Zeyu Wang

Zeyu Wang

MPhil student at HKUST(GZ)

About me

I am Zeyu Wang, an MPhil student at The Hong Kong University of Science and Technology (Guangzhou). I received my bachelor’s degree in Computer Science and Technology from Taiyuan University of Technology.

My interests lie in large language model post-training, reinforcement learning, retrieval-augmented generation (RAG), and AI agents. My work spans controllable creative writing at IDEA, search-oriented language models at Baidu, and agent systems for campus information and customer support. During my undergraduate studies, I also worked on audio-visual speaker tracking in the Multimodal Human-Robot Interaction (MIHRI) Lab.

I enjoy building open-source tools and sharing what I learn through GitHub and my blog.

Education

  • The Hong Kong University of Science and Technology (Guangzhou) — Master of Philosophy (MPhil) in Artificial Intelligence, September 2026–June 2028 (expected).
  • Taiyuan University of Technology — Bachelor’s in Computer Science and Technology, September 2022–June 2026. Comprehensive Assessment Rank: 2/149.

Research & Publications

IDEA — Controllable Creative Writing

July 2025–April 2026

Worked on improving short-form video script generation, moving from a hosted R1 API to a local Qwen2.5-7B model. Combined expert-guided cold-start data, distillation, and data filtering with GRPO using natural-language critiques as a non-rule-based reward, improving informativeness and rhetorical quality.

From Homogeneity to Diversity: Self-Refined GRPO for Controllable Creative Writing
Minna Peng, Wei Tan, Zeyu Wang, Yongquan Hu, Ruizhi Huang, Yang Jie.
Accepted to Findings of EMNLP 2026.

MIHRI Lab, Taiyuan University of Technology — Audio-Visual Speaker Tracking

May 2024–December 2024

Worked on a three-stage multimodal distillation framework using SiamFC and stGCF, with symmetric cross-attention for audio-visual fusion. Achieved a mean absolute tracking error of 3.02 pixels on AV16.3. I was the third author.

Multi-Stage Multimodal Distillation for Audio-Visual Speaker Tracking
Yidi Li, Wenkai Zhao, Zeyu Wang, Zhenhuan Xu, Bin Ren, Nicu Sebe.
IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2025. DOI

Internships

Baidu — NLP Algorithm Intern

May 2025–July 2025

Improved e-commerce search term extraction from user queries, increasing micro-F1 from 83.2% to 87.3%. Applied a Search-R1-inspired approach to retrieval decisions and search term generation, training Qwen2.5-7B with GRPO and rule-based rewards on 10,000 GPT-4o-assisted annotations, including 3,000 cold-start examples.

SentimenTrader — AI Agent Algorithm Intern

October 2024–April 2025

Built an AutoAgent for financial decision support, combining RAG over articles and YouTube content with iterative web operations for stock backtesting. Improved retrieval with parallel semantic and keyword queries, filtering, and reranking.

Strikingly — RAG Development Intern

December 2023–June 2024

Developed GPT-4-based translation and customer-support applications using Milvus, domain terminology, and Agentic RAG. Improved customer-service efficiency by 60%+, with 92% recall and 0.78 MRR for support QA.

Projects & Open Source

  • 2026 Tencent Rhino-Bird Open-Source Talent Program (Hunyuan LLM track) — One of only three participants in the Hunyuan LLM track to receive the Open-Source Task cash award. Official announcement
  • TYUT AI Counselor — Initiator and lead developer. Built a campus information assistant with hybrid retrieval and prompt-injection screening; used 100,000+ times across the university and featured in Shanxi Daily.
  • Home Agent — Team leader and lead developer. Extended Blockly to support Xiaomi smart-home APIs and fine-tuned Qwen3 with LoRA to translate natural language into JSON blocks. Won the Hardcore Technology Award (1/98) at the Qwen On-Device AI Innovation Challenge in December 2025.
  • AI Interviewer — Team leader and lead developer. Built a multi-agent workflow for resume assessment, RAG-based question retrieval, answer evaluation, and scoring.
  • self-llm — Core contributor. Contributed Qwen3 architecture analysis, Kimi-VL-A3B technical report explanations, OpenELM LoRA fine-tuning, and deployment guides for WebDemo, vLLM, and FastAPI.
  • Mutual-AI — Core contributor in algorithms and infrastructure. Built interactive CV and NLP demos with CNNs, BERT, and FastAPI, deployed across five servers.

Selected Awards

  • Open-Source Task cash award — 2026 Tencent Rhino-Bird Open-Source Talent Program; one of only three recipients in the Hunyuan LLM track. Official announcement
  • National Second Prize & First Place in Shanxi Province — National College Student Career Planning Competition, 2026.
  • Hardcore Technology Award (1/98) — Qwen On-Device AI Innovation Challenge, 2025; team leader.
  • AMD Hardware Optimization Award (2/50) — Alibaba ModelScope MCP & Agent Challenge, 2025; team leader.
  • Shanxi Provincial Special Prize — Challenge Cup, 2025; team leader.
  • National Special Prize (1/380) — China University Students’ Intelligent Lighting and Intelligent Wearable Innovation and Entrepreneurship Competition, 2024; team leader.
  • International First Prize — International Youth Artificial Intelligence Competition, 2024; team leader.
  • National Second Prize — China Undergraduate Mathematical Contest in Modeling, 2023.