Master's Student · Engineer · Builder

Hi, I am Zhe Huang.

I am based in Beijing and currently studying Intelligence Science and Technology at Beijing University of Posts and Telecommunications. I like turning research ideas into working systems, useful tools, and clear experiments.

Portrait of Zhe Huang
Beijing, China
Now BUPT, M.Eng.

Intelligence Science and Technology, expected June 2026.

Home Base Beijing, China

Working across academic projects, engineering practice, and open-source tools.

Tools Python & AI Systems

PyTorch, vLLM, Verl, Linux, data pipelines, and practical automation.

About

A computer science student who enjoys building things that work.

I studied Computer Science and Technology at Beijing Forestry University and now pursue a master's degree at BUPT. My day-to-day work sits between machine learning research and engineering: reading papers, reproducing methods, running experiments, writing code, and making systems easier to use.

  • Comfortable moving from problem formulation to implementation and evaluation.
  • Interested in AI systems, multimodal models, recommendation, and developer tools.
  • Prefer practical artifacts: runnable code, reproducible experiments, and clear notes.

Work

Projects I have worked on.

CARE

Competence-aware reward shaping for adaptive reasoning length in Video-MLLMs.

  • Designed dynamic rewards for GRPO-based reasoning control.
  • Improved reasoning efficiency and answer accuracy in video QA.
  • Validated adaptive behavior through ablations and case analyses.
View code

VideoDR Benchmark

Open-web video deep research benchmark for agentic video reasoning.

  • Co-developed benchmark tasks for open-world video reasoning.
  • Built agents combining video understanding, web retrieval, and multi-hop reasoning.
  • Designed evaluation for retrieval consistency and evidence grounding.
View code

Graph-Augmented Product Bundling

LLM and graph architecture for product bundle construction.

  • Combined Qwen2, LightGCN, and text-form graph injection.
  • Integrated MCTS strategies during decoding.
  • Evaluated on Alibaba POG and Steam datasets.

Dialogue Generation

Iterative prompt-enhanced large model dialogue generation.

  • Built vLLM-based inference pipelines.
  • Used multi-sample generation and draft optimization.
  • Evaluated improvements on DailyDialog with Alpaca-7B and Vicuna-13B.

Writing

Publications and manuscripts.

Preprint · 2026

Watching, Reasoning, and Searching: A Video Deep Research Benchmark on Open Web for Agentic Video Reasoning

Liu, C., Yu, X., Chang, Z., Huang, Z., Zhang, S., Lian, H., Wang, K., Xu, R., Hu, S., Hou, J., Peng, H., Qin, C., Hu, X., Peng, H., Chen, R., & Wang, H.

doi.org/10.48550/ARXIV.2601.06943

Under Review · ICML 2026

Reasoning as Intersection: VideoCFR for Visual Focus Alignment in Video-MLLMs

A weakly supervised approach for aligning model focus with key video frames in video reasoning.

ICCCS · 2024

Enhancing Dialogue Generation Through Iterative Prompt with Difference Analogy Learning

Wang, M., Huang, Z., Yuan, C., Liu, H., Liu, T., & Wang, X. 2024 9th International Conference on Computer and Communication Systems, 1363-1368.

doi.org/10.1109/ICCCS61882.2024.10603103

In Preparation

CARE: Competence-Aware Reward Shaping for Adaptive Reasoning Length in Video-MLLMs

Adaptive reward shaping for efficient reasoning control in video question answering and reasoning tasks.

Experience and Skills

From experiments to working pipelines.

Oct 2024 - Dec 2025

Algorithm Engineering Intern

Beijing Zhi Jing Ling Technology Co., Ltd.

Worked on graph-augmented LLM product bundling, including method design, implementation, evaluation, and decoding optimization.

Python PyTorch NumPy Pandas PEFT LLaMA Factory Verl vLLM Linux Agent Frameworks