable alignment (ASFT, ICLR 2026, merged into LLaMA-Factory) and instruction synthesis (TAG-INSTRUCT, ACL 2025 Findings).
Alongside that I work on the PlanGPT series: domain models for urban planning, from the original PlanGPT to the vision-language PlanGPT-VL, with further work underway. Separately, I have shipped three RAG systems for Chinese planning institutes as technical lead.
InterestsData selection & synthesis·LLM post-training and alignment·Domain-specific and multimodal LLMs·Spatiotemporal intelligence
News
- 2026.04
- InstructDiff was accepted to the ACL 2026 Main Conference.
- 2026.02
- ASFT was merged into LLaMA-Factory.
- 2026.02
- Started as an intern with the Seed Speech team at ByteDance Seed, working on Seed-Audio data.
- 2026.01
- ASFT was accepted to ICLR 2026.
- 2025.11
- PlanGPT-VL appeared at the EMNLP 2025 Industry Track.
- 2025.09
- Began my M.S. at Peking University, Shenzhen Graduate School.
Selected Publications
InstructDiff: Domain-Adaptive Data Selection via Contrastive Entropy for Efficient LLM Fine-Tuning
ACL 2026
Junyou Su*, He Zhu*, Xiao Luo, Liyu Zhang, Hong-Yu Zhou, Yun Chen, Peng Li, Yang Liu, Guanhua Chen
Annual Meeting of the Association for Computational Linguistics — Main Conference
- Warm-up calibration plus distribution-aware filtering: a contrastive-entropy signal picks the most learnable samples, replacing hand-written domain heuristics.
- Selecting 10% of the data beats training on all of it — +17% relative on math reasoning and +52% on general instruction following, reproduced on medical QA and code.
ACL Anthology
Anchored Supervised Fine-Tuning
ICLR 2026
He Zhu*, Junyou Su*, Peng Lai*, Ren Ma, Wenjia Zhang, Linyi Yang, Guanhua Chen
International Conference on Learning Representations·Merged into LLaMA-Factory.
- Dynamic sample weighting with lightweight KL anchoring removes the training drift of Dynamic Fine-Tuning.
- Casts DFT as reward-weighted regression, which yields a tighter lower bound and steadier convergence than plain SFT.
- Gains on math, medical QA and code at about 3% of full RL compute; ASFT-LoRA halves GPU memory while keeping them.
OpenReview
arXiv
code
PlanGPT-VL: Enhancing Urban Planning with Domain-Specific Vision-Language Models
EMNLP 2025
He Zhu*, Junyou Su*, Minxin Chen*, Wen Wang, Yijie Deng, Guanhua Chen, Wenjia Zhang
Conference on Empirical Methods in Natural Language Processing — Industry Track
- The PlanAnno-V pipeline synthesises high-quality visual question answering data for urban-planning maps.
- A critical-point thinking mechanism cuts hallucination through structured verification of each answer.
- Released PlanBench-V, a benchmark for reading national land-use planning maps.
ACL Anthology
project page
TAG-INSTRUCT: Controlled Instruction Complexity Enhancement through Structure-based Augmentation
ACL 2025 Findings
He Zhu, Zhiwen Ruan, Junyou Su, Xingwei He, Yun Chen, Wenjia Zhang, Guanhua Chen
Findings of the Association for Computational Linguistics
- Compresses instructions into a compact tag space, then expands them again under reinforcement-learning-guided difficulty control.
- Makes instruction complexity a dial rather than a side effect of prompt rewriting, with better stability across benchmarks.
ACL Anthology
DUET: Relational Query-Level Uncertainty at the Assistant Boundary
Under review
Junyou Su, Wen Wang, Yurun Wang, Junyang Chen, Wenjia Zhang
- Reads only the next-token distributions a target model and a few fixed peers place at the end of the prompt, so nothing is generated first.
- Train-free and label-free: cross-model agreement adds 2.45 macro AUROC over a confidence-only control on the same prefills.
* equal contribution (co-first author).
Preprints
FANNO: Augmenting High-Quality Instruction Data with Open-Sourced LLMs Only
arXiv:2408.01323
He Zhu, Junyou Su, Tianle Lun, Yicheng Tao, Wenjia Zhang, Zipei Fan, Guanhua Chen
- End-to-end instruction synthesis from raw documents using open models only: pre-filtering, seed generation, then UCB-driven augmentation.
- 10K synthesised instructions outperform Alpaca-GPT4 and WizardLM on AlpacaEval 2.0 and Arena-Hard.
arXiv
PlanGPT: Enhancing Urban Planning with Tailored Language Model and Efficient Retrieval
arXiv:2402.19273
He Zhu, Wenjia Zhang, Nuoxian Huang, …, Junyou Su, … (12 authors)
- A tailored language model and retrieval stack for Chinese urban-planning documents — the first entry in the PlanGPT line of work.
arXiv
project page
Education
M.S., research track · Shenzhen Graduate School · Advisor: Prof. Wenjia Zhang
- Focus: data synthesis, domain-specific LLMs, model post-training.
B.Eng. in Computer Science and Technology · Advisor: Prof. Xuan Song
- GPA 3.87/4.00, top 5%; University Outstanding Graduate and University Outstanding Thesis Award.
Experience
ByteDance Seed — Foundation Model Intern
2026.02 – present
Seed Speech team · China
- Data for speech foundation models (Seed-Audio): building and running the pipeline behind them.
Supervisors: Prof. Ryosuke Shibasaki, Dr. Zipei Fan · Tokyo, Japan / remote
- Fine-tuned LLMs for ship ETA and trajectory prediction, then added a maritime-news analysis agent that improves ETA accuracy.
- With the University of Tokyo, built Mobility Agent on LangChain — an LLM agent with a spatiotemporal trajectory-analysis toolset.
Delivered Systems
RAG assistant for planning knowledge
2025.03 – 2025.06
Beijing Municipal Institute of City Planning & Design · technical lead
- Large-scale OCR ingestion of scanned planning documents into a structured knowledge base, with hierarchical retrieval for precise low-latency answers and voice input for hands-free use.
RAG question answering and automated document review
2024.09 – 2024.12
China Academy of Urban Planning & Design, Shenzhen · technical lead
- LLM-driven partitioning of institutional documents into sub-knowledge bases plus hierarchical retrieval; shipped question answering over internal and web sources and automated compliance review.
Other project work
- Shenzhen Marine Development Research Promotion Center — co-built a RAG agent for automated review of planning documents.
- Undergraduate thesis (2024.09–2025.06, university Outstanding Thesis Award) — an LLM agent that mines maritime news to improve ship arrival-time prediction.
- Mobility Agent (2024.10–2025.03, with the University of Tokyo) — LangChain workflow and toolset for natural-language spatiotemporal analysis.
Skills
- Languages
- Java, Python > C++, C, SQL > Go, Vue, JavaScript > Verilog, MATLAB
- LLM
- Data synthesis and selection, SFT / LoRA, RL post-training, prompt engineering, retrieval-augmented generation, agents and multi-agent simulation, deep research
- Engineering
- Full-stack web development; distributed training and evaluation pipelines; LLaMA-Factory contributor
Honors & Awards
- 2025
- SUSTech Outstanding Graduate and Outstanding Undergraduate Thesis Award, both university level
- 2023, 2024
- SUSTech Outstanding Student Award, university level
- 2022–2024
- SUSTech Outstanding Student Scholarship, 3rd Prize
- 2023
- National College Mathematical Modeling Contest, Provincial 3rd Prize
- 2021
- SUSTech Freshman Scholarship, 2nd Prize
